Spaces:
Runtime error
Runtime error
copied from dreamcreature main repo
Browse files- .gitignore +8 -0
- __init__.py +0 -0
- app.py +215 -0
- data/cub200_2011/class_names.txt +200 -0
- data/cub200_2011/pretrained_kmeans.pth +3 -0
- data/cub200_2011/train.txt +0 -0
- data/cub200_2011/train_caps_better_m8_k256.txt +0 -0
- data/dogs/class_names.txt +120 -0
- data/dogs/pretrained_kmeans.pth +3 -0
- data/dogs/train.txt +0 -0
- data/dogs/train_caps_better_m8_k256.txt +0 -0
- dreamcreature/__init__.py +0 -0
- dreamcreature/attn_processor.py +143 -0
- dreamcreature/dataset.py +155 -0
- dreamcreature/dino.py +42 -0
- dreamcreature/kmeans_segmentation.py +188 -0
- dreamcreature/loss.py +73 -0
- dreamcreature/mapper.py +75 -0
- dreamcreature/pipeline.py +771 -0
- dreamcreature/pipeline_xl.py +895 -0
- dreamcreature/text_encoder.py +201 -0
- dreamcreature/tokenizer.py +109 -0
- gradio_demo_cub200.py +166 -0
- gradio_demo_dog.py +166 -0
- requirements.txt +12 -0
- run_sd_sup.sh +12 -0
- run_sd_unsup.sh +12 -0
- run_sdxl_sup.sh +13 -0
- run_sdxl_unsup.sh +13 -0
- train_dreamcreature_sd.py +1122 -0
- train_dreamcreature_sdxl.py +1539 -0
- train_kmeans_segmentation.ipynb +578 -0
- utils.py +127 -0
.gitignore
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
.idea/
|
| 2 |
+
src/data/cub200_2011/train
|
| 3 |
+
src/data/dogs/Images
|
| 4 |
+
__pycache__
|
| 5 |
+
*/.ipynb_checkpoints
|
| 6 |
+
/.ipynb_checkpoints/requirements-checkpoint.txt
|
| 7 |
+
|
| 8 |
+
*.bin
|
__init__.py
ADDED
|
File without changes
|
app.py
ADDED
|
@@ -0,0 +1,215 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gc
|
| 2 |
+
import os
|
| 3 |
+
import re
|
| 4 |
+
import shutil
|
| 5 |
+
|
| 6 |
+
import gradio as gr
|
| 7 |
+
import requests
|
| 8 |
+
import torch
|
| 9 |
+
|
| 10 |
+
from dreamcreature.pipeline import create_args, load_pipeline
|
| 11 |
+
|
| 12 |
+
CUB_DESCRIPTION = """
|
| 13 |
+
# DreamCreature (CUB-200-2011)
|
| 14 |
+
To create your own creature, you can type:
|
| 15 |
+
|
| 16 |
+
`"a photo of a <head:id> <wing:id> bird"` where `id` ranges from 1~200 (200 classes corresponding to CUB-200-2011)
|
| 17 |
+
|
| 18 |
+
For instance `"a photo of a <head:17> <wing:18> bird"` using head of `cardinal (17)` and wing of `spotted catbird (18)`
|
| 19 |
+
|
| 20 |
+
Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/cub200_2011/class_names.txt
|
| 21 |
+
|
| 22 |
+
You can also try any prompt you like such as:
|
| 23 |
+
|
| 24 |
+
Sub-concept transfer: `"a photo of a <wing:17> cat"`
|
| 25 |
+
|
| 26 |
+
Inspiring design: `"a photo of a <head:101> <wing:191> teddy bear"`
|
| 27 |
+
|
| 28 |
+
(Experimental) You can also use two parts together such as:
|
| 29 |
+
|
| 30 |
+
`"a photo of a <head:17> <head:18> bird"` mixing head of `cardinal (17)` and `spotted catbird (18)`
|
| 31 |
+
|
| 32 |
+
The current available parts are: `head`, `body`, `wing`, `tail`, and `leg`
|
| 33 |
+
|
| 34 |
+
"""
|
| 35 |
+
|
| 36 |
+
DOG_DESCRIPTION = """
|
| 37 |
+
# DreamCreature (Stanford Dogs)
|
| 38 |
+
To create your own creature, you can type:
|
| 39 |
+
|
| 40 |
+
`"a photo of a <nose:id> <ear:id> dog"` where `id` ranges from 0~119 (120 classes corresponding to Stanford Dogs)
|
| 41 |
+
|
| 42 |
+
For instance `"a photo of a <nose:2> <ear:112> dog"` using head of `maltese dog (2)` and wing of `cardigan (112)`
|
| 43 |
+
|
| 44 |
+
Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/dogs/class_names.txt
|
| 45 |
+
|
| 46 |
+
Sub-concept transfer: `"a photo of a <ear:112> cat"`
|
| 47 |
+
|
| 48 |
+
Inspiring design: `"a photo of a <eye:38> <body:38> teddy bear"`
|
| 49 |
+
|
| 50 |
+
(Experimental) You can also use two parts together such as:
|
| 51 |
+
|
| 52 |
+
`"a photo of a <nose:1> <nose:112> dog"` mixing head of `maltese dog (2)` and `spotted cardigan (112)`
|
| 53 |
+
|
| 54 |
+
The current available parts are: `eye`, `neck`, `ear`, `body`, `leg`, `nose` and `forehead`
|
| 55 |
+
|
| 56 |
+
"""
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
def prepare_pipeline(model_name):
|
| 60 |
+
is_cub = 'cub' in model_name
|
| 61 |
+
|
| 62 |
+
checkpoint_name = {
|
| 63 |
+
'dreamcreature-sd1.5-cub200': 'checkpoint-74900',
|
| 64 |
+
'dreamcreature-sd1.5-dog': 'checkpoint-150000'
|
| 65 |
+
}
|
| 66 |
+
|
| 67 |
+
repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
|
| 68 |
+
file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
|
| 69 |
+
local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
|
| 70 |
+
os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
|
| 71 |
+
download_file(file_url, local_path)
|
| 72 |
+
|
| 73 |
+
file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
|
| 74 |
+
local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
|
| 75 |
+
download_file(file_url, local_path)
|
| 76 |
+
|
| 77 |
+
OUTPUT_DIR = model_name
|
| 78 |
+
|
| 79 |
+
args = create_args(OUTPUT_DIR)
|
| 80 |
+
if 'dpo' in OUTPUT_DIR:
|
| 81 |
+
args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"
|
| 82 |
+
|
| 83 |
+
pipe = load_pipeline(args, torch.float16, 'cuda')
|
| 84 |
+
pipe = pipe.to(torch.float16)
|
| 85 |
+
|
| 86 |
+
pipe.verbose = True
|
| 87 |
+
pipe.v = 're'
|
| 88 |
+
|
| 89 |
+
if is_cub:
|
| 90 |
+
pipe.num_k_per_part = 200
|
| 91 |
+
|
| 92 |
+
MAPPING = {
|
| 93 |
+
'body': 0,
|
| 94 |
+
'tail': 1,
|
| 95 |
+
'head': 2,
|
| 96 |
+
'wing': 4,
|
| 97 |
+
'leg': 6
|
| 98 |
+
}
|
| 99 |
+
|
| 100 |
+
ID2NAME = open('data/cub200_2011/class_names.txt').readlines()
|
| 101 |
+
ID2NAME = [line.strip() for line in ID2NAME]
|
| 102 |
+
|
| 103 |
+
else:
|
| 104 |
+
pipe.num_k_per_part = 120
|
| 105 |
+
|
| 106 |
+
MAPPING = {
|
| 107 |
+
'eye': 0,
|
| 108 |
+
'neck': 2,
|
| 109 |
+
'ear': 3,
|
| 110 |
+
'body': 4,
|
| 111 |
+
'leg': 5,
|
| 112 |
+
'nose': 6,
|
| 113 |
+
'forehead': 7
|
| 114 |
+
}
|
| 115 |
+
|
| 116 |
+
ID2NAME = open('data/dogs/class_names.txt').readlines()
|
| 117 |
+
ID2NAME = [line.strip() for line in ID2NAME]
|
| 118 |
+
|
| 119 |
+
return pipe, MAPPING, ID2NAME
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
def download_file(url, local_path):
|
| 123 |
+
if os.path.exists(local_path):
|
| 124 |
+
return
|
| 125 |
+
|
| 126 |
+
with requests.get(url, stream=True) as r:
|
| 127 |
+
with open(local_path, 'wb') as f:
|
| 128 |
+
shutil.copyfileobj(r.raw, f)
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
def process_text(text, MAPPING, ID2NAME):
|
| 132 |
+
pattern = r"<([^:>]+):(\d+)>"
|
| 133 |
+
result = text
|
| 134 |
+
offset = 0
|
| 135 |
+
|
| 136 |
+
part2id = []
|
| 137 |
+
|
| 138 |
+
for match in re.finditer(pattern, text):
|
| 139 |
+
key = match.group(1)
|
| 140 |
+
clsid = int(match.group(2))
|
| 141 |
+
clsid = min(max(clsid, 1), 200) # must be 1~200
|
| 142 |
+
|
| 143 |
+
replacement = f"<{MAPPING[key]}:{clsid - 1}>"
|
| 144 |
+
start, end = match.span()
|
| 145 |
+
|
| 146 |
+
# Adjust the start and end positions based on the offset from previous replacements
|
| 147 |
+
start += offset
|
| 148 |
+
end += offset
|
| 149 |
+
|
| 150 |
+
# Replace the matched text with the replacement
|
| 151 |
+
result = result[:start] + replacement + result[end:]
|
| 152 |
+
|
| 153 |
+
# Update the offset for the next replacement
|
| 154 |
+
offset += len(replacement) - (end - start)
|
| 155 |
+
|
| 156 |
+
part2id.append(f'{key}: {ID2NAME[clsid - 1]}')
|
| 157 |
+
|
| 158 |
+
return result, part2id
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
def generate_images(model_name, prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
|
| 162 |
+
generator = torch.Generator(device='cuda')
|
| 163 |
+
generator = generator.manual_seed(int(seed))
|
| 164 |
+
|
| 165 |
+
try:
|
| 166 |
+
pipe, MAPPING, ID2NAME = prepare_pipeline(model_name)
|
| 167 |
+
|
| 168 |
+
prompt, part2id = process_text(prompt, MAPPING, ID2NAME)
|
| 169 |
+
negative_prompt, _ = process_text(negative_prompt, MAPPING, ID2NAME)
|
| 170 |
+
|
| 171 |
+
images = pipe(prompt,
|
| 172 |
+
negative_prompt=negative_prompt, generator=generator,
|
| 173 |
+
num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
|
| 174 |
+
num_images_per_prompt=num_images).images
|
| 175 |
+
|
| 176 |
+
del pipe
|
| 177 |
+
except Exception as e:
|
| 178 |
+
raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
|
| 179 |
+
f"The error message: {e}")
|
| 180 |
+
finally:
|
| 181 |
+
gc.collect()
|
| 182 |
+
torch.cuda.empty_cache()
|
| 183 |
+
|
| 184 |
+
return images, '; '.join(part2id)
|
| 185 |
+
|
| 186 |
+
|
| 187 |
+
with gr.Blocks(title="DreamCreature") as demo:
|
| 188 |
+
with gr.Row():
|
| 189 |
+
main_desc = gr.Markdown(CUB_DESCRIPTION)
|
| 190 |
+
with gr.Column():
|
| 191 |
+
with gr.Row():
|
| 192 |
+
with gr.Group():
|
| 193 |
+
dropdown = gr.Dropdown(choices=["dreamcreature-sd1.5-cub200",
|
| 194 |
+
"dreamcreature-sd1.5-dog"],
|
| 195 |
+
value="dreamcreature-sd1.5-cub200")
|
| 196 |
+
prompt = gr.Textbox(label="Prompt", value="a photo of a <head:101> <wing:191> teddy bear")
|
| 197 |
+
negative_prompt = gr.Textbox(label="Negative Prompt",
|
| 198 |
+
value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
|
| 199 |
+
num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
|
| 200 |
+
guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
|
| 201 |
+
num_images = gr.Slider(minimum=1, maximum=4, step=1, value=4, label="Number of Images")
|
| 202 |
+
seed = gr.Number(label="Seed", value=777881414)
|
| 203 |
+
button = gr.Button()
|
| 204 |
+
|
| 205 |
+
with gr.Column():
|
| 206 |
+
output_images = gr.Gallery(columns=4, label='Output')
|
| 207 |
+
markdown_labels = gr.Markdown("")
|
| 208 |
+
|
| 209 |
+
dropdown.change(fn=lambda x: {'dreamcreature-sd1.5-cub200': CUB_DESCRIPTION,
|
| 210 |
+
'dreamcreature-sd1.5-dog': DOG_DESCRIPTION}[x], inputs=dropdown, outputs=main_desc)
|
| 211 |
+
button.click(fn=generate_images,
|
| 212 |
+
inputs=[dropdown, prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
|
| 213 |
+
seed], outputs=[output_images, markdown_labels], show_progress=True)
|
| 214 |
+
|
| 215 |
+
demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')
|
data/cub200_2011/class_names.txt
ADDED
|
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
black footed albatross
|
| 2 |
+
laysan albatross
|
| 3 |
+
sooty albatross
|
| 4 |
+
groove billed ani
|
| 5 |
+
crested auklet
|
| 6 |
+
least auklet
|
| 7 |
+
parakeet auklet
|
| 8 |
+
rhinoceros auklet
|
| 9 |
+
brewer blackbird
|
| 10 |
+
red winged blackbird
|
| 11 |
+
rusty blackbird
|
| 12 |
+
yellow headed blackbird
|
| 13 |
+
bobolink
|
| 14 |
+
indigo bunting
|
| 15 |
+
lazuli bunting
|
| 16 |
+
painted bunting
|
| 17 |
+
cardinal
|
| 18 |
+
spotted catbird
|
| 19 |
+
gray catbird
|
| 20 |
+
yellow breasted chat
|
| 21 |
+
eastern towhee
|
| 22 |
+
chuck will widow
|
| 23 |
+
brandt cormorant
|
| 24 |
+
red faced cormorant
|
| 25 |
+
pelagic cormorant
|
| 26 |
+
bronzed cowbird
|
| 27 |
+
shiny cowbird
|
| 28 |
+
brown creeper
|
| 29 |
+
american crow
|
| 30 |
+
fish crow
|
| 31 |
+
black billed cuckoo
|
| 32 |
+
mangrove cuckoo
|
| 33 |
+
yellow billed cuckoo
|
| 34 |
+
gray crowned rosy finch
|
| 35 |
+
purple finch
|
| 36 |
+
northern flicker
|
| 37 |
+
acadian flycatcher
|
| 38 |
+
great crested flycatcher
|
| 39 |
+
least flycatcher
|
| 40 |
+
olive sided flycatcher
|
| 41 |
+
scissor tailed flycatcher
|
| 42 |
+
vermilion flycatcher
|
| 43 |
+
yellow bellied flycatcher
|
| 44 |
+
frigatebird
|
| 45 |
+
northern fulmar
|
| 46 |
+
gadwall
|
| 47 |
+
american goldfinch
|
| 48 |
+
european goldfinch
|
| 49 |
+
boat tailed grackle
|
| 50 |
+
eared grebe
|
| 51 |
+
horned grebe
|
| 52 |
+
pied billed grebe
|
| 53 |
+
western grebe
|
| 54 |
+
blue grosbeak
|
| 55 |
+
evening grosbeak
|
| 56 |
+
pine grosbeak
|
| 57 |
+
rose breasted grosbeak
|
| 58 |
+
pigeon guillemot
|
| 59 |
+
california gull
|
| 60 |
+
glaucous winged gull
|
| 61 |
+
heermann gull
|
| 62 |
+
herring gull
|
| 63 |
+
ivory gull
|
| 64 |
+
ring billed gull
|
| 65 |
+
slaty backed gull
|
| 66 |
+
western gull
|
| 67 |
+
anna hummingbird
|
| 68 |
+
ruby throated hummingbird
|
| 69 |
+
rufous hummingbird
|
| 70 |
+
green violetear
|
| 71 |
+
long tailed jaeger
|
| 72 |
+
pomarine jaeger
|
| 73 |
+
blue jay
|
| 74 |
+
florida jay
|
| 75 |
+
green jay
|
| 76 |
+
dark eyed junco
|
| 77 |
+
tropical kingbird
|
| 78 |
+
gray kingbird
|
| 79 |
+
belted kingfisher
|
| 80 |
+
green kingfisher
|
| 81 |
+
pied kingfisher
|
| 82 |
+
ringed kingfisher
|
| 83 |
+
white breasted kingfisher
|
| 84 |
+
red legged kittiwake
|
| 85 |
+
horned lark
|
| 86 |
+
pacific loon
|
| 87 |
+
mallard
|
| 88 |
+
western meadowlark
|
| 89 |
+
hooded merganser
|
| 90 |
+
red breasted merganser
|
| 91 |
+
mockingbird
|
| 92 |
+
nighthawk
|
| 93 |
+
clark nutcracker
|
| 94 |
+
white breasted nuthatch
|
| 95 |
+
baltimore oriole
|
| 96 |
+
hooded oriole
|
| 97 |
+
orchard oriole
|
| 98 |
+
scott oriole
|
| 99 |
+
ovenbird
|
| 100 |
+
brown pelican
|
| 101 |
+
white pelican
|
| 102 |
+
western wood pewee
|
| 103 |
+
sayornis
|
| 104 |
+
american pipit
|
| 105 |
+
whip poor will
|
| 106 |
+
horned puffin
|
| 107 |
+
common raven
|
| 108 |
+
white necked raven
|
| 109 |
+
american redstart
|
| 110 |
+
geococcyx
|
| 111 |
+
loggerhead shrike
|
| 112 |
+
great grey shrike
|
| 113 |
+
baird sparrow
|
| 114 |
+
black throated sparrow
|
| 115 |
+
brewer sparrow
|
| 116 |
+
chipping sparrow
|
| 117 |
+
clay colored sparrow
|
| 118 |
+
house sparrow
|
| 119 |
+
field sparrow
|
| 120 |
+
fox sparrow
|
| 121 |
+
grasshopper sparrow
|
| 122 |
+
harris sparrow
|
| 123 |
+
henslow sparrow
|
| 124 |
+
le conte sparrow
|
| 125 |
+
lincoln sparrow
|
| 126 |
+
nelson sharp tailed sparrow
|
| 127 |
+
savannah sparrow
|
| 128 |
+
seaside sparrow
|
| 129 |
+
song sparrow
|
| 130 |
+
tree sparrow
|
| 131 |
+
vesper sparrow
|
| 132 |
+
white crowned sparrow
|
| 133 |
+
white throated sparrow
|
| 134 |
+
cape glossy starling
|
| 135 |
+
bank swallow
|
| 136 |
+
barn swallow
|
| 137 |
+
cliff swallow
|
| 138 |
+
tree swallow
|
| 139 |
+
scarlet tanager
|
| 140 |
+
summer tanager
|
| 141 |
+
artic tern
|
| 142 |
+
black tern
|
| 143 |
+
caspian tern
|
| 144 |
+
common tern
|
| 145 |
+
elegant tern
|
| 146 |
+
forsters tern
|
| 147 |
+
least tern
|
| 148 |
+
green tailed towhee
|
| 149 |
+
brown thrasher
|
| 150 |
+
sage thrasher
|
| 151 |
+
black capped vireo
|
| 152 |
+
blue headed vireo
|
| 153 |
+
philadelphia vireo
|
| 154 |
+
red eyed vireo
|
| 155 |
+
warbling vireo
|
| 156 |
+
white eyed vireo
|
| 157 |
+
yellow throated vireo
|
| 158 |
+
bay breasted warbler
|
| 159 |
+
black and white warbler
|
| 160 |
+
black throated blue warbler
|
| 161 |
+
blue winged warbler
|
| 162 |
+
canada warbler
|
| 163 |
+
cape may warbler
|
| 164 |
+
cerulean warbler
|
| 165 |
+
chestnut sided warbler
|
| 166 |
+
golden winged warbler
|
| 167 |
+
hooded warbler
|
| 168 |
+
kentucky warbler
|
| 169 |
+
magnolia warbler
|
| 170 |
+
mourning warbler
|
| 171 |
+
myrtle warbler
|
| 172 |
+
nashville warbler
|
| 173 |
+
orange crowned warbler
|
| 174 |
+
palm warbler
|
| 175 |
+
pine warbler
|
| 176 |
+
prairie warbler
|
| 177 |
+
prothonotary warbler
|
| 178 |
+
swainson warbler
|
| 179 |
+
tennessee warbler
|
| 180 |
+
wilson warbler
|
| 181 |
+
worm eating warbler
|
| 182 |
+
yellow warbler
|
| 183 |
+
northern waterthrush
|
| 184 |
+
louisiana waterthrush
|
| 185 |
+
bohemian waxwing
|
| 186 |
+
cedar waxwing
|
| 187 |
+
american three toed woodpecker
|
| 188 |
+
pileated woodpecker
|
| 189 |
+
red bellied woodpecker
|
| 190 |
+
red cockaded woodpecker
|
| 191 |
+
red headed woodpecker
|
| 192 |
+
downy woodpecker
|
| 193 |
+
bewick wren
|
| 194 |
+
cactus wren
|
| 195 |
+
carolina wren
|
| 196 |
+
house wren
|
| 197 |
+
marsh wren
|
| 198 |
+
rock wren
|
| 199 |
+
winter wren
|
| 200 |
+
common yellowthroat
|
data/cub200_2011/pretrained_kmeans.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:00b2ff84f80daa3cdbd4b18e4088fd900b70a8a192a70957c34e4369d6065e65
|
| 3 |
+
size 6874495
|
data/cub200_2011/train.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/cub200_2011/train_caps_better_m8_k256.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/dogs/class_names.txt
ADDED
|
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Chihuahua
|
| 2 |
+
Japanese spaniel
|
| 3 |
+
Maltese dog
|
| 4 |
+
Pekinese
|
| 5 |
+
Shih Tzu
|
| 6 |
+
Blenheim spaniel
|
| 7 |
+
papillon
|
| 8 |
+
toy terrier
|
| 9 |
+
Rhodesian ridgeback
|
| 10 |
+
Afghan hound
|
| 11 |
+
basset
|
| 12 |
+
beagle
|
| 13 |
+
bloodhound
|
| 14 |
+
bluetick
|
| 15 |
+
black and tan coonhound
|
| 16 |
+
Walker hound
|
| 17 |
+
English foxhound
|
| 18 |
+
redbone
|
| 19 |
+
borzoi
|
| 20 |
+
Irish wolfhound
|
| 21 |
+
Italian greyhound
|
| 22 |
+
whippet
|
| 23 |
+
Ibizan hound
|
| 24 |
+
Norwegian elkhound
|
| 25 |
+
otterhound
|
| 26 |
+
Saluki
|
| 27 |
+
Scottish deerhound
|
| 28 |
+
Weimaraner
|
| 29 |
+
Staffordshire bullterrier
|
| 30 |
+
American Staffordshire terrier
|
| 31 |
+
Bedlington terrier
|
| 32 |
+
Border terrier
|
| 33 |
+
Kerry blue terrier
|
| 34 |
+
Irish terrier
|
| 35 |
+
Norfolk terrier
|
| 36 |
+
Norwich terrier
|
| 37 |
+
Yorkshire terrier
|
| 38 |
+
wire haired fox terrier
|
| 39 |
+
Lakeland terrier
|
| 40 |
+
Sealyham terrier
|
| 41 |
+
Airedale
|
| 42 |
+
cairn
|
| 43 |
+
Australian terrier
|
| 44 |
+
Dandie Dinmont
|
| 45 |
+
Boston bull
|
| 46 |
+
miniature schnauzer
|
| 47 |
+
giant schnauzer
|
| 48 |
+
standard schnauzer
|
| 49 |
+
Scotch terrier
|
| 50 |
+
Tibetan terrier
|
| 51 |
+
silky terrier
|
| 52 |
+
soft coated wheaten terrier
|
| 53 |
+
West Highland white terrier
|
| 54 |
+
Lhasa
|
| 55 |
+
flat coated retriever
|
| 56 |
+
curly coated retriever
|
| 57 |
+
golden retriever
|
| 58 |
+
Labrador retriever
|
| 59 |
+
Chesapeake Bay retriever
|
| 60 |
+
German short haired pointer
|
| 61 |
+
vizsla
|
| 62 |
+
English setter
|
| 63 |
+
Irish setter
|
| 64 |
+
Gordon setter
|
| 65 |
+
Brittany spaniel
|
| 66 |
+
clumber
|
| 67 |
+
English springer
|
| 68 |
+
Welsh springer spaniel
|
| 69 |
+
cocker spaniel
|
| 70 |
+
Sussex spaniel
|
| 71 |
+
Irish water spaniel
|
| 72 |
+
kuvasz
|
| 73 |
+
schipperke
|
| 74 |
+
groenendael
|
| 75 |
+
malinois
|
| 76 |
+
briard
|
| 77 |
+
kelpie
|
| 78 |
+
komondor
|
| 79 |
+
Old English sheepdog
|
| 80 |
+
Shetland sheepdog
|
| 81 |
+
collie
|
| 82 |
+
Border collie
|
| 83 |
+
Bouvier des Flandres
|
| 84 |
+
Rottweiler
|
| 85 |
+
German shepherd
|
| 86 |
+
Doberman
|
| 87 |
+
miniature pinscher
|
| 88 |
+
Greater Swiss Mountain dog
|
| 89 |
+
Bernese mountain dog
|
| 90 |
+
Appenzeller
|
| 91 |
+
EntleBucher
|
| 92 |
+
boxer
|
| 93 |
+
bull mastiff
|
| 94 |
+
Tibetan mastiff
|
| 95 |
+
French bulldog
|
| 96 |
+
Great Dane
|
| 97 |
+
Saint Bernard
|
| 98 |
+
Eskimo dog
|
| 99 |
+
malamute
|
| 100 |
+
Siberian husky
|
| 101 |
+
affenpinscher
|
| 102 |
+
basenji
|
| 103 |
+
pug
|
| 104 |
+
Leonberg
|
| 105 |
+
Newfoundland
|
| 106 |
+
Great Pyrenees
|
| 107 |
+
Samoyed
|
| 108 |
+
Pomeranian
|
| 109 |
+
chow
|
| 110 |
+
keeshond
|
| 111 |
+
Brabancon griffon
|
| 112 |
+
Pembroke
|
| 113 |
+
Cardigan
|
| 114 |
+
toy poodle
|
| 115 |
+
miniature poodle
|
| 116 |
+
standard poodle
|
| 117 |
+
Mexican hairless
|
| 118 |
+
dingo
|
| 119 |
+
dhole
|
| 120 |
+
African hunting dog
|
data/dogs/pretrained_kmeans.pth
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1bf723669b2d6dad50d58a6d7b3dad7fafa6b49d8a3fca3fd5b713662ccd4b88
|
| 3 |
+
size 6874495
|
data/dogs/train.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
data/dogs/train_caps_better_m8_k256.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
dreamcreature/__init__.py
ADDED
|
File without changes
|
dreamcreature/attn_processor.py
ADDED
|
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from diffusers.models.attention_processor import *
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
class LoRAAttnProcessorCustom(nn.Module, AttnProcessor):
|
| 5 |
+
r"""
|
| 6 |
+
Processor for implementing the LoRA attention mechanism.
|
| 7 |
+
|
| 8 |
+
Args:
|
| 9 |
+
hidden_size (`int`, *optional*):
|
| 10 |
+
The hidden size of the attention layer.
|
| 11 |
+
cross_attention_dim (`int`, *optional*):
|
| 12 |
+
The number of channels in the `encoder_hidden_states`.
|
| 13 |
+
rank (`int`, defaults to 4):
|
| 14 |
+
The dimension of the LoRA update matrices.
|
| 15 |
+
network_alpha (`int`, *optional*):
|
| 16 |
+
Equivalent to `alpha` but it's usage is specific to Kohya (A1111) style LoRAs.
|
| 17 |
+
"""
|
| 18 |
+
|
| 19 |
+
def __init__(self, hidden_size, cross_attention_dim=None, rank=4, network_alpha=None, **kwargs):
|
| 20 |
+
super().__init__()
|
| 21 |
+
|
| 22 |
+
self.hidden_size = hidden_size
|
| 23 |
+
self.cross_attention_dim = cross_attention_dim
|
| 24 |
+
self.rank = rank
|
| 25 |
+
|
| 26 |
+
q_rank = kwargs.pop("q_rank", None)
|
| 27 |
+
q_hidden_size = kwargs.pop("q_hidden_size", None)
|
| 28 |
+
q_rank = q_rank if q_rank is not None else rank
|
| 29 |
+
q_hidden_size = q_hidden_size if q_hidden_size is not None else hidden_size
|
| 30 |
+
|
| 31 |
+
v_rank = kwargs.pop("v_rank", None)
|
| 32 |
+
v_hidden_size = kwargs.pop("v_hidden_size", None)
|
| 33 |
+
v_rank = v_rank if v_rank is not None else rank
|
| 34 |
+
v_hidden_size = v_hidden_size if v_hidden_size is not None else hidden_size
|
| 35 |
+
|
| 36 |
+
out_rank = kwargs.pop("out_rank", None)
|
| 37 |
+
out_hidden_size = kwargs.pop("out_hidden_size", None)
|
| 38 |
+
out_rank = out_rank if out_rank is not None else rank
|
| 39 |
+
out_hidden_size = out_hidden_size if out_hidden_size is not None else hidden_size
|
| 40 |
+
|
| 41 |
+
self.to_q_lora = LoRALinearLayer(q_hidden_size, q_hidden_size, q_rank, network_alpha)
|
| 42 |
+
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
|
| 43 |
+
self.to_v_lora = LoRALinearLayer(cross_attention_dim or v_hidden_size, v_hidden_size, v_rank, network_alpha)
|
| 44 |
+
self.to_out_lora = LoRALinearLayer(out_hidden_size, out_hidden_size, out_rank, network_alpha)
|
| 45 |
+
|
| 46 |
+
def __call__(self, attn: Attention, hidden_states, *args, **kwargs):
|
| 47 |
+
self_cls_name = self.__class__.__name__
|
| 48 |
+
deprecate(
|
| 49 |
+
self_cls_name,
|
| 50 |
+
"0.26.0",
|
| 51 |
+
(
|
| 52 |
+
f"Make sure use {self_cls_name[4:]} instead by setting"
|
| 53 |
+
"LoRA layers to `self.{to_q,to_k,to_v,to_out[0]}.lora_layer` respectively. This will be done automatically when using"
|
| 54 |
+
" `LoraLoaderMixin.load_lora_weights`"
|
| 55 |
+
),
|
| 56 |
+
)
|
| 57 |
+
attn.to_q.lora_layer = self.to_q_lora.to(hidden_states.device)
|
| 58 |
+
attn.to_k.lora_layer = self.to_k_lora.to(hidden_states.device)
|
| 59 |
+
attn.to_v.lora_layer = self.to_v_lora.to(hidden_states.device)
|
| 60 |
+
attn.to_out[0].lora_layer = self.to_out_lora.to(hidden_states.device)
|
| 61 |
+
|
| 62 |
+
attn._modules.pop("processor")
|
| 63 |
+
attn.processor = AttnProcessorCustom(16)
|
| 64 |
+
return attn.processor(attn, hidden_states, *args, **kwargs)
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
class AttnProcessorCustom(AttnProcessor):
|
| 68 |
+
r"""
|
| 69 |
+
Default processor for performing attention-related computations.
|
| 70 |
+
"""
|
| 71 |
+
|
| 72 |
+
def __init__(self, attn_size):
|
| 73 |
+
self.attn_size = attn_size
|
| 74 |
+
|
| 75 |
+
def __call__(
|
| 76 |
+
self,
|
| 77 |
+
attn: Attention,
|
| 78 |
+
hidden_states,
|
| 79 |
+
encoder_hidden_states=None,
|
| 80 |
+
attention_mask=None,
|
| 81 |
+
temb=None,
|
| 82 |
+
scale=1.0,
|
| 83 |
+
):
|
| 84 |
+
residual = hidden_states
|
| 85 |
+
|
| 86 |
+
args = () if USE_PEFT_BACKEND else (scale,)
|
| 87 |
+
|
| 88 |
+
if attn.spatial_norm is not None:
|
| 89 |
+
hidden_states = attn.spatial_norm(hidden_states, temb)
|
| 90 |
+
|
| 91 |
+
input_ndim = hidden_states.ndim
|
| 92 |
+
|
| 93 |
+
if input_ndim == 4:
|
| 94 |
+
batch_size, channel, height, width = hidden_states.shape
|
| 95 |
+
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
|
| 96 |
+
|
| 97 |
+
batch_size, sequence_length, _ = (
|
| 98 |
+
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
| 99 |
+
)
|
| 100 |
+
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
|
| 101 |
+
|
| 102 |
+
if attn.group_norm is not None:
|
| 103 |
+
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
|
| 104 |
+
|
| 105 |
+
query = attn.to_q(hidden_states, *args)
|
| 106 |
+
|
| 107 |
+
if encoder_hidden_states is None:
|
| 108 |
+
encoder_hidden_states = hidden_states
|
| 109 |
+
elif attn.norm_cross:
|
| 110 |
+
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
|
| 111 |
+
|
| 112 |
+
key = attn.to_k(encoder_hidden_states, *args)
|
| 113 |
+
value = attn.to_v(encoder_hidden_states, *args)
|
| 114 |
+
|
| 115 |
+
query = attn.head_to_batch_dim(query)
|
| 116 |
+
key = attn.head_to_batch_dim(key)
|
| 117 |
+
value = attn.head_to_batch_dim(value)
|
| 118 |
+
|
| 119 |
+
attn_size = self.attn_size
|
| 120 |
+
attention_probs = attn.get_attention_scores(query, key, attention_mask)
|
| 121 |
+
if attention_probs.size(2) == 77 and attention_probs.size(1) == (attn_size * attn_size): # (B*Head,HW,L)
|
| 122 |
+
attn_probs_cache = attention_probs.reshape(batch_size, -1, attn_size, attn_size, 77)
|
| 123 |
+
attn.attn_probs = attn_probs_cache
|
| 124 |
+
else:
|
| 125 |
+
attn.attn_probs = None
|
| 126 |
+
|
| 127 |
+
hidden_states = torch.bmm(attention_probs, value)
|
| 128 |
+
hidden_states = attn.batch_to_head_dim(hidden_states)
|
| 129 |
+
|
| 130 |
+
# linear proj
|
| 131 |
+
hidden_states = attn.to_out[0](hidden_states, *args)
|
| 132 |
+
# dropout
|
| 133 |
+
hidden_states = attn.to_out[1](hidden_states)
|
| 134 |
+
|
| 135 |
+
if input_ndim == 4:
|
| 136 |
+
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
|
| 137 |
+
|
| 138 |
+
if attn.residual_connection:
|
| 139 |
+
hidden_states = hidden_states + residual
|
| 140 |
+
|
| 141 |
+
hidden_states = hidden_states / attn.rescale_output_factor
|
| 142 |
+
|
| 143 |
+
return hidden_states
|
dreamcreature/dataset.py
ADDED
|
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import random
|
| 3 |
+
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
from PIL import Image
|
| 7 |
+
from torch.utils.data import Dataset
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
class ImageDataset(Dataset):
|
| 11 |
+
|
| 12 |
+
def __init__(self,
|
| 13 |
+
rootdir,
|
| 14 |
+
filename='train.txt',
|
| 15 |
+
path_prefix='',
|
| 16 |
+
transform=None,
|
| 17 |
+
target_transform=None):
|
| 18 |
+
super().__init__()
|
| 19 |
+
|
| 20 |
+
self.rootdir = rootdir
|
| 21 |
+
self.filename = filename
|
| 22 |
+
self.path_prefix = path_prefix
|
| 23 |
+
|
| 24 |
+
self.image_paths = []
|
| 25 |
+
self.image_labels = []
|
| 26 |
+
|
| 27 |
+
filename = os.path.join(self.rootdir, self.filename)
|
| 28 |
+
|
| 29 |
+
with open(filename, 'r') as f:
|
| 30 |
+
while True:
|
| 31 |
+
lines = f.readline()
|
| 32 |
+
if not lines:
|
| 33 |
+
break
|
| 34 |
+
|
| 35 |
+
lines = lines.strip()
|
| 36 |
+
split_lines = lines.split(' ')
|
| 37 |
+
path_tmp = split_lines[0]
|
| 38 |
+
label_tmp = split_lines[1:]
|
| 39 |
+
self.is_onehot = len(label_tmp) != 1
|
| 40 |
+
if not self.is_onehot:
|
| 41 |
+
label_tmp = label_tmp[0]
|
| 42 |
+
self.image_paths.append(path_tmp)
|
| 43 |
+
self.image_labels.append(label_tmp)
|
| 44 |
+
|
| 45 |
+
self.image_paths = np.array(self.image_paths)
|
| 46 |
+
self.image_labels = np.array(self.image_labels, dtype=np.float32)
|
| 47 |
+
|
| 48 |
+
self.transform = transform
|
| 49 |
+
self.target_transform = target_transform
|
| 50 |
+
|
| 51 |
+
def __getitem__(self, index):
|
| 52 |
+
"""
|
| 53 |
+
Args:
|
| 54 |
+
index (int): Index
|
| 55 |
+
Returns:
|
| 56 |
+
tuple: (image, target) where target is index of the target class.
|
| 57 |
+
"""
|
| 58 |
+
path, target = self.image_paths[index], self.image_labels[index]
|
| 59 |
+
target = torch.tensor(target)
|
| 60 |
+
|
| 61 |
+
img = Image.open(f'{self.path_prefix}{path}').convert('RGB')
|
| 62 |
+
|
| 63 |
+
if self.transform is not None:
|
| 64 |
+
img = self.transform(img)
|
| 65 |
+
|
| 66 |
+
if self.target_transform is not None:
|
| 67 |
+
target = self.target_transform(target)
|
| 68 |
+
|
| 69 |
+
return img, target, index
|
| 70 |
+
|
| 71 |
+
def __len__(self):
|
| 72 |
+
return len(self.image_paths)
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
class DreamCreatureDataset(ImageDataset):
|
| 76 |
+
|
| 77 |
+
def __init__(self,
|
| 78 |
+
rootdir,
|
| 79 |
+
filename='train.txt',
|
| 80 |
+
path_prefix='',
|
| 81 |
+
code_filename='train_caps.txt',
|
| 82 |
+
num_parts=8, num_k_per_part=256, repeat=1,
|
| 83 |
+
use_gt_label=False,
|
| 84 |
+
bg_code=7,
|
| 85 |
+
transform=None,
|
| 86 |
+
target_transform=None):
|
| 87 |
+
super().__init__(rootdir, filename, path_prefix, transform, target_transform)
|
| 88 |
+
|
| 89 |
+
self.image_codes = np.array(open(rootdir + '/' + code_filename).readlines())
|
| 90 |
+
self.num_parts = num_parts
|
| 91 |
+
self.num_k_per_part = num_k_per_part
|
| 92 |
+
self.repeat = repeat
|
| 93 |
+
self.use_gt_label = use_gt_label
|
| 94 |
+
self.bg_code = bg_code
|
| 95 |
+
|
| 96 |
+
def filter_by_class(self, target):
|
| 97 |
+
target_mask = self.image_labels == target
|
| 98 |
+
self.image_paths = self.image_paths[target_mask]
|
| 99 |
+
self.image_codes = self.image_codes[target_mask]
|
| 100 |
+
self.image_labels = self.image_labels[target_mask]
|
| 101 |
+
|
| 102 |
+
def set_max_samples(self, n, seed):
|
| 103 |
+
np.random.seed(seed)
|
| 104 |
+
rand_idx = np.arange(len(self.image_paths))
|
| 105 |
+
np.random.shuffle(rand_idx)
|
| 106 |
+
|
| 107 |
+
self.image_paths = self.image_paths[rand_idx[:n]]
|
| 108 |
+
self.image_codes = self.image_codes[rand_idx[:n]]
|
| 109 |
+
self.image_labels = self.image_labels[rand_idx[:n]]
|
| 110 |
+
|
| 111 |
+
def __len__(self):
|
| 112 |
+
return len(self.image_paths) * self.repeat
|
| 113 |
+
|
| 114 |
+
def __getitem__(self, index):
|
| 115 |
+
"""
|
| 116 |
+
Args:
|
| 117 |
+
index (int): Index
|
| 118 |
+
Returns:
|
| 119 |
+
tuple: (image, target) where target is index of the target class.
|
| 120 |
+
"""
|
| 121 |
+
index = index % len(self.image_paths)
|
| 122 |
+
path, target = self.image_paths[index], self.image_labels[index]
|
| 123 |
+
target = torch.tensor(target)
|
| 124 |
+
|
| 125 |
+
img = Image.open(f'{self.path_prefix}{path}').convert('RGB')
|
| 126 |
+
|
| 127 |
+
cap = self.image_codes[index].strip()
|
| 128 |
+
|
| 129 |
+
if self.transform is not None:
|
| 130 |
+
img = self.transform(img)
|
| 131 |
+
|
| 132 |
+
if self.target_transform is not None:
|
| 133 |
+
target = self.target_transform(target)
|
| 134 |
+
|
| 135 |
+
appeared = []
|
| 136 |
+
|
| 137 |
+
code = torch.ones(self.num_parts) * self.num_k_per_part # represents not exists
|
| 138 |
+
splits = cap.strip().replace('.', '').split(' ')
|
| 139 |
+
for c in splits:
|
| 140 |
+
idx, intval = c.split(':')
|
| 141 |
+
appeared.append(int(idx))
|
| 142 |
+
if self.use_gt_label and self.bg_code != int(idx):
|
| 143 |
+
code[int(idx)] = target
|
| 144 |
+
else:
|
| 145 |
+
code[int(idx)] = int(intval)
|
| 146 |
+
|
| 147 |
+
example = {
|
| 148 |
+
'pixel_values': img,
|
| 149 |
+
'captions': cap,
|
| 150 |
+
'codes': code,
|
| 151 |
+
'labels': target,
|
| 152 |
+
'appeared': appeared
|
| 153 |
+
}
|
| 154 |
+
|
| 155 |
+
return example
|
dreamcreature/dino.py
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch.nn as nn
|
| 2 |
+
from transformers import AutoModel, AutoImageProcessor
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
class DINO(nn.Module):
|
| 6 |
+
NAMES = {
|
| 7 |
+
'dino': 'facebook/dino-vits16',
|
| 8 |
+
'dinov2': 'facebook/dinov2-base'
|
| 9 |
+
}
|
| 10 |
+
|
| 11 |
+
def __init__(self, name='facebook/dinov2-base', **kwargs):
|
| 12 |
+
super().__init__()
|
| 13 |
+
|
| 14 |
+
self.model = AutoModel.from_pretrained(name)
|
| 15 |
+
self.processor = AutoImageProcessor.from_pretrained(name)
|
| 16 |
+
|
| 17 |
+
def forward(self, image):
|
| 18 |
+
vit_output = self.model(image,
|
| 19 |
+
output_hidden_states=True,
|
| 20 |
+
return_dict=True)
|
| 21 |
+
|
| 22 |
+
outputs = {}
|
| 23 |
+
for i in range(1, len(vit_output.hidden_states)):
|
| 24 |
+
outputs[f'block{i}'] = vit_output.hidden_states[i][:, 0] # get cls only
|
| 25 |
+
outputs['feats'] = outputs[f'block{i}']
|
| 26 |
+
return outputs
|
| 27 |
+
|
| 28 |
+
def preprocess(self, image, size=None):
|
| 29 |
+
inputs = self.processor(images=image, return_tensors="pt", size=size)
|
| 30 |
+
return inputs['pixel_values']
|
| 31 |
+
|
| 32 |
+
def get_feat_maps(self, image, index=-1):
|
| 33 |
+
vit_output = self.model(image,
|
| 34 |
+
output_hidden_states=True,
|
| 35 |
+
return_dict=True)
|
| 36 |
+
|
| 37 |
+
last_hidden_states = vit_output.hidden_states[index]
|
| 38 |
+
|
| 39 |
+
B, T, C = last_hidden_states.size()
|
| 40 |
+
HW = int((T - 1) ** 0.5)
|
| 41 |
+
|
| 42 |
+
return last_hidden_states[:, 1:, :].reshape(B, HW, HW, C).permute(0, 3, 1, 2) # (B, C, H, W)
|
dreamcreature/kmeans_segmentation.py
ADDED
|
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
from typing import List
|
| 3 |
+
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
import torchpq
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class KMeansSegmentation:
|
| 10 |
+
FOREGROUND = 'foreground_background'
|
| 11 |
+
COARSE = 'coarse_kmeans'
|
| 12 |
+
FINE = 'fine_kmeans'
|
| 13 |
+
|
| 14 |
+
def __init__(self, path, foreground_idx=0, background_code=7, M=8, K=256):
|
| 15 |
+
if not os.path.exists(path):
|
| 16 |
+
raise FileNotFoundError(f'please train {path}')
|
| 17 |
+
kmeans = torch.load(path)
|
| 18 |
+
|
| 19 |
+
self.foreground_idx = foreground_idx
|
| 20 |
+
self.kmeans = kmeans
|
| 21 |
+
self.background_code = background_code
|
| 22 |
+
self.M = M
|
| 23 |
+
self.K = K
|
| 24 |
+
|
| 25 |
+
self.fg: torchpq.clustering.KMeans = kmeans[KMeansSegmentation.FOREGROUND]
|
| 26 |
+
self.coarse: torchpq.clustering.KMeans = kmeans[KMeansSegmentation.COARSE]
|
| 27 |
+
self.fine: List[torchpq.clustering.KMeans] = kmeans[KMeansSegmentation.FINE]
|
| 28 |
+
|
| 29 |
+
def obtain_fine_feats(self, prompt, filter_idxs=[]):
|
| 30 |
+
if isinstance(prompt, str):
|
| 31 |
+
code = np.zeros((self.M,), dtype=int)
|
| 32 |
+
splits = prompt.strip().split(' ')
|
| 33 |
+
for s in splits:
|
| 34 |
+
m, k = s.split(':')
|
| 35 |
+
code[int(m)] = int(k)
|
| 36 |
+
else:
|
| 37 |
+
code = prompt
|
| 38 |
+
|
| 39 |
+
fine_feats = []
|
| 40 |
+
for m in range(self.M):
|
| 41 |
+
fine_feats.append(self.fine[m].centroids.cpu().t()[code[m]])
|
| 42 |
+
fine_feats = torch.stack(fine_feats, dim=0)
|
| 43 |
+
|
| 44 |
+
if len(filter_idxs) != 0:
|
| 45 |
+
new_fine_feats = []
|
| 46 |
+
|
| 47 |
+
for m in range(self.M):
|
| 48 |
+
if m not in filter_idxs:
|
| 49 |
+
new_fine_feats.append(fine_feats[m])
|
| 50 |
+
|
| 51 |
+
fine_feats = torch.stack(new_fine_feats, dim=0)
|
| 52 |
+
|
| 53 |
+
return fine_feats
|
| 54 |
+
|
| 55 |
+
def get_segmask(self, feat_map, with_appeared_tokens=False):
|
| 56 |
+
N, C, H, W = feat_map.size()
|
| 57 |
+
query = feat_map.cuda().reshape(N, C, H * W).permute(0, 2, 1) # (N, H*W, C)
|
| 58 |
+
|
| 59 |
+
fg_labels = self.fg.predict(query.reshape(N * H * W, C).t().contiguous()) # (N*H*W)
|
| 60 |
+
fg_labels = fg_labels.reshape(N, H * W)
|
| 61 |
+
|
| 62 |
+
fg_idx = self.foreground_idx
|
| 63 |
+
bg_idx = 1 - self.foreground_idx
|
| 64 |
+
|
| 65 |
+
nobg = []
|
| 66 |
+
bgmean = []
|
| 67 |
+
|
| 68 |
+
for i in range(N):
|
| 69 |
+
bgnorm_mean = query[i][fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)
|
| 70 |
+
|
| 71 |
+
if fg_idx == 0:
|
| 72 |
+
bg_mask = fg_labels[i]
|
| 73 |
+
else:
|
| 74 |
+
bg_mask = 1 - fg_labels[i]
|
| 75 |
+
|
| 76 |
+
bg_mask = bg_mask.unsqueeze(1)
|
| 77 |
+
nobg.append(query[i] * (1 - bg_mask) + (-1 * bg_mask))
|
| 78 |
+
bgmean.append(bgnorm_mean)
|
| 79 |
+
|
| 80 |
+
nobg = torch.stack(nobg, dim=0) # (B, H*W, C)
|
| 81 |
+
coarse_labels = self.coarse.predict(nobg.reshape(N * H * W, 768).t().contiguous())
|
| 82 |
+
coarse_labels = coarse_labels.reshape(N, H, W)
|
| 83 |
+
|
| 84 |
+
segmasks = []
|
| 85 |
+
for m in range(self.M):
|
| 86 |
+
mask = (coarse_labels == m).float() # (N, H, W)
|
| 87 |
+
segmasks.append(mask)
|
| 88 |
+
segmasks = torch.stack(segmasks, dim=1) # (N, M, H, W)
|
| 89 |
+
|
| 90 |
+
if with_appeared_tokens:
|
| 91 |
+
appeared_tokens = []
|
| 92 |
+
for i in range(N):
|
| 93 |
+
appeared_tokens.append(torch.unique(coarse_labels[i].reshape(-1)).tolist())
|
| 94 |
+
return segmasks, appeared_tokens
|
| 95 |
+
|
| 96 |
+
return segmasks
|
| 97 |
+
|
| 98 |
+
def predict(self, feat_map, disable=True, filter_idxs=[]):
|
| 99 |
+
# feat_map: (B, C, H, W)
|
| 100 |
+
|
| 101 |
+
N, C, H, W = feat_map.size()
|
| 102 |
+
query = feat_map.reshape(N, C, H * W).permute(0, 2, 1) # (N, H*W, C)
|
| 103 |
+
|
| 104 |
+
fg_labels = self.fg.predict(query.reshape(N * H * W, C).t().contiguous().cuda()).cpu() # (N*H*W)
|
| 105 |
+
fg_labels = fg_labels.reshape(N, H * W)
|
| 106 |
+
|
| 107 |
+
fg_idx = self.foreground_idx
|
| 108 |
+
bg_idx = 1 - self.foreground_idx
|
| 109 |
+
|
| 110 |
+
nobg = []
|
| 111 |
+
bgmean = []
|
| 112 |
+
|
| 113 |
+
for i in range(N):
|
| 114 |
+
bgnorm_mean = query[i][fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)
|
| 115 |
+
|
| 116 |
+
if fg_idx == 0:
|
| 117 |
+
bg_mask = fg_labels[i]
|
| 118 |
+
else:
|
| 119 |
+
bg_mask = 1 - fg_labels[i]
|
| 120 |
+
|
| 121 |
+
bg_mask = bg_mask.unsqueeze(1)
|
| 122 |
+
nobg.append(query[i] * (1 - bg_mask) + (-1 * bg_mask))
|
| 123 |
+
bgmean.append(bgnorm_mean)
|
| 124 |
+
|
| 125 |
+
nobg = torch.stack(nobg, dim=0) # (B, H*W, C)
|
| 126 |
+
bgmean = torch.cat(bgmean, dim=0)
|
| 127 |
+
|
| 128 |
+
coarse_labels = self.coarse.predict(nobg.reshape(N * H * W, 768).t().contiguous().cuda()).cpu()
|
| 129 |
+
coarse_labels = coarse_labels.reshape(N, H * W)
|
| 130 |
+
|
| 131 |
+
from tqdm.auto import tqdm
|
| 132 |
+
|
| 133 |
+
fgmean = []
|
| 134 |
+
M = self.M
|
| 135 |
+
|
| 136 |
+
locs = np.zeros((N, M, 2))
|
| 137 |
+
|
| 138 |
+
for i in tqdm(range(N), disable=disable):
|
| 139 |
+
mean_feats = []
|
| 140 |
+
for m in range(M):
|
| 141 |
+
coarse_mask = coarse_labels[i] == m
|
| 142 |
+
if coarse_mask.sum().item() == 0:
|
| 143 |
+
m_mean_feats = torch.zeros(1, C)
|
| 144 |
+
else:
|
| 145 |
+
locs[i, m] = (coarse_mask.reshape(H, W).nonzero().float().add(0.5).mean(dim=0) / H).cpu().numpy()
|
| 146 |
+
m_mean_feats = query[i][coarse_mask].mean(dim=0, keepdim=True) # (H*W,C) -> (1,C)
|
| 147 |
+
|
| 148 |
+
mean_feats.append(m_mean_feats)
|
| 149 |
+
|
| 150 |
+
mean_feats = torch.cat(mean_feats, dim=0)
|
| 151 |
+
fgmean.append(mean_feats)
|
| 152 |
+
|
| 153 |
+
fgmean = torch.stack(fgmean, dim=0) # (N, M, C)
|
| 154 |
+
final_labels = torch.ones(N, M) * self.K
|
| 155 |
+
|
| 156 |
+
for m in range(M):
|
| 157 |
+
fine_kmeans = self.fine[m]
|
| 158 |
+
|
| 159 |
+
if m == self.background_code:
|
| 160 |
+
fine_labels = fine_kmeans.predict(bgmean.t().contiguous().cuda()).cpu()
|
| 161 |
+
final_labels[:, m] = fine_labels
|
| 162 |
+
else:
|
| 163 |
+
fine_inp = fgmean[:, m].reshape(N, C)
|
| 164 |
+
is_zero = fine_inp.sum(dim=1) == 0
|
| 165 |
+
fine_labels = fine_kmeans.predict(fine_inp.t().contiguous().cuda()).cpu()
|
| 166 |
+
fine_labels[is_zero] = self.K
|
| 167 |
+
|
| 168 |
+
final_labels[:, m] = fine_labels
|
| 169 |
+
|
| 170 |
+
fgmean[:, self.background_code] = bgmean
|
| 171 |
+
fine_prompts = []
|
| 172 |
+
|
| 173 |
+
for fine_label in final_labels:
|
| 174 |
+
prompt_dict = {k: int(v) for k, v in enumerate(list(fine_label))}
|
| 175 |
+
if len(filter_idxs) != 0:
|
| 176 |
+
for i in filter_idxs:
|
| 177 |
+
del prompt_dict[i]
|
| 178 |
+
prompt = ' '.join([f'{k}:{v}' for k, v in prompt_dict.items() if v != self.K])
|
| 179 |
+
fine_prompts.append(prompt)
|
| 180 |
+
|
| 181 |
+
return {
|
| 182 |
+
'features': fgmean,
|
| 183 |
+
'fg_labels': fg_labels,
|
| 184 |
+
'coarse_labels': coarse_labels,
|
| 185 |
+
'fine_labels': final_labels,
|
| 186 |
+
'fine_prompts': fine_prompts,
|
| 187 |
+
'location': locs
|
| 188 |
+
}
|
dreamcreature/loss.py
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn.functional as F
|
| 3 |
+
from diffusers import UNet2DConditionModel
|
| 4 |
+
from diffusers.models.attention_processor import Attention
|
| 5 |
+
|
| 6 |
+
from dreamcreature.dino import DINO
|
| 7 |
+
from dreamcreature.kmeans_segmentation import KMeansSegmentation
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
def dreamcreature_loss(batch,
|
| 11 |
+
unet: UNet2DConditionModel,
|
| 12 |
+
dino: DINO,
|
| 13 |
+
seg: KMeansSegmentation,
|
| 14 |
+
placeholder_token_ids,
|
| 15 |
+
accelerator):
|
| 16 |
+
attn_probs = {}
|
| 17 |
+
|
| 18 |
+
for name, module in unet.named_modules():
|
| 19 |
+
if isinstance(module, Attention) and module.attn_probs is not None:
|
| 20 |
+
a = module.attn_probs.mean(dim=1) # (B,Head,H,W,77) -> (B,H,W,77)
|
| 21 |
+
attn_probs[name] = a
|
| 22 |
+
|
| 23 |
+
avg_attn_map = []
|
| 24 |
+
for name in attn_probs:
|
| 25 |
+
avg_attn_map.append(attn_probs[name])
|
| 26 |
+
|
| 27 |
+
avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0) # (L,B,H,W,77) -> (B,H,W,77)
|
| 28 |
+
B, H, W, seq_length = avg_attn_map.size()
|
| 29 |
+
located_attn_map = []
|
| 30 |
+
|
| 31 |
+
# locate the attn map
|
| 32 |
+
for i, placeholder_token_id in enumerate(placeholder_token_ids):
|
| 33 |
+
for bi in range(B):
|
| 34 |
+
if "input_ids" in batch:
|
| 35 |
+
learnable_idx = (batch["input_ids"][bi] == placeholder_token_id).nonzero(as_tuple=True)[0]
|
| 36 |
+
else:
|
| 37 |
+
learnable_idx = (batch["input_ids_one"][bi] == placeholder_token_id).nonzero(as_tuple=True)[0]
|
| 38 |
+
|
| 39 |
+
if len(learnable_idx) != 0: # only assign if found
|
| 40 |
+
if len(learnable_idx) == 1:
|
| 41 |
+
offset_learnable_idx = learnable_idx
|
| 42 |
+
else: # if there is two and above.
|
| 43 |
+
raise NotImplementedError
|
| 44 |
+
|
| 45 |
+
located_map = avg_attn_map[bi, :, :, offset_learnable_idx]
|
| 46 |
+
located_attn_map.append(located_map)
|
| 47 |
+
else:
|
| 48 |
+
located_attn_map.append(torch.zeros(H, W, 1).to(accelerator.device))
|
| 49 |
+
|
| 50 |
+
M = len(placeholder_token_ids)
|
| 51 |
+
located_attn_map = torch.stack(located_attn_map, dim=0).reshape(M, B, H, W).transpose(0, 1) # (B, M, 16, 16)
|
| 52 |
+
|
| 53 |
+
raw_images = batch['raw_images']
|
| 54 |
+
dino_input = dino.preprocess(raw_images, size=224)
|
| 55 |
+
with torch.no_grad():
|
| 56 |
+
dino_ft = dino.get_feat_maps(dino_input)
|
| 57 |
+
segmasks, appeared_tokens = seg.get_segmask(dino_ft, True) # (B, M, H, W)
|
| 58 |
+
segmasks = segmasks.to(located_attn_map.dtype)
|
| 59 |
+
if H != 16: # for res 1024
|
| 60 |
+
segmasks = F.interpolate(segmasks, (H, W), mode='nearest')
|
| 61 |
+
|
| 62 |
+
masks = []
|
| 63 |
+
for i, appeared in enumerate(appeared_tokens):
|
| 64 |
+
mask = (segmasks[i, appeared].sum(dim=0) > 0).float() # (A, H, W) -> (H, W)
|
| 65 |
+
masks.append(mask)
|
| 66 |
+
masks = torch.stack(masks, dim=0) # (B, H, W)
|
| 67 |
+
batch['masks'] = masks
|
| 68 |
+
|
| 69 |
+
norm_map = located_attn_map / located_attn_map.sum(dim=1, keepdim=True).clamp(min=1e-6)
|
| 70 |
+
# if norm_map is assigned manually, means the sub-concept token is not found, hence no gradient will be backprop
|
| 71 |
+
attn_loss = F.binary_cross_entropy(norm_map.clamp(min=0, max=1),
|
| 72 |
+
segmasks.clamp(min=0, max=1))
|
| 73 |
+
return attn_loss, located_attn_map.detach().max()
|
dreamcreature/mapper.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Optional
|
| 2 |
+
|
| 3 |
+
import torch
|
| 4 |
+
from torch import nn
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
class TokenMapper(nn.Module):
|
| 8 |
+
def __init__(self,
|
| 9 |
+
num_parts,
|
| 10 |
+
num_k_per_part,
|
| 11 |
+
out_dims,
|
| 12 |
+
projection_nlayers=1,
|
| 13 |
+
projection_activation=nn.ReLU(),
|
| 14 |
+
with_pe=True):
|
| 15 |
+
super().__init__()
|
| 16 |
+
|
| 17 |
+
self.num_parts = num_parts
|
| 18 |
+
self.num_k_per_part = num_k_per_part
|
| 19 |
+
self.with_pe = with_pe
|
| 20 |
+
self.out_dims = out_dims
|
| 21 |
+
|
| 22 |
+
self.embedding = nn.Embedding((self.num_k_per_part + 1) * num_parts, out_dims)
|
| 23 |
+
if with_pe:
|
| 24 |
+
self.pe = nn.Parameter(torch.randn(num_parts, out_dims))
|
| 25 |
+
else:
|
| 26 |
+
self.register_buffer('pe', torch.zeros(num_parts, out_dims))
|
| 27 |
+
|
| 28 |
+
if projection_nlayers == 0:
|
| 29 |
+
self.projection = nn.Identity()
|
| 30 |
+
else:
|
| 31 |
+
projections = []
|
| 32 |
+
for i in range(projection_nlayers - 1):
|
| 33 |
+
projections.append(nn.Linear(out_dims, out_dims))
|
| 34 |
+
projections.append(projection_activation)
|
| 35 |
+
|
| 36 |
+
projections.append(nn.Linear(out_dims, out_dims))
|
| 37 |
+
self.projection = nn.Sequential(*projections)
|
| 38 |
+
|
| 39 |
+
def get_all_embeddings(self, no_projection=False, no_pe=False):
|
| 40 |
+
idx = torch.arange(self.num_parts * (self.num_k_per_part + 1)).long().to(self.embedding.weight.device)
|
| 41 |
+
idx = idx.reshape(self.num_parts, self.num_k_per_part + 1)
|
| 42 |
+
emb = self.embedding(idx) # (K, N, d)
|
| 43 |
+
|
| 44 |
+
if not no_pe:
|
| 45 |
+
emb_pe = emb + self.pe.unsqueeze(1)
|
| 46 |
+
else:
|
| 47 |
+
emb_pe = emb
|
| 48 |
+
|
| 49 |
+
if not no_projection:
|
| 50 |
+
projected = self.projection(emb_pe)
|
| 51 |
+
else:
|
| 52 |
+
projected = emb_pe
|
| 53 |
+
|
| 54 |
+
return projected
|
| 55 |
+
|
| 56 |
+
def forward(self, hashes, index: Optional[torch.Tensor] = None):
|
| 57 |
+
B = hashes.size(0)
|
| 58 |
+
|
| 59 |
+
# 0, 257, 514, ...
|
| 60 |
+
if index is None:
|
| 61 |
+
offset = torch.arange(self.num_parts, device=hashes.device) * (self.num_k_per_part + 1)
|
| 62 |
+
hashes = self.embedding(hashes.long() + offset.reshape(1, -1)) # (B, N, d)
|
| 63 |
+
else:
|
| 64 |
+
offset = index.reshape(-1) * (self.num_k_per_part + 1)
|
| 65 |
+
hashes = self.embedding(hashes.long() + offset.reshape(B, -1).long()) # (B, N, d)
|
| 66 |
+
|
| 67 |
+
if index is not None:
|
| 68 |
+
pe = self.pe[index.reshape(-1)] # index must be equal size
|
| 69 |
+
pe = pe.reshape(B, -1, self.out_dims)
|
| 70 |
+
hashes = hashes + pe
|
| 71 |
+
else:
|
| 72 |
+
hashes = hashes + self.pe.unsqueeze(0).repeat(B, 1, 1)
|
| 73 |
+
projected = self.projection(hashes)
|
| 74 |
+
|
| 75 |
+
return projected
|
dreamcreature/pipeline.py
ADDED
|
@@ -0,0 +1,771 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import re
|
| 3 |
+
|
| 4 |
+
from diffusers.loaders import AttnProcsLayers
|
| 5 |
+
from diffusers.models.attention_processor import Attention
|
| 6 |
+
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import *
|
| 7 |
+
from omegaconf import OmegaConf
|
| 8 |
+
|
| 9 |
+
from dreamcreature.attn_processor import LoRAAttnProcessorCustom
|
| 10 |
+
from dreamcreature.mapper import TokenMapper
|
| 11 |
+
from dreamcreature.text_encoder import CustomCLIPTextModel
|
| 12 |
+
from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
|
| 13 |
+
from utils import add_tokens, get_attn_processors
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def setup_attn_processor(unet, **kwargs):
|
| 17 |
+
lora_attn_procs = {}
|
| 18 |
+
for name in unet.attn_processors.keys():
|
| 19 |
+
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
|
| 20 |
+
if name.startswith("mid_block"):
|
| 21 |
+
hidden_size = unet.config.block_out_channels[-1]
|
| 22 |
+
elif name.startswith("up_blocks"):
|
| 23 |
+
block_id = int(name[len("up_blocks.")])
|
| 24 |
+
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
|
| 25 |
+
elif name.startswith("down_blocks"):
|
| 26 |
+
block_id = int(name[len("down_blocks.")])
|
| 27 |
+
hidden_size = unet.config.block_out_channels[block_id]
|
| 28 |
+
|
| 29 |
+
lora_attn_procs[name] = LoRAAttnProcessorCustom(
|
| 30 |
+
hidden_size=hidden_size,
|
| 31 |
+
cross_attention_dim=cross_attention_dim,
|
| 32 |
+
rank=kwargs['rank'],
|
| 33 |
+
)
|
| 34 |
+
|
| 35 |
+
unet.set_attn_processor(lora_attn_procs)
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def load_attn_processor(unet, filename):
|
| 39 |
+
lora_layers = AttnProcsLayers(get_attn_processors(unet))
|
| 40 |
+
lora_layers.load_state_dict(torch.load(filename))
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def convert_prompt_re(prompt: str):
|
| 44 |
+
pattern = r"<(\d+):(\d+)>"
|
| 45 |
+
result = prompt
|
| 46 |
+
offset = 0
|
| 47 |
+
|
| 48 |
+
ints = []
|
| 49 |
+
parts_i = []
|
| 50 |
+
|
| 51 |
+
for match in re.finditer(pattern, prompt):
|
| 52 |
+
i = int(match.group(1))
|
| 53 |
+
b = int(match.group(2))
|
| 54 |
+
|
| 55 |
+
replacement = f"<part>_{i}"
|
| 56 |
+
start, end = match.span()
|
| 57 |
+
|
| 58 |
+
# Adjust the start and end positions based on the offset from previous replacements
|
| 59 |
+
start += offset
|
| 60 |
+
end += offset
|
| 61 |
+
|
| 62 |
+
# Replace the matched text with the replacement
|
| 63 |
+
result = result[:start] + replacement + result[end:]
|
| 64 |
+
|
| 65 |
+
# Update the offset for the next replacement
|
| 66 |
+
offset += len(replacement) - (end - start)
|
| 67 |
+
|
| 68 |
+
parts_i.append(i)
|
| 69 |
+
ints.append(b)
|
| 70 |
+
|
| 71 |
+
result = result.strip()
|
| 72 |
+
|
| 73 |
+
if len(ints) == 0:
|
| 74 |
+
return result, None, None
|
| 75 |
+
|
| 76 |
+
ints = torch.tensor(ints) # (nparts,)
|
| 77 |
+
return result, ints, parts_i
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def convert_prompt(prompt: str, replace_token: bool = False, v='v1'):
|
| 81 |
+
r"""
|
| 82 |
+
Parameters:
|
| 83 |
+
prompt (`str`):
|
| 84 |
+
The prompt to guide the image generation.
|
| 85 |
+
|
| 86 |
+
Returns:
|
| 87 |
+
`str`: The converted prompt
|
| 88 |
+
"""
|
| 89 |
+
if v == 're':
|
| 90 |
+
return convert_prompt_re(prompt)
|
| 91 |
+
|
| 92 |
+
if ':' not in prompt:
|
| 93 |
+
return prompt, None, None
|
| 94 |
+
|
| 95 |
+
splits = prompt.replace('.', '').strip().split(' ')
|
| 96 |
+
# v1: a photo of a 0:1 1:24 ...
|
| 97 |
+
# v2: a photo of a <0:1> <1:24> ...
|
| 98 |
+
ints = []
|
| 99 |
+
noncode_start = ''
|
| 100 |
+
noncode_end = ''
|
| 101 |
+
parts = ''
|
| 102 |
+
parts_i = []
|
| 103 |
+
split_tokens = []
|
| 104 |
+
for b in splits:
|
| 105 |
+
if ':' not in b:
|
| 106 |
+
split_tokens.append(b)
|
| 107 |
+
continue
|
| 108 |
+
|
| 109 |
+
if v == 'v1':
|
| 110 |
+
i, b = b.strip().split(':')
|
| 111 |
+
has_comma = ',' in b
|
| 112 |
+
if has_comma:
|
| 113 |
+
b = b[:-1]
|
| 114 |
+
intb = int(b)
|
| 115 |
+
parts += f'<part>_{i} '
|
| 116 |
+
split_tokens.append(f'<part>_{i}')
|
| 117 |
+
if has_comma:
|
| 118 |
+
split_tokens.append(',')
|
| 119 |
+
else:
|
| 120 |
+
if b[0] == '<':
|
| 121 |
+
if '>' not in b: # no closing >, ignore
|
| 122 |
+
split_tokens.append(b)
|
| 123 |
+
continue
|
| 124 |
+
|
| 125 |
+
i, b = b[1:].strip().split(':')
|
| 126 |
+
token_to_add = ''
|
| 127 |
+
if b[-1] in [',', '.']:
|
| 128 |
+
token_to_add = b[-1]
|
| 129 |
+
b = b[:-1]
|
| 130 |
+
|
| 131 |
+
if b[-1] == '>':
|
| 132 |
+
b = b[:-1]
|
| 133 |
+
else: # not >, just search for the first >
|
| 134 |
+
for ci, char in enumerate(b):
|
| 135 |
+
if char == '>':
|
| 136 |
+
token_to_add = b[ci + 1:] + token_to_add
|
| 137 |
+
b = b[:ci] # skip >
|
| 138 |
+
break
|
| 139 |
+
else: # has : but not start with <
|
| 140 |
+
split_tokens.append(b)
|
| 141 |
+
continue
|
| 142 |
+
|
| 143 |
+
intb = abs(int(b)) # just force negative one to positive
|
| 144 |
+
|
| 145 |
+
parts += f'<part>_{i} '
|
| 146 |
+
split_tokens.append(f'<part>_{i}')
|
| 147 |
+
if len(token_to_add) != 0:
|
| 148 |
+
split_tokens.append(token_to_add)
|
| 149 |
+
|
| 150 |
+
try:
|
| 151 |
+
int(i)
|
| 152 |
+
except:
|
| 153 |
+
raise ValueError(f'cannot cast `part` properly, please make sure input is correct')
|
| 154 |
+
|
| 155 |
+
parts_i.append(int(i))
|
| 156 |
+
ints.append(intb)
|
| 157 |
+
|
| 158 |
+
ints = torch.tensor(ints) # (nparts,)
|
| 159 |
+
|
| 160 |
+
if replace_token:
|
| 161 |
+
new_caption = f'{noncode_start.strip()} <part> {noncode_end.strip()}'
|
| 162 |
+
else:
|
| 163 |
+
new_caption = ' '.join(split_tokens)
|
| 164 |
+
|
| 165 |
+
new_caption = new_caption.strip()
|
| 166 |
+
|
| 167 |
+
return new_caption, ints, parts_i
|
| 168 |
+
|
| 169 |
+
|
| 170 |
+
class DreamCreatureSDPipeline(StableDiffusionPipeline):
|
| 171 |
+
def _maybe_convert_prompt(self, prompt: str, tokenizer: MultiTokenCLIPTokenizer):
|
| 172 |
+
r"""
|
| 173 |
+
Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
|
| 174 |
+
to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
|
| 175 |
+
is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
|
| 176 |
+
inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
|
| 177 |
+
|
| 178 |
+
Parameters:
|
| 179 |
+
prompt (`str`):
|
| 180 |
+
The prompt to guide the image generation.
|
| 181 |
+
tokenizer (`PreTrainedTokenizer`):
|
| 182 |
+
The tokenizer responsible for encoding the prompt into input tokens.
|
| 183 |
+
|
| 184 |
+
Returns:
|
| 185 |
+
`str`: The converted prompt
|
| 186 |
+
"""
|
| 187 |
+
if hasattr(self, 'replace_token'):
|
| 188 |
+
replace_token = self.replace_token
|
| 189 |
+
else:
|
| 190 |
+
replace_token = True
|
| 191 |
+
|
| 192 |
+
if hasattr(self, 'v'):
|
| 193 |
+
v = self.v
|
| 194 |
+
else:
|
| 195 |
+
v = 'v1'
|
| 196 |
+
|
| 197 |
+
new_caption, code, parts_i = convert_prompt(prompt, replace_token, v)
|
| 198 |
+
if hasattr(self, 'num_k_per_part'):
|
| 199 |
+
if code is not None and any(code >= self.num_k_per_part):
|
| 200 |
+
raise ValueError(f'`id` cannot more than {self.num_k_per_part}')
|
| 201 |
+
|
| 202 |
+
if hasattr(self, 'verbose') and self.verbose:
|
| 203 |
+
print(new_caption)
|
| 204 |
+
|
| 205 |
+
return new_caption, code, parts_i
|
| 206 |
+
|
| 207 |
+
def compute_prompt_embeddings(self, prompts, device):
|
| 208 |
+
# textual inversion: procecss multi-vector tokens if necessary
|
| 209 |
+
if not isinstance(prompts, List):
|
| 210 |
+
prompts = [prompts]
|
| 211 |
+
|
| 212 |
+
prompt_embeds_concat = []
|
| 213 |
+
for prompt in prompts:
|
| 214 |
+
prompt, code, parts_i = self.maybe_convert_prompt(prompt, self.tokenizer)
|
| 215 |
+
|
| 216 |
+
if hasattr(self, 'replace_token'):
|
| 217 |
+
replace_token = self.replace_token
|
| 218 |
+
else:
|
| 219 |
+
replace_token = True
|
| 220 |
+
|
| 221 |
+
text_inputs = self.tokenizer(
|
| 222 |
+
prompt,
|
| 223 |
+
replace_token=replace_token,
|
| 224 |
+
padding="max_length",
|
| 225 |
+
max_length=self.tokenizer.model_max_length,
|
| 226 |
+
truncation=True,
|
| 227 |
+
return_tensors="pt",
|
| 228 |
+
)
|
| 229 |
+
text_input_ids = text_inputs.input_ids
|
| 230 |
+
if hasattr(self, 'verbose') and self.verbose:
|
| 231 |
+
print(text_input_ids)
|
| 232 |
+
|
| 233 |
+
untruncated_ids = self.tokenizer(prompt,
|
| 234 |
+
replace_token=replace_token,
|
| 235 |
+
padding="longest",
|
| 236 |
+
return_tensors="pt").input_ids
|
| 237 |
+
|
| 238 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
| 239 |
+
text_input_ids, untruncated_ids
|
| 240 |
+
):
|
| 241 |
+
removed_text = self.tokenizer.batch_decode(
|
| 242 |
+
untruncated_ids[:, self.tokenizer.model_max_length - 1: -1]
|
| 243 |
+
)
|
| 244 |
+
logger.warning(
|
| 245 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
| 246 |
+
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
|
| 247 |
+
)
|
| 248 |
+
|
| 249 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
| 250 |
+
attention_mask = text_inputs.attention_mask.to(device)
|
| 251 |
+
else:
|
| 252 |
+
attention_mask = None
|
| 253 |
+
|
| 254 |
+
if code is None:
|
| 255 |
+
modified_hs = None
|
| 256 |
+
else:
|
| 257 |
+
placeholder_token_ids = self.placeholder_token_ids
|
| 258 |
+
placeholder_token_ids = [placeholder_token_ids[i] for i in parts_i] # follow the order of prompt's i
|
| 259 |
+
mapper_outputs = self.simple_mapper(code.unsqueeze(0).to(device), torch.tensor(parts_i).to(device))
|
| 260 |
+
modified_hs = self.text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids.to(device),
|
| 261 |
+
None,
|
| 262 |
+
mapper_outputs,
|
| 263 |
+
placeholder_token_ids)
|
| 264 |
+
|
| 265 |
+
prompt_embeds = self.text_encoder(
|
| 266 |
+
text_input_ids.to(device),
|
| 267 |
+
attention_mask=attention_mask,
|
| 268 |
+
hidden_states=modified_hs
|
| 269 |
+
)
|
| 270 |
+
prompt_embeds = prompt_embeds[0]
|
| 271 |
+
prompt_embeds_concat.append(prompt_embeds)
|
| 272 |
+
|
| 273 |
+
if len(prompt_embeds_concat) == 1:
|
| 274 |
+
return prompt_embeds_concat[0]
|
| 275 |
+
else:
|
| 276 |
+
return torch.cat(prompt_embeds_concat, dim=0)
|
| 277 |
+
|
| 278 |
+
def encode_prompt(
|
| 279 |
+
self,
|
| 280 |
+
prompt,
|
| 281 |
+
device,
|
| 282 |
+
num_images_per_prompt,
|
| 283 |
+
do_classifier_free_guidance,
|
| 284 |
+
negative_prompt=None,
|
| 285 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 286 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 287 |
+
lora_scale: Optional[float] = None,
|
| 288 |
+
):
|
| 289 |
+
r"""
|
| 290 |
+
Encodes the prompt into text encoder hidden states.
|
| 291 |
+
|
| 292 |
+
Args:
|
| 293 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 294 |
+
prompt to be encoded
|
| 295 |
+
device: (`torch.device`):
|
| 296 |
+
torch device
|
| 297 |
+
num_images_per_prompt (`int`):
|
| 298 |
+
number of images that should be generated per prompt
|
| 299 |
+
do_classifier_free_guidance (`bool`):
|
| 300 |
+
whether to use classifier free guidance or not
|
| 301 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 302 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 303 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
| 304 |
+
less than `1`).
|
| 305 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 306 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
| 307 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
| 308 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 309 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 310 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
| 311 |
+
argument.
|
| 312 |
+
lora_scale (`float`, *optional*):
|
| 313 |
+
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
| 314 |
+
"""
|
| 315 |
+
# set lora scale so that monkey patched LoRA
|
| 316 |
+
# function of text encoder can correctly access it
|
| 317 |
+
if lora_scale is not None and isinstance(self, LoraLoaderMixin):
|
| 318 |
+
self._lora_scale = lora_scale
|
| 319 |
+
|
| 320 |
+
if prompt is not None and isinstance(prompt, str):
|
| 321 |
+
batch_size = 1
|
| 322 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 323 |
+
batch_size = len(prompt)
|
| 324 |
+
else:
|
| 325 |
+
batch_size = prompt_embeds.shape[0]
|
| 326 |
+
|
| 327 |
+
if prompt_embeds is None:
|
| 328 |
+
prompt_embeds = self.compute_prompt_embeddings(prompt, device)
|
| 329 |
+
|
| 330 |
+
# if self.text_encoder is not None:
|
| 331 |
+
# prompt_embeds_dtype = self.text_encoder.dtype
|
| 332 |
+
# elif self.unet is not None:
|
| 333 |
+
# prompt_embeds_dtype = self.unet.dtype
|
| 334 |
+
# else:
|
| 335 |
+
# prompt_embeds_dtype = prompt_embeds.dtype
|
| 336 |
+
|
| 337 |
+
prompt_embeds_dtype = self.unet.dtype # should be unet only because this is unet's condition input
|
| 338 |
+
prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
|
| 339 |
+
|
| 340 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 341 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
| 342 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 343 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
| 344 |
+
|
| 345 |
+
# get unconditional embeddings for classifier free guidance
|
| 346 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None:
|
| 347 |
+
uncond_tokens: List[str]
|
| 348 |
+
if negative_prompt is None:
|
| 349 |
+
uncond_tokens = [""] * batch_size
|
| 350 |
+
elif prompt is not None and type(prompt) is not type(negative_prompt):
|
| 351 |
+
raise TypeError(
|
| 352 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
| 353 |
+
f" {type(prompt)}."
|
| 354 |
+
)
|
| 355 |
+
elif isinstance(negative_prompt, str):
|
| 356 |
+
uncond_tokens = [negative_prompt] * batch_size
|
| 357 |
+
elif batch_size != len(negative_prompt):
|
| 358 |
+
raise ValueError(
|
| 359 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
| 360 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
| 361 |
+
" the batch size of `prompt`."
|
| 362 |
+
)
|
| 363 |
+
else:
|
| 364 |
+
uncond_tokens = negative_prompt
|
| 365 |
+
|
| 366 |
+
negative_prompt_embeds = []
|
| 367 |
+
for u_tokens in uncond_tokens:
|
| 368 |
+
negative_prompt_embeds.append(self.compute_prompt_embeddings(u_tokens, device))
|
| 369 |
+
negative_prompt_embeds = torch.cat(negative_prompt_embeds, dim=0)
|
| 370 |
+
|
| 371 |
+
# if isinstance(self, TextualInversionLoaderMixin):
|
| 372 |
+
# uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
|
| 373 |
+
|
| 374 |
+
# max_length = prompt_embeds.shape[1]
|
| 375 |
+
# uncond_input = self.tokenizer(
|
| 376 |
+
# uncond_tokens,
|
| 377 |
+
# padding="max_length",
|
| 378 |
+
# max_length=max_length,
|
| 379 |
+
# truncation=True,
|
| 380 |
+
# return_tensors="pt",
|
| 381 |
+
# )
|
| 382 |
+
#
|
| 383 |
+
# if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
| 384 |
+
# attention_mask = uncond_input.attention_mask.to(device)
|
| 385 |
+
# else:
|
| 386 |
+
# attention_mask = None
|
| 387 |
+
#
|
| 388 |
+
# negative_prompt_embeds = self.text_encoder(
|
| 389 |
+
# uncond_input.input_ids.to(device),
|
| 390 |
+
# attention_mask=attention_mask,
|
| 391 |
+
# )
|
| 392 |
+
# negative_prompt_embeds = negative_prompt_embeds[0]
|
| 393 |
+
|
| 394 |
+
if do_classifier_free_guidance:
|
| 395 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
| 396 |
+
seq_len = negative_prompt_embeds.shape[1]
|
| 397 |
+
|
| 398 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
|
| 399 |
+
|
| 400 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 401 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
| 402 |
+
|
| 403 |
+
return prompt_embeds, negative_prompt_embeds
|
| 404 |
+
|
| 405 |
+
@torch.no_grad()
|
| 406 |
+
def obtain_attention_map(self, image, prompt, timesteps):
|
| 407 |
+
prompt, codes, index = self.maybe_convert_prompt(prompt, self.tokenizer)
|
| 408 |
+
|
| 409 |
+
if hasattr(self, 'replace_token'):
|
| 410 |
+
replace_token = self.replace_token
|
| 411 |
+
else:
|
| 412 |
+
replace_token = True
|
| 413 |
+
|
| 414 |
+
text_inputs = self.tokenizer(
|
| 415 |
+
prompt,
|
| 416 |
+
replace_token=replace_token,
|
| 417 |
+
padding="max_length",
|
| 418 |
+
max_length=self.tokenizer.model_max_length,
|
| 419 |
+
truncation=True,
|
| 420 |
+
return_tensors="pt",
|
| 421 |
+
)
|
| 422 |
+
input_ids = text_inputs.input_ids
|
| 423 |
+
|
| 424 |
+
placeholder_token_ids = self.placeholder_token_ids
|
| 425 |
+
placeholder_token_ids = [placeholder_token_ids[i] for i in index]
|
| 426 |
+
|
| 427 |
+
# forward an image, denoise it and obtain the attention map
|
| 428 |
+
device = self._execution_device
|
| 429 |
+
|
| 430 |
+
latents = self.vae.encode(image.to(device, dtype=self.weight_dtype)).latent_dist.sample()
|
| 431 |
+
latents = latents * self.vae.config.scaling_factor
|
| 432 |
+
|
| 433 |
+
# bsz = latents.shape[0]
|
| 434 |
+
# Sample a random timestep for each image
|
| 435 |
+
# timesteps = torch.randint(0, self.noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
|
| 436 |
+
timesteps = timesteps.long().to(latents.device)
|
| 437 |
+
|
| 438 |
+
# Add noise to the latents according to the noise magnitude at each timestep
|
| 439 |
+
# (this is the forward diffusion process)
|
| 440 |
+
noise = torch.randn_like(latents)
|
| 441 |
+
noisy_latents = self.noise_scheduler.add_noise(latents, noise, timesteps)
|
| 442 |
+
|
| 443 |
+
mapper_outputs = self.simple_mapper(codes.unsqueeze(0).to(device), torch.tensor(index).to(device))
|
| 444 |
+
# print(mapper_outputs.size(), batch["input_ids"].size())
|
| 445 |
+
modified_hs = self.text_encoder.text_model.forward_embeddings_with_mapper(input_ids.to(device),
|
| 446 |
+
None,
|
| 447 |
+
mapper_outputs,
|
| 448 |
+
placeholder_token_ids)
|
| 449 |
+
encoder_hidden_states = self.text_encoder(input_ids, hidden_states=modified_hs)[0]
|
| 450 |
+
model_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states.to(dtype=self.weight_dtype)).sample
|
| 451 |
+
|
| 452 |
+
attn_probs = {}
|
| 453 |
+
|
| 454 |
+
for name, module in self.unet.named_modules():
|
| 455 |
+
if isinstance(module, Attention) and module.attn_probs is not None:
|
| 456 |
+
a = module.attn_probs[0].mean(dim=0) # (2,Head,H,W,77)->(H,W,77)
|
| 457 |
+
attn_probs[name] = a
|
| 458 |
+
|
| 459 |
+
avg_attn_map = []
|
| 460 |
+
for name in attn_probs:
|
| 461 |
+
avg_attn_map.append(attn_probs[name])
|
| 462 |
+
avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0) # (5,B,H,W,77) -> (B,H,W,77)
|
| 463 |
+
|
| 464 |
+
return attn_probs, avg_attn_map, input_ids
|
| 465 |
+
|
| 466 |
+
@torch.no_grad()
|
| 467 |
+
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
| 468 |
+
def __call__(
|
| 469 |
+
self,
|
| 470 |
+
prompt: Union[str, List[str]] = None,
|
| 471 |
+
height: Optional[int] = None,
|
| 472 |
+
width: Optional[int] = None,
|
| 473 |
+
num_inference_steps: int = 50,
|
| 474 |
+
guidance_scale: float = 7.5,
|
| 475 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 476 |
+
num_images_per_prompt: Optional[int] = 1,
|
| 477 |
+
eta: float = 0.0,
|
| 478 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
| 479 |
+
latents: Optional[torch.FloatTensor] = None,
|
| 480 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 481 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 482 |
+
output_type: Optional[str] = "pil",
|
| 483 |
+
return_dict: bool = True,
|
| 484 |
+
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
| 485 |
+
callback_steps: int = 1,
|
| 486 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 487 |
+
guidance_rescale: float = 0.0,
|
| 488 |
+
get_attention_map: bool = False
|
| 489 |
+
):
|
| 490 |
+
r"""
|
| 491 |
+
The call function to the pipeline for generation.
|
| 492 |
+
|
| 493 |
+
Args:
|
| 494 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 495 |
+
The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
|
| 496 |
+
height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
|
| 497 |
+
The height in pixels of the generated image.
|
| 498 |
+
width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
|
| 499 |
+
The width in pixels of the generated image.
|
| 500 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
| 501 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
| 502 |
+
expense of slower inference.
|
| 503 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
| 504 |
+
A higher guidance scale value encourages the model to generate images closely linked to the text
|
| 505 |
+
`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
|
| 506 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 507 |
+
The prompt or prompts to guide what to not include in image generation. If not defined, you need to
|
| 508 |
+
pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
|
| 509 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
| 510 |
+
The number of images to generate per prompt.
|
| 511 |
+
eta (`float`, *optional*, defaults to 0.0):
|
| 512 |
+
Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
|
| 513 |
+
to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
|
| 514 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
| 515 |
+
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
|
| 516 |
+
generation deterministic.
|
| 517 |
+
latents (`torch.FloatTensor`, *optional*):
|
| 518 |
+
Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
|
| 519 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
| 520 |
+
tensor is generated by sampling using the supplied random `generator`.
|
| 521 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 522 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
|
| 523 |
+
provided, text embeddings are generated from the `prompt` input argument.
|
| 524 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 525 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
|
| 526 |
+
not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
|
| 527 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
| 528 |
+
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
|
| 529 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 530 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
| 531 |
+
plain tuple.
|
| 532 |
+
callback (`Callable`, *optional*):
|
| 533 |
+
A function that calls every `callback_steps` steps during inference. The function is called with the
|
| 534 |
+
following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
| 535 |
+
callback_steps (`int`, *optional*, defaults to 1):
|
| 536 |
+
The frequency at which the `callback` function is called. If not specified, the callback is called at
|
| 537 |
+
every step.
|
| 538 |
+
cross_attention_kwargs (`dict`, *optional*):
|
| 539 |
+
A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
|
| 540 |
+
[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
| 541 |
+
guidance_rescale (`float`, *optional*, defaults to 0.7):
|
| 542 |
+
Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
|
| 543 |
+
Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
|
| 544 |
+
using zero terminal SNR.
|
| 545 |
+
|
| 546 |
+
Examples:
|
| 547 |
+
|
| 548 |
+
Returns:
|
| 549 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
| 550 |
+
If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
|
| 551 |
+
otherwise a `tuple` is returned where the first element is a list with the generated images and the
|
| 552 |
+
second element is a list of `bool`s indicating whether the corresponding generated image contains
|
| 553 |
+
"not-safe-for-work" (nsfw) content.
|
| 554 |
+
"""
|
| 555 |
+
# 0. Default height and width to unet
|
| 556 |
+
height = height or self.unet.config.sample_size * self.vae_scale_factor
|
| 557 |
+
width = width or self.unet.config.sample_size * self.vae_scale_factor
|
| 558 |
+
|
| 559 |
+
# 1. Check inputs. Raise error if not correct
|
| 560 |
+
self.check_inputs(
|
| 561 |
+
prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
|
| 562 |
+
)
|
| 563 |
+
|
| 564 |
+
# 2. Define call parameters
|
| 565 |
+
if prompt is not None and isinstance(prompt, str):
|
| 566 |
+
batch_size = 1
|
| 567 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 568 |
+
batch_size = len(prompt)
|
| 569 |
+
else:
|
| 570 |
+
batch_size = prompt_embeds.shape[0]
|
| 571 |
+
|
| 572 |
+
device = self._execution_device
|
| 573 |
+
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
| 574 |
+
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
| 575 |
+
# corresponds to doing no classifier free guidance.
|
| 576 |
+
do_classifier_free_guidance = guidance_scale > 1.0
|
| 577 |
+
|
| 578 |
+
# 3. Encode input prompt
|
| 579 |
+
text_encoder_lora_scale = (
|
| 580 |
+
cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
|
| 581 |
+
)
|
| 582 |
+
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
|
| 583 |
+
prompt,
|
| 584 |
+
device,
|
| 585 |
+
num_images_per_prompt,
|
| 586 |
+
do_classifier_free_guidance,
|
| 587 |
+
negative_prompt,
|
| 588 |
+
prompt_embeds=prompt_embeds,
|
| 589 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
| 590 |
+
lora_scale=text_encoder_lora_scale,
|
| 591 |
+
)
|
| 592 |
+
# For classifier free guidance, we need to do two forward passes.
|
| 593 |
+
# Here we concatenate the unconditional and text embeddings into a single batch
|
| 594 |
+
# to avoid doing two forward passes
|
| 595 |
+
if do_classifier_free_guidance:
|
| 596 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
|
| 597 |
+
|
| 598 |
+
# 4. Prepare timesteps
|
| 599 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
| 600 |
+
timesteps = self.scheduler.timesteps
|
| 601 |
+
|
| 602 |
+
# 5. Prepare latent variables
|
| 603 |
+
num_channels_latents = self.unet.config.in_channels
|
| 604 |
+
latents = self.prepare_latents(
|
| 605 |
+
batch_size * num_images_per_prompt,
|
| 606 |
+
num_channels_latents,
|
| 607 |
+
height,
|
| 608 |
+
width,
|
| 609 |
+
prompt_embeds.dtype,
|
| 610 |
+
device,
|
| 611 |
+
generator,
|
| 612 |
+
latents,
|
| 613 |
+
)
|
| 614 |
+
|
| 615 |
+
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
| 616 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
| 617 |
+
|
| 618 |
+
# 7. Denoising loop
|
| 619 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
| 620 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
| 621 |
+
if get_attention_map:
|
| 622 |
+
attn_maps = {} # each t one attn map
|
| 623 |
+
|
| 624 |
+
for i, t in enumerate(timesteps):
|
| 625 |
+
# expand the latents if we are doing classifier free guidance
|
| 626 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
| 627 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
| 628 |
+
|
| 629 |
+
# predict the noise residual
|
| 630 |
+
noise_pred = self.unet(
|
| 631 |
+
latent_model_input,
|
| 632 |
+
t,
|
| 633 |
+
encoder_hidden_states=prompt_embeds,
|
| 634 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
| 635 |
+
return_dict=False,
|
| 636 |
+
)[0]
|
| 637 |
+
if get_attention_map:
|
| 638 |
+
attn_probs = {}
|
| 639 |
+
|
| 640 |
+
for name, module in self.unet.named_modules():
|
| 641 |
+
if isinstance(module, Attention) and module.attn_probs is not None:
|
| 642 |
+
a = module.attn_probs[1].mean(dim=0) # (2,Head,H,W,77)->(H,W,77)
|
| 643 |
+
attn_probs[name] = a
|
| 644 |
+
|
| 645 |
+
attn_maps[i] = attn_probs
|
| 646 |
+
|
| 647 |
+
# perform guidance
|
| 648 |
+
if do_classifier_free_guidance:
|
| 649 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 650 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
| 651 |
+
|
| 652 |
+
if do_classifier_free_guidance and guidance_rescale > 0.0:
|
| 653 |
+
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
|
| 654 |
+
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
|
| 655 |
+
|
| 656 |
+
# compute the previous noisy sample x_t -> x_t-1
|
| 657 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
| 658 |
+
|
| 659 |
+
# call the callback, if provided
|
| 660 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
| 661 |
+
progress_bar.update()
|
| 662 |
+
if callback is not None and i % callback_steps == 0:
|
| 663 |
+
callback(i, t, latents)
|
| 664 |
+
|
| 665 |
+
if get_attention_map:
|
| 666 |
+
output_maps = {}
|
| 667 |
+
for name in attn_probs.keys():
|
| 668 |
+
timeavg_maps = []
|
| 669 |
+
for i in attn_maps.keys():
|
| 670 |
+
timeavg_maps.append(attn_maps[i][name])
|
| 671 |
+
timeavg_maps = torch.stack(timeavg_maps, dim=0).mean(dim=0)
|
| 672 |
+
output_maps[name] = timeavg_maps
|
| 673 |
+
|
| 674 |
+
avg_attn_map = []
|
| 675 |
+
for name in attn_probs:
|
| 676 |
+
avg_attn_map.append(attn_probs[name])
|
| 677 |
+
avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0) # (5,B,H,W,77) -> (B,H,W,77)
|
| 678 |
+
output_maps['avg'] = avg_attn_map
|
| 679 |
+
|
| 680 |
+
del attn_maps
|
| 681 |
+
del attn_probs
|
| 682 |
+
self.attn_maps = output_maps
|
| 683 |
+
|
| 684 |
+
if not output_type == "latent":
|
| 685 |
+
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
| 686 |
+
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
|
| 687 |
+
else:
|
| 688 |
+
image = latents
|
| 689 |
+
has_nsfw_concept = None
|
| 690 |
+
|
| 691 |
+
if has_nsfw_concept is None:
|
| 692 |
+
do_denormalize = [True] * image.shape[0]
|
| 693 |
+
else:
|
| 694 |
+
do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
|
| 695 |
+
|
| 696 |
+
image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
|
| 697 |
+
|
| 698 |
+
# Offload last model to CPU
|
| 699 |
+
if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
|
| 700 |
+
self.final_offload_hook.offload()
|
| 701 |
+
|
| 702 |
+
if not return_dict:
|
| 703 |
+
return (image, has_nsfw_concept)
|
| 704 |
+
|
| 705 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
| 706 |
+
|
| 707 |
+
|
| 708 |
+
def create_args(output_dir, num_parts=8, num_k_per_part=256):
|
| 709 |
+
args = OmegaConf.create({
|
| 710 |
+
'pretrained_model_name_or_path': 'runwayml/stable-diffusion-v1-5',
|
| 711 |
+
'num_parts': num_parts,
|
| 712 |
+
'num_k_per_part': num_k_per_part,
|
| 713 |
+
'revision': None,
|
| 714 |
+
'variant': None,
|
| 715 |
+
'rank': 4,
|
| 716 |
+
'projection_nlayers': 1,
|
| 717 |
+
'output_dir': output_dir
|
| 718 |
+
})
|
| 719 |
+
folders = sorted(os.listdir(args.output_dir))
|
| 720 |
+
cps = [int(f.split('-')[1]) for f in folders if 'checkpoint' in f and '.ipynb' not in f]
|
| 721 |
+
maxcp = max(cps)
|
| 722 |
+
|
| 723 |
+
args.maxcp = maxcp
|
| 724 |
+
args.unet_path = None
|
| 725 |
+
return args
|
| 726 |
+
|
| 727 |
+
|
| 728 |
+
def load_pipeline(args, weight_dtype=torch.float16, device=torch.device('cuda')):
|
| 729 |
+
tokenizer = MultiTokenCLIPTokenizer.from_pretrained(
|
| 730 |
+
args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
|
| 731 |
+
)
|
| 732 |
+
text_encoder = CustomCLIPTextModel.from_pretrained(
|
| 733 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
|
| 734 |
+
)
|
| 735 |
+
unet_path = args.unet_path if args.unet_path is not None else args.pretrained_model_name_or_path
|
| 736 |
+
unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
|
| 737 |
+
unet_path, subfolder="unet", revision=args.revision
|
| 738 |
+
)
|
| 739 |
+
pipeline = DreamCreatureSDPipeline.from_pretrained(
|
| 740 |
+
args.pretrained_model_name_or_path,
|
| 741 |
+
unet=unet,
|
| 742 |
+
text_encoder=text_encoder,
|
| 743 |
+
tokenizer=tokenizer,
|
| 744 |
+
revision=args.revision,
|
| 745 |
+
torch_dtype=weight_dtype,
|
| 746 |
+
)
|
| 747 |
+
pipeline.num_k_per_part = args.num_k_per_part
|
| 748 |
+
pipeline.num_parts = args.num_parts
|
| 749 |
+
placeholder_token = "<part>"
|
| 750 |
+
initializer_token = None
|
| 751 |
+
placeholder_token_ids = add_tokens(tokenizer,
|
| 752 |
+
text_encoder,
|
| 753 |
+
placeholder_token,
|
| 754 |
+
args.num_parts,
|
| 755 |
+
initializer_token)
|
| 756 |
+
pipeline.placeholder_token_ids = placeholder_token_ids
|
| 757 |
+
pipeline.simple_mapper = TokenMapper(args.num_parts,
|
| 758 |
+
args.num_k_per_part,
|
| 759 |
+
768,
|
| 760 |
+
args.projection_nlayers)
|
| 761 |
+
pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + f'/checkpoint-{args.maxcp}/pytorch_model_1.bin',
|
| 762 |
+
map_location='cpu'))
|
| 763 |
+
pipeline.simple_mapper.to(device)
|
| 764 |
+
pipeline.replace_token = False
|
| 765 |
+
pipeline = pipeline.to(device)
|
| 766 |
+
|
| 767 |
+
# load attention processors
|
| 768 |
+
# pipeline.unet.load_attn_procs(args.output_dir, use_safetensors=not args.custom_diffusion)
|
| 769 |
+
setup_attn_processor(pipeline.unet, rank=args.rank)
|
| 770 |
+
load_attn_processor(pipeline.unet, args.output_dir + f'/checkpoint-{args.maxcp}/pytorch_model.bin')
|
| 771 |
+
return pipeline
|
dreamcreature/pipeline_xl.py
ADDED
|
@@ -0,0 +1,895 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
|
| 3 |
+
from diffusers.models.attention_processor import Attention
|
| 4 |
+
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import *
|
| 5 |
+
from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl import (
|
| 6 |
+
StableDiffusionXLPipeline,
|
| 7 |
+
StableDiffusionXLPipelineOutput,
|
| 8 |
+
XLA_AVAILABLE,
|
| 9 |
+
Tuple,
|
| 10 |
+
StableDiffusionXLLoraLoaderMixin,
|
| 11 |
+
PipelineImageInput
|
| 12 |
+
)
|
| 13 |
+
from omegaconf import OmegaConf
|
| 14 |
+
|
| 15 |
+
from dreamcreature.attn_processor import AttnProcessorCustom
|
| 16 |
+
from dreamcreature.mapper import TokenMapper
|
| 17 |
+
from dreamcreature.pipeline import convert_prompt
|
| 18 |
+
from dreamcreature.text_encoder import CustomCLIPTextModel, CustomCLIPTextModelWithProjection
|
| 19 |
+
from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
|
| 20 |
+
from utils import add_tokens
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def init_for_pipeline(args):
|
| 24 |
+
tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
|
| 25 |
+
args.pretrained_model_name_or_path,
|
| 26 |
+
subfolder="tokenizer",
|
| 27 |
+
revision=args.revision,
|
| 28 |
+
use_fast=False,
|
| 29 |
+
)
|
| 30 |
+
tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
|
| 31 |
+
args.pretrained_model_name_or_path,
|
| 32 |
+
subfolder="tokenizer_2",
|
| 33 |
+
revision=args.revision,
|
| 34 |
+
use_fast=False,
|
| 35 |
+
)
|
| 36 |
+
text_encoder_cls_one = CustomCLIPTextModel
|
| 37 |
+
text_encoder_cls_two = CustomCLIPTextModelWithProjection
|
| 38 |
+
text_encoder_one = text_encoder_cls_one.from_pretrained(
|
| 39 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
|
| 40 |
+
)
|
| 41 |
+
text_encoder_two = text_encoder_cls_two.from_pretrained(
|
| 42 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
OUT_DIMS = 768 + 1280 # 2048
|
| 46 |
+
simple_mapper = TokenMapper(args.num_parts,
|
| 47 |
+
args.num_k_per_part,
|
| 48 |
+
OUT_DIMS,
|
| 49 |
+
args.projection_nlayers)
|
| 50 |
+
return text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
class DreamCreatureSDXLPipeline(StableDiffusionXLPipeline):
|
| 54 |
+
def _maybe_convert_prompt(self, prompt: str, tokenizer: MultiTokenCLIPTokenizer):
|
| 55 |
+
r"""
|
| 56 |
+
Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
|
| 57 |
+
to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
|
| 58 |
+
is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
|
| 59 |
+
inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
|
| 60 |
+
|
| 61 |
+
Parameters:
|
| 62 |
+
prompt (`str`):
|
| 63 |
+
The prompt to guide the image generation.
|
| 64 |
+
tokenizer (`PreTrainedTokenizer`):
|
| 65 |
+
The tokenizer responsible for encoding the prompt into input tokens.
|
| 66 |
+
|
| 67 |
+
Returns:
|
| 68 |
+
`str`: The converted prompt
|
| 69 |
+
"""
|
| 70 |
+
if hasattr(self, 'replace_token'):
|
| 71 |
+
replace_token = self.replace_token
|
| 72 |
+
else:
|
| 73 |
+
replace_token = True
|
| 74 |
+
|
| 75 |
+
new_caption, code, parts_i = convert_prompt(prompt, replace_token)
|
| 76 |
+
|
| 77 |
+
if hasattr(self, 'verbose') and self.verbose:
|
| 78 |
+
print(new_caption)
|
| 79 |
+
|
| 80 |
+
return new_caption, code, parts_i
|
| 81 |
+
|
| 82 |
+
def compute_prompt_embeddings(self, prompts, text_encoder, tokenizer, device, index):
|
| 83 |
+
# textual inversion: procecss multi-vector tokens if necessary
|
| 84 |
+
if not isinstance(prompts, List):
|
| 85 |
+
prompts = [prompts]
|
| 86 |
+
|
| 87 |
+
prompt_embeds_concat = []
|
| 88 |
+
pooled_prompt_embeds_concat = []
|
| 89 |
+
for prompt in prompts:
|
| 90 |
+
prompt, code, parts_i = self.maybe_convert_prompt(prompt, tokenizer)
|
| 91 |
+
|
| 92 |
+
if hasattr(self, 'replace_token'):
|
| 93 |
+
replace_token = self.replace_token
|
| 94 |
+
else:
|
| 95 |
+
replace_token = True
|
| 96 |
+
|
| 97 |
+
text_inputs = tokenizer(
|
| 98 |
+
prompt,
|
| 99 |
+
replace_token=replace_token,
|
| 100 |
+
padding="max_length",
|
| 101 |
+
max_length=self.tokenizer.model_max_length,
|
| 102 |
+
truncation=True,
|
| 103 |
+
return_tensors="pt",
|
| 104 |
+
)
|
| 105 |
+
text_input_ids = text_inputs.input_ids
|
| 106 |
+
if hasattr(self, 'verbose') and self.verbose:
|
| 107 |
+
print(text_input_ids)
|
| 108 |
+
|
| 109 |
+
untruncated_ids = tokenizer(prompt,
|
| 110 |
+
replace_token=replace_token,
|
| 111 |
+
padding="longest",
|
| 112 |
+
return_tensors="pt").input_ids
|
| 113 |
+
|
| 114 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
| 115 |
+
text_input_ids, untruncated_ids
|
| 116 |
+
):
|
| 117 |
+
removed_text = tokenizer.batch_decode(
|
| 118 |
+
untruncated_ids[:, tokenizer.model_max_length - 1: -1]
|
| 119 |
+
)
|
| 120 |
+
logger.warning(
|
| 121 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
| 122 |
+
f" {tokenizer.model_max_length} tokens: {removed_text}"
|
| 123 |
+
)
|
| 124 |
+
|
| 125 |
+
if hasattr(text_encoder.config, "use_attention_mask") and text_encoder.config.use_attention_mask:
|
| 126 |
+
attention_mask = text_inputs.attention_mask.to(device)
|
| 127 |
+
else:
|
| 128 |
+
attention_mask = None
|
| 129 |
+
|
| 130 |
+
if code is None:
|
| 131 |
+
modified_hs = None
|
| 132 |
+
else:
|
| 133 |
+
placeholder_token_ids = self.placeholder_token_ids
|
| 134 |
+
placeholder_token_ids = [placeholder_token_ids[i] for i in parts_i] # follow the order of prompt's i
|
| 135 |
+
mapper_outputs = self.simple_mapper(code.unsqueeze(0).to(device), torch.tensor(parts_i).to(device))
|
| 136 |
+
if index == 0: # first encoder
|
| 137 |
+
mapper_outputs = mapper_outputs[..., :768]
|
| 138 |
+
else:
|
| 139 |
+
mapper_outputs = mapper_outputs[..., 768:]
|
| 140 |
+
modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids.to(device),
|
| 141 |
+
None,
|
| 142 |
+
mapper_outputs,
|
| 143 |
+
placeholder_token_ids)
|
| 144 |
+
|
| 145 |
+
prompt_embeds = text_encoder(
|
| 146 |
+
text_input_ids.to(device),
|
| 147 |
+
output_hidden_states=True,
|
| 148 |
+
attention_mask=attention_mask,
|
| 149 |
+
hidden_states=modified_hs
|
| 150 |
+
)
|
| 151 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 152 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
| 153 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
| 154 |
+
|
| 155 |
+
pooled_prompt_embeds_concat.append(pooled_prompt_embeds)
|
| 156 |
+
prompt_embeds_concat.append(prompt_embeds)
|
| 157 |
+
|
| 158 |
+
if len(prompt_embeds_concat) == 1:
|
| 159 |
+
return prompt_embeds_concat[0], pooled_prompt_embeds_concat[0]
|
| 160 |
+
else:
|
| 161 |
+
return torch.cat(prompt_embeds_concat, dim=0), torch.cat(pooled_prompt_embeds_concat, dim=0)
|
| 162 |
+
|
| 163 |
+
def encode_prompt(
|
| 164 |
+
self,
|
| 165 |
+
prompt: str,
|
| 166 |
+
prompt_2: Optional[str] = None,
|
| 167 |
+
device: Optional[torch.device] = None,
|
| 168 |
+
num_images_per_prompt: int = 1,
|
| 169 |
+
do_classifier_free_guidance: bool = True,
|
| 170 |
+
negative_prompt: Optional[str] = None,
|
| 171 |
+
negative_prompt_2: Optional[str] = None,
|
| 172 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 173 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 174 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 175 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 176 |
+
lora_scale: Optional[float] = None,
|
| 177 |
+
clip_skip: Optional[int] = None,
|
| 178 |
+
):
|
| 179 |
+
r"""
|
| 180 |
+
Encodes the prompt into text encoder hidden states.
|
| 181 |
+
|
| 182 |
+
Args:
|
| 183 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 184 |
+
prompt to be encoded
|
| 185 |
+
prompt_2 (`str` or `List[str]`, *optional*):
|
| 186 |
+
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
| 187 |
+
used in both text-encoders
|
| 188 |
+
device: (`torch.device`):
|
| 189 |
+
torch device
|
| 190 |
+
num_images_per_prompt (`int`):
|
| 191 |
+
number of images that should be generated per prompt
|
| 192 |
+
do_classifier_free_guidance (`bool`):
|
| 193 |
+
whether to use classifier free guidance or not
|
| 194 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 195 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 196 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
| 197 |
+
less than `1`).
|
| 198 |
+
negative_prompt_2 (`str` or `List[str]`, *optional*):
|
| 199 |
+
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
|
| 200 |
+
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
|
| 201 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 202 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
| 203 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
| 204 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 205 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 206 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
| 207 |
+
argument.
|
| 208 |
+
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 209 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
| 210 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
| 211 |
+
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 212 |
+
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 213 |
+
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
|
| 214 |
+
input argument.
|
| 215 |
+
lora_scale (`float`, *optional*):
|
| 216 |
+
A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
| 217 |
+
clip_skip (`int`, *optional*):
|
| 218 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
| 219 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
| 220 |
+
"""
|
| 221 |
+
device = device or self._execution_device
|
| 222 |
+
|
| 223 |
+
# set lora scale so that monkey patched LoRA
|
| 224 |
+
# function of text encoder can correctly access it
|
| 225 |
+
if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
|
| 226 |
+
self._lora_scale = lora_scale
|
| 227 |
+
|
| 228 |
+
# dynamically adjust the LoRA scale
|
| 229 |
+
if self.text_encoder is not None:
|
| 230 |
+
if not USE_PEFT_BACKEND:
|
| 231 |
+
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
| 232 |
+
else:
|
| 233 |
+
scale_lora_layers(self.text_encoder, lora_scale)
|
| 234 |
+
|
| 235 |
+
if self.text_encoder_2 is not None:
|
| 236 |
+
if not USE_PEFT_BACKEND:
|
| 237 |
+
adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
|
| 238 |
+
else:
|
| 239 |
+
scale_lora_layers(self.text_encoder_2, lora_scale)
|
| 240 |
+
|
| 241 |
+
prompt = [prompt] if isinstance(prompt, str) else prompt
|
| 242 |
+
|
| 243 |
+
if prompt is not None:
|
| 244 |
+
batch_size = len(prompt)
|
| 245 |
+
else:
|
| 246 |
+
batch_size = prompt_embeds.shape[0]
|
| 247 |
+
|
| 248 |
+
# Define tokenizers and text encoders
|
| 249 |
+
tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
|
| 250 |
+
text_encoders = (
|
| 251 |
+
[self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
|
| 252 |
+
)
|
| 253 |
+
|
| 254 |
+
if prompt_embeds is None:
|
| 255 |
+
prompt_2 = prompt_2 or prompt
|
| 256 |
+
prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
|
| 257 |
+
|
| 258 |
+
# textual inversion: procecss multi-vector tokens if necessary
|
| 259 |
+
prompt_embeds_list = []
|
| 260 |
+
prompts = [prompt, prompt_2]
|
| 261 |
+
index = 0
|
| 262 |
+
for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
|
| 263 |
+
prompt_embeds, pooled_prompt_embeds = self.compute_prompt_embeddings(prompt,
|
| 264 |
+
text_encoder,
|
| 265 |
+
tokenizer,
|
| 266 |
+
device,
|
| 267 |
+
index)
|
| 268 |
+
prompt_embeds_list.append(prompt_embeds)
|
| 269 |
+
index += 1
|
| 270 |
+
|
| 271 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
| 272 |
+
|
| 273 |
+
# get unconditional embeddings for classifier free guidance
|
| 274 |
+
zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
|
| 275 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
|
| 276 |
+
negative_prompt_embeds = torch.zeros_like(prompt_embeds)
|
| 277 |
+
negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
|
| 278 |
+
elif do_classifier_free_guidance and negative_prompt_embeds is None:
|
| 279 |
+
negative_prompt = negative_prompt or ""
|
| 280 |
+
negative_prompt_2 = negative_prompt_2 or negative_prompt
|
| 281 |
+
|
| 282 |
+
# normalize str to list
|
| 283 |
+
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
|
| 284 |
+
negative_prompt_2 = (
|
| 285 |
+
batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
|
| 286 |
+
)
|
| 287 |
+
|
| 288 |
+
uncond_tokens: List[str]
|
| 289 |
+
if prompt is not None and type(prompt) is not type(negative_prompt):
|
| 290 |
+
raise TypeError(
|
| 291 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
| 292 |
+
f" {type(prompt)}."
|
| 293 |
+
)
|
| 294 |
+
elif batch_size != len(negative_prompt):
|
| 295 |
+
raise ValueError(
|
| 296 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
| 297 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
| 298 |
+
" the batch size of `prompt`."
|
| 299 |
+
)
|
| 300 |
+
else:
|
| 301 |
+
uncond_tokens = [negative_prompt, negative_prompt_2]
|
| 302 |
+
|
| 303 |
+
negative_prompt_embeds_list = []
|
| 304 |
+
index = 0
|
| 305 |
+
for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
|
| 306 |
+
negative_prompt_embeds, negative_pooled_prompt_embeds = self.compute_prompt_embeddings(negative_prompt,
|
| 307 |
+
text_encoder,
|
| 308 |
+
tokenizer,
|
| 309 |
+
device,
|
| 310 |
+
index)
|
| 311 |
+
negative_prompt_embeds_list.append(negative_prompt_embeds)
|
| 312 |
+
index += 1
|
| 313 |
+
|
| 314 |
+
negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
|
| 315 |
+
|
| 316 |
+
if self.text_encoder_2 is not None:
|
| 317 |
+
prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
| 318 |
+
else:
|
| 319 |
+
prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
|
| 320 |
+
|
| 321 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 322 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
| 323 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 324 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
| 325 |
+
|
| 326 |
+
if do_classifier_free_guidance:
|
| 327 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
| 328 |
+
seq_len = negative_prompt_embeds.shape[1]
|
| 329 |
+
|
| 330 |
+
if self.text_encoder_2 is not None:
|
| 331 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
|
| 332 |
+
else:
|
| 333 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
|
| 334 |
+
|
| 335 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 336 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
| 337 |
+
|
| 338 |
+
pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
| 339 |
+
bs_embed * num_images_per_prompt, -1
|
| 340 |
+
)
|
| 341 |
+
if do_classifier_free_guidance:
|
| 342 |
+
negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
|
| 343 |
+
bs_embed * num_images_per_prompt, -1
|
| 344 |
+
)
|
| 345 |
+
|
| 346 |
+
if self.text_encoder is not None:
|
| 347 |
+
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
|
| 348 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
| 349 |
+
unscale_lora_layers(self.text_encoder, lora_scale)
|
| 350 |
+
|
| 351 |
+
if self.text_encoder_2 is not None:
|
| 352 |
+
if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
|
| 353 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
| 354 |
+
unscale_lora_layers(self.text_encoder_2, lora_scale)
|
| 355 |
+
|
| 356 |
+
return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
|
| 357 |
+
|
| 358 |
+
@torch.no_grad()
|
| 359 |
+
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
| 360 |
+
def __call__(
|
| 361 |
+
self,
|
| 362 |
+
prompt: Union[str, List[str]] = None,
|
| 363 |
+
prompt_2: Optional[Union[str, List[str]]] = None,
|
| 364 |
+
height: Optional[int] = None,
|
| 365 |
+
width: Optional[int] = None,
|
| 366 |
+
num_inference_steps: int = 50,
|
| 367 |
+
timesteps: List[int] = None,
|
| 368 |
+
denoising_end: Optional[float] = None,
|
| 369 |
+
guidance_scale: float = 5.0,
|
| 370 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 371 |
+
negative_prompt_2: Optional[Union[str, List[str]]] = None,
|
| 372 |
+
num_images_per_prompt: Optional[int] = 1,
|
| 373 |
+
eta: float = 0.0,
|
| 374 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
| 375 |
+
latents: Optional[torch.FloatTensor] = None,
|
| 376 |
+
prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 377 |
+
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 378 |
+
pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 379 |
+
negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
|
| 380 |
+
ip_adapter_image: Optional[PipelineImageInput] = None,
|
| 381 |
+
output_type: Optional[str] = "pil",
|
| 382 |
+
return_dict: bool = True,
|
| 383 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
| 384 |
+
guidance_rescale: float = 0.0,
|
| 385 |
+
original_size: Optional[Tuple[int, int]] = None,
|
| 386 |
+
crops_coords_top_left: Tuple[int, int] = (0, 0),
|
| 387 |
+
target_size: Optional[Tuple[int, int]] = None,
|
| 388 |
+
negative_original_size: Optional[Tuple[int, int]] = None,
|
| 389 |
+
negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
|
| 390 |
+
negative_target_size: Optional[Tuple[int, int]] = None,
|
| 391 |
+
clip_skip: Optional[int] = None,
|
| 392 |
+
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
| 393 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
| 394 |
+
get_attention_map: bool = False,
|
| 395 |
+
**kwargs,
|
| 396 |
+
):
|
| 397 |
+
r"""
|
| 398 |
+
Function invoked when calling the pipeline for generation.
|
| 399 |
+
|
| 400 |
+
Args:
|
| 401 |
+
prompt (`str` or `List[str]`, *optional*):
|
| 402 |
+
The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
|
| 403 |
+
instead.
|
| 404 |
+
prompt_2 (`str` or `List[str]`, *optional*):
|
| 405 |
+
The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
|
| 406 |
+
used in both text-encoders
|
| 407 |
+
height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
| 408 |
+
The height in pixels of the generated image. This is set to 1024 by default for the best results.
|
| 409 |
+
Anything below 512 pixels won't work well for
|
| 410 |
+
[stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
|
| 411 |
+
and checkpoints that are not specifically fine-tuned on low resolutions.
|
| 412 |
+
width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
|
| 413 |
+
The width in pixels of the generated image. This is set to 1024 by default for the best results.
|
| 414 |
+
Anything below 512 pixels won't work well for
|
| 415 |
+
[stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
|
| 416 |
+
and checkpoints that are not specifically fine-tuned on low resolutions.
|
| 417 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
| 418 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
| 419 |
+
expense of slower inference.
|
| 420 |
+
timesteps (`List[int]`, *optional*):
|
| 421 |
+
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
|
| 422 |
+
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
|
| 423 |
+
passed will be used. Must be in descending order.
|
| 424 |
+
denoising_end (`float`, *optional*):
|
| 425 |
+
When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
|
| 426 |
+
completed before it is intentionally prematurely terminated. As a result, the returned sample will
|
| 427 |
+
still retain a substantial amount of noise as determined by the discrete timesteps selected by the
|
| 428 |
+
scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
|
| 429 |
+
"Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
|
| 430 |
+
Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
|
| 431 |
+
guidance_scale (`float`, *optional*, defaults to 5.0):
|
| 432 |
+
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
| 433 |
+
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
| 434 |
+
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
| 435 |
+
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
| 436 |
+
usually at the expense of lower image quality.
|
| 437 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
| 438 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
| 439 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
| 440 |
+
less than `1`).
|
| 441 |
+
negative_prompt_2 (`str` or `List[str]`, *optional*):
|
| 442 |
+
The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
|
| 443 |
+
`text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
|
| 444 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
| 445 |
+
The number of images to generate per prompt.
|
| 446 |
+
eta (`float`, *optional*, defaults to 0.0):
|
| 447 |
+
Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
|
| 448 |
+
[`schedulers.DDIMScheduler`], will be ignored for others.
|
| 449 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
| 450 |
+
One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
|
| 451 |
+
to make generation deterministic.
|
| 452 |
+
latents (`torch.FloatTensor`, *optional*):
|
| 453 |
+
Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
|
| 454 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
| 455 |
+
tensor will ge generated by sampling using the supplied random `generator`.
|
| 456 |
+
prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 457 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
| 458 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
| 459 |
+
negative_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 460 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 461 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
| 462 |
+
argument.
|
| 463 |
+
pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 464 |
+
Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
|
| 465 |
+
If not provided, pooled text embeddings will be generated from `prompt` input argument.
|
| 466 |
+
negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
|
| 467 |
+
Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
| 468 |
+
weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
|
| 469 |
+
input argument.
|
| 470 |
+
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
|
| 471 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
| 472 |
+
The output format of the generate image. Choose between
|
| 473 |
+
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
| 474 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
| 475 |
+
Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
|
| 476 |
+
of a plain tuple.
|
| 477 |
+
cross_attention_kwargs (`dict`, *optional*):
|
| 478 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
| 479 |
+
`self.processor` in
|
| 480 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
| 481 |
+
guidance_rescale (`float`, *optional*, defaults to 0.0):
|
| 482 |
+
Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
|
| 483 |
+
Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
|
| 484 |
+
[Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
|
| 485 |
+
Guidance rescale factor should fix overexposure when using zero terminal SNR.
|
| 486 |
+
original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
| 487 |
+
If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
|
| 488 |
+
`original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
|
| 489 |
+
explained in section 2.2 of
|
| 490 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
| 491 |
+
crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
|
| 492 |
+
`crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
|
| 493 |
+
`crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
|
| 494 |
+
`crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
|
| 495 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
| 496 |
+
target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
| 497 |
+
For most cases, `target_size` should be set to the desired height and width of the generated image. If
|
| 498 |
+
not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
|
| 499 |
+
section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
|
| 500 |
+
negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
| 501 |
+
To negatively condition the generation process based on a specific image resolution. Part of SDXL's
|
| 502 |
+
micro-conditioning as explained in section 2.2 of
|
| 503 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
|
| 504 |
+
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
|
| 505 |
+
negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
|
| 506 |
+
To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
|
| 507 |
+
micro-conditioning as explained in section 2.2 of
|
| 508 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
|
| 509 |
+
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
|
| 510 |
+
negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
|
| 511 |
+
To negatively condition the generation process based on a target image resolution. It should be as same
|
| 512 |
+
as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
|
| 513 |
+
[https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
|
| 514 |
+
information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
|
| 515 |
+
callback_on_step_end (`Callable`, *optional*):
|
| 516 |
+
A function that calls at the end of each denoising steps during the inference. The function is called
|
| 517 |
+
with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
|
| 518 |
+
callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
|
| 519 |
+
`callback_on_step_end_tensor_inputs`.
|
| 520 |
+
callback_on_step_end_tensor_inputs (`List`, *optional*):
|
| 521 |
+
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
| 522 |
+
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
| 523 |
+
`._callback_tensor_inputs` attribute of your pipeline class.
|
| 524 |
+
|
| 525 |
+
Examples:
|
| 526 |
+
|
| 527 |
+
Returns:
|
| 528 |
+
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
|
| 529 |
+
[`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
|
| 530 |
+
`tuple`. When returning a tuple, the first element is a list with the generated images.
|
| 531 |
+
"""
|
| 532 |
+
|
| 533 |
+
callback = kwargs.pop("callback", None)
|
| 534 |
+
callback_steps = kwargs.pop("callback_steps", None)
|
| 535 |
+
|
| 536 |
+
if callback is not None:
|
| 537 |
+
deprecate(
|
| 538 |
+
"callback",
|
| 539 |
+
"1.0.0",
|
| 540 |
+
"Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
|
| 541 |
+
)
|
| 542 |
+
if callback_steps is not None:
|
| 543 |
+
deprecate(
|
| 544 |
+
"callback_steps",
|
| 545 |
+
"1.0.0",
|
| 546 |
+
"Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
|
| 547 |
+
)
|
| 548 |
+
|
| 549 |
+
# 0. Default height and width to unet
|
| 550 |
+
height = height or self.default_sample_size * self.vae_scale_factor
|
| 551 |
+
width = width or self.default_sample_size * self.vae_scale_factor
|
| 552 |
+
|
| 553 |
+
original_size = original_size or (height, width)
|
| 554 |
+
target_size = target_size or (height, width)
|
| 555 |
+
|
| 556 |
+
# 1. Check inputs. Raise error if not correct
|
| 557 |
+
self.check_inputs(
|
| 558 |
+
prompt,
|
| 559 |
+
prompt_2,
|
| 560 |
+
height,
|
| 561 |
+
width,
|
| 562 |
+
callback_steps,
|
| 563 |
+
negative_prompt,
|
| 564 |
+
negative_prompt_2,
|
| 565 |
+
prompt_embeds,
|
| 566 |
+
negative_prompt_embeds,
|
| 567 |
+
pooled_prompt_embeds,
|
| 568 |
+
negative_pooled_prompt_embeds,
|
| 569 |
+
callback_on_step_end_tensor_inputs,
|
| 570 |
+
)
|
| 571 |
+
|
| 572 |
+
self._guidance_scale = guidance_scale
|
| 573 |
+
self._guidance_rescale = guidance_rescale
|
| 574 |
+
self._clip_skip = clip_skip
|
| 575 |
+
self._cross_attention_kwargs = cross_attention_kwargs
|
| 576 |
+
self._denoising_end = denoising_end
|
| 577 |
+
|
| 578 |
+
# 2. Define call parameters
|
| 579 |
+
if prompt is not None and isinstance(prompt, str):
|
| 580 |
+
batch_size = 1
|
| 581 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 582 |
+
batch_size = len(prompt)
|
| 583 |
+
else:
|
| 584 |
+
batch_size = prompt_embeds.shape[0]
|
| 585 |
+
|
| 586 |
+
device = self._execution_device
|
| 587 |
+
|
| 588 |
+
# 3. Encode input prompt
|
| 589 |
+
lora_scale = (
|
| 590 |
+
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
|
| 591 |
+
)
|
| 592 |
+
|
| 593 |
+
(
|
| 594 |
+
prompt_embeds,
|
| 595 |
+
negative_prompt_embeds,
|
| 596 |
+
pooled_prompt_embeds,
|
| 597 |
+
negative_pooled_prompt_embeds,
|
| 598 |
+
) = self.encode_prompt(
|
| 599 |
+
prompt=prompt,
|
| 600 |
+
prompt_2=prompt_2,
|
| 601 |
+
device=device,
|
| 602 |
+
num_images_per_prompt=num_images_per_prompt,
|
| 603 |
+
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
| 604 |
+
negative_prompt=negative_prompt,
|
| 605 |
+
negative_prompt_2=negative_prompt_2,
|
| 606 |
+
prompt_embeds=prompt_embeds,
|
| 607 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
| 608 |
+
pooled_prompt_embeds=pooled_prompt_embeds,
|
| 609 |
+
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
| 610 |
+
lora_scale=lora_scale,
|
| 611 |
+
clip_skip=self.clip_skip,
|
| 612 |
+
)
|
| 613 |
+
|
| 614 |
+
# 4. Prepare timesteps
|
| 615 |
+
timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
|
| 616 |
+
|
| 617 |
+
# 5. Prepare latent variables
|
| 618 |
+
num_channels_latents = self.unet.config.in_channels
|
| 619 |
+
latents = self.prepare_latents(
|
| 620 |
+
batch_size * num_images_per_prompt,
|
| 621 |
+
num_channels_latents,
|
| 622 |
+
height,
|
| 623 |
+
width,
|
| 624 |
+
prompt_embeds.dtype,
|
| 625 |
+
device,
|
| 626 |
+
generator,
|
| 627 |
+
latents,
|
| 628 |
+
)
|
| 629 |
+
|
| 630 |
+
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
| 631 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
| 632 |
+
|
| 633 |
+
# 7. Prepare added time ids & embeddings
|
| 634 |
+
add_text_embeds = pooled_prompt_embeds
|
| 635 |
+
if self.text_encoder_2 is None:
|
| 636 |
+
text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
|
| 637 |
+
else:
|
| 638 |
+
text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
|
| 639 |
+
|
| 640 |
+
add_time_ids = self._get_add_time_ids(
|
| 641 |
+
original_size,
|
| 642 |
+
crops_coords_top_left,
|
| 643 |
+
target_size,
|
| 644 |
+
dtype=prompt_embeds.dtype,
|
| 645 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
| 646 |
+
)
|
| 647 |
+
if negative_original_size is not None and negative_target_size is not None:
|
| 648 |
+
negative_add_time_ids = self._get_add_time_ids(
|
| 649 |
+
negative_original_size,
|
| 650 |
+
negative_crops_coords_top_left,
|
| 651 |
+
negative_target_size,
|
| 652 |
+
dtype=prompt_embeds.dtype,
|
| 653 |
+
text_encoder_projection_dim=text_encoder_projection_dim,
|
| 654 |
+
)
|
| 655 |
+
else:
|
| 656 |
+
negative_add_time_ids = add_time_ids
|
| 657 |
+
|
| 658 |
+
if self.do_classifier_free_guidance:
|
| 659 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
|
| 660 |
+
add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
|
| 661 |
+
add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
|
| 662 |
+
|
| 663 |
+
prompt_embeds = prompt_embeds.to(device)
|
| 664 |
+
add_text_embeds = add_text_embeds.to(device)
|
| 665 |
+
add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
|
| 666 |
+
|
| 667 |
+
if ip_adapter_image is not None:
|
| 668 |
+
output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
|
| 669 |
+
image_embeds, negative_image_embeds = self.encode_image(
|
| 670 |
+
ip_adapter_image, device, num_images_per_prompt, output_hidden_state
|
| 671 |
+
)
|
| 672 |
+
if self.do_classifier_free_guidance:
|
| 673 |
+
image_embeds = torch.cat([negative_image_embeds, image_embeds])
|
| 674 |
+
image_embeds = image_embeds.to(device)
|
| 675 |
+
|
| 676 |
+
# 8. Denoising loop
|
| 677 |
+
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
| 678 |
+
|
| 679 |
+
# 8.1 Apply denoising_end
|
| 680 |
+
if (
|
| 681 |
+
self.denoising_end is not None
|
| 682 |
+
and isinstance(self.denoising_end, float)
|
| 683 |
+
and self.denoising_end > 0
|
| 684 |
+
and self.denoising_end < 1
|
| 685 |
+
):
|
| 686 |
+
discrete_timestep_cutoff = int(
|
| 687 |
+
round(
|
| 688 |
+
self.scheduler.config.num_train_timesteps
|
| 689 |
+
- (self.denoising_end * self.scheduler.config.num_train_timesteps)
|
| 690 |
+
)
|
| 691 |
+
)
|
| 692 |
+
num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
|
| 693 |
+
timesteps = timesteps[:num_inference_steps]
|
| 694 |
+
|
| 695 |
+
# 9. Optionally get Guidance Scale Embedding
|
| 696 |
+
timestep_cond = None
|
| 697 |
+
if self.unet.config.time_cond_proj_dim is not None:
|
| 698 |
+
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
|
| 699 |
+
timestep_cond = self.get_guidance_scale_embedding(
|
| 700 |
+
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
|
| 701 |
+
).to(device=device, dtype=latents.dtype)
|
| 702 |
+
|
| 703 |
+
self._num_timesteps = len(timesteps)
|
| 704 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
| 705 |
+
if get_attention_map:
|
| 706 |
+
attn_maps = {} # each t one attn map
|
| 707 |
+
|
| 708 |
+
for i, t in enumerate(timesteps):
|
| 709 |
+
# expand the latents if we are doing classifier free guidance
|
| 710 |
+
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
|
| 711 |
+
|
| 712 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
| 713 |
+
|
| 714 |
+
# predict the noise residual
|
| 715 |
+
added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
|
| 716 |
+
if ip_adapter_image is not None:
|
| 717 |
+
added_cond_kwargs["image_embeds"] = image_embeds
|
| 718 |
+
noise_pred = self.unet(
|
| 719 |
+
latent_model_input,
|
| 720 |
+
t,
|
| 721 |
+
encoder_hidden_states=prompt_embeds,
|
| 722 |
+
timestep_cond=timestep_cond,
|
| 723 |
+
cross_attention_kwargs=self.cross_attention_kwargs,
|
| 724 |
+
added_cond_kwargs=added_cond_kwargs,
|
| 725 |
+
return_dict=False,
|
| 726 |
+
)[0]
|
| 727 |
+
if get_attention_map:
|
| 728 |
+
attn_probs = {}
|
| 729 |
+
|
| 730 |
+
for name, module in self.unet.named_modules():
|
| 731 |
+
if isinstance(module, Attention) and module.attn_probs is not None:
|
| 732 |
+
# take 1 because we are taking the noise_pred_text
|
| 733 |
+
a = module.attn_probs[1].mean(dim=0) # (2,Head,H,W,77)->(H,W,77)
|
| 734 |
+
attn_probs[name] = a
|
| 735 |
+
|
| 736 |
+
attn_maps[i] = attn_probs
|
| 737 |
+
|
| 738 |
+
# perform guidance
|
| 739 |
+
if self.do_classifier_free_guidance:
|
| 740 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
| 741 |
+
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
| 742 |
+
|
| 743 |
+
if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
|
| 744 |
+
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
|
| 745 |
+
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
|
| 746 |
+
|
| 747 |
+
# compute the previous noisy sample x_t -> x_t-1
|
| 748 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
| 749 |
+
|
| 750 |
+
if callback_on_step_end is not None:
|
| 751 |
+
callback_kwargs = {}
|
| 752 |
+
for k in callback_on_step_end_tensor_inputs:
|
| 753 |
+
callback_kwargs[k] = locals()[k]
|
| 754 |
+
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
| 755 |
+
|
| 756 |
+
latents = callback_outputs.pop("latents", latents)
|
| 757 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
| 758 |
+
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
|
| 759 |
+
add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
|
| 760 |
+
negative_pooled_prompt_embeds = callback_outputs.pop(
|
| 761 |
+
"negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
|
| 762 |
+
)
|
| 763 |
+
add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
|
| 764 |
+
negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
|
| 765 |
+
|
| 766 |
+
# call the callback, if provided
|
| 767 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
| 768 |
+
progress_bar.update()
|
| 769 |
+
if callback is not None and i % callback_steps == 0:
|
| 770 |
+
step_idx = i // getattr(self.scheduler, "order", 1)
|
| 771 |
+
callback(step_idx, t, latents)
|
| 772 |
+
|
| 773 |
+
if XLA_AVAILABLE:
|
| 774 |
+
import torch_xla.core.xla_model as xm
|
| 775 |
+
xm.mark_step()
|
| 776 |
+
|
| 777 |
+
if get_attention_map:
|
| 778 |
+
output_maps = {}
|
| 779 |
+
for name in attn_probs.keys():
|
| 780 |
+
timeavg_maps = []
|
| 781 |
+
for i in attn_maps.keys():
|
| 782 |
+
timeavg_maps.append(attn_maps[i][name])
|
| 783 |
+
timeavg_maps = torch.stack(timeavg_maps, dim=0).mean(dim=0)
|
| 784 |
+
output_maps[name] = timeavg_maps
|
| 785 |
+
|
| 786 |
+
avg_attn_map = []
|
| 787 |
+
for name in attn_probs:
|
| 788 |
+
avg_attn_map.append(attn_probs[name])
|
| 789 |
+
avg_attn_map = torch.stack(avg_attn_map, dim=0).mean(dim=0) # (5,B,H,W,77) -> (B,H,W,77)
|
| 790 |
+
output_maps['avg'] = avg_attn_map
|
| 791 |
+
|
| 792 |
+
del attn_maps
|
| 793 |
+
del attn_probs
|
| 794 |
+
self.attn_maps = output_maps
|
| 795 |
+
|
| 796 |
+
if not output_type == "latent":
|
| 797 |
+
# make sure the VAE is in float32 mode, as it overflows in float16
|
| 798 |
+
needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
|
| 799 |
+
|
| 800 |
+
if needs_upcasting:
|
| 801 |
+
self.upcast_vae()
|
| 802 |
+
latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
|
| 803 |
+
|
| 804 |
+
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
|
| 805 |
+
|
| 806 |
+
# cast back to fp16 if needed
|
| 807 |
+
if needs_upcasting:
|
| 808 |
+
self.vae.to(dtype=torch.float16)
|
| 809 |
+
else:
|
| 810 |
+
image = latents
|
| 811 |
+
|
| 812 |
+
if not output_type == "latent":
|
| 813 |
+
# apply watermark if available
|
| 814 |
+
if self.watermark is not None:
|
| 815 |
+
image = self.watermark.apply_watermark(image)
|
| 816 |
+
|
| 817 |
+
image = self.image_processor.postprocess(image, output_type=output_type)
|
| 818 |
+
|
| 819 |
+
# Offload all models
|
| 820 |
+
self.maybe_free_model_hooks()
|
| 821 |
+
|
| 822 |
+
if not return_dict:
|
| 823 |
+
return (image,)
|
| 824 |
+
|
| 825 |
+
return StableDiffusionXLPipelineOutput(images=image)
|
| 826 |
+
|
| 827 |
+
|
| 828 |
+
def create_args_xl(output_dir, num_parts=8, num_k_per_part=256):
|
| 829 |
+
args = OmegaConf.create({
|
| 830 |
+
'pretrained_model_name_or_path': 'stabilityai/stable-diffusion-xl-base-1.0',
|
| 831 |
+
'num_parts': num_parts,
|
| 832 |
+
'num_k_per_part': num_k_per_part,
|
| 833 |
+
'revision': None,
|
| 834 |
+
'variant': None,
|
| 835 |
+
'rank': 4,
|
| 836 |
+
'projection_nlayers': 1,
|
| 837 |
+
'output_dir': output_dir
|
| 838 |
+
})
|
| 839 |
+
folders = sorted(os.listdir(args.output_dir))
|
| 840 |
+
cps = [int(f.split('-')[1]) for f in folders if 'checkpoint' in f and '.ipynb' not in f]
|
| 841 |
+
maxcp = max(cps)
|
| 842 |
+
|
| 843 |
+
args.maxcp = maxcp
|
| 844 |
+
return args
|
| 845 |
+
|
| 846 |
+
|
| 847 |
+
def load_pipeline_xl(args, weight_dtype=torch.float16, device=torch.device('cuda')):
|
| 848 |
+
text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper = init_for_pipeline(args)
|
| 849 |
+
placeholder_token = "<part>"
|
| 850 |
+
initializer_token = None
|
| 851 |
+
placeholder_token_ids_one = add_tokens(tokenizer_one,
|
| 852 |
+
text_encoder_one,
|
| 853 |
+
placeholder_token,
|
| 854 |
+
args.num_parts,
|
| 855 |
+
initializer_token)
|
| 856 |
+
placeholder_token_ids_two = add_tokens(tokenizer_two,
|
| 857 |
+
text_encoder_two,
|
| 858 |
+
placeholder_token,
|
| 859 |
+
args.num_parts,
|
| 860 |
+
initializer_token)
|
| 861 |
+
|
| 862 |
+
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
|
| 863 |
+
pipeline = DreamCreatureSDXLPipeline.from_pretrained(
|
| 864 |
+
args.pretrained_model_name_or_path,
|
| 865 |
+
vae=vae,
|
| 866 |
+
tokenizer=tokenizer_one,
|
| 867 |
+
tokenizer_2=tokenizer_two,
|
| 868 |
+
text_encoder=text_encoder_one,
|
| 869 |
+
text_encoder_2=text_encoder_two,
|
| 870 |
+
revision=args.revision,
|
| 871 |
+
variant=args.variant,
|
| 872 |
+
torch_dtype=weight_dtype,
|
| 873 |
+
)
|
| 874 |
+
pipeline.placeholder_token_ids = placeholder_token_ids_one
|
| 875 |
+
pipeline.replace_token = False
|
| 876 |
+
pipeline.simple_mapper = simple_mapper
|
| 877 |
+
pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + f'/checkpoint-{args.maxcp}/hash_mapper.pth',
|
| 878 |
+
map_location='cpu'))
|
| 879 |
+
|
| 880 |
+
pipeline.simple_mapper.to(device)
|
| 881 |
+
pipeline = pipeline.to(device)
|
| 882 |
+
|
| 883 |
+
# load attention processors
|
| 884 |
+
pipeline.load_lora_weights(args.output_dir + f'/checkpoint-{args.maxcp}')
|
| 885 |
+
|
| 886 |
+
def setup_attn_processors(unet, attn_size):
|
| 887 |
+
attn_procs = {}
|
| 888 |
+
for name in unet.attn_processors.keys():
|
| 889 |
+
attn_procs[name] = AttnProcessorCustom(attn_size)
|
| 890 |
+
unet.set_attn_processor(attn_procs)
|
| 891 |
+
|
| 892 |
+
pipeline = pipeline.to(weight_dtype)
|
| 893 |
+
setup_attn_processors(pipeline.unet, 16)
|
| 894 |
+
|
| 895 |
+
return pipeline
|
dreamcreature/text_encoder.py
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from collections import defaultdict
|
| 2 |
+
from typing import Optional, Tuple, Union
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
import torch.utils.checkpoint
|
| 6 |
+
from transformers.modeling_outputs import BaseModelOutputWithPooling
|
| 7 |
+
from transformers.models.clip.modeling_clip import (CLIPTextConfig,
|
| 8 |
+
CLIPTextModel,
|
| 9 |
+
CLIPTextModelWithProjection,
|
| 10 |
+
CLIPTextModelOutput)
|
| 11 |
+
from transformers.models.clip.modeling_clip import (CLIPTextTransformer,
|
| 12 |
+
_prepare_4d_attention_mask,
|
| 13 |
+
_create_4d_causal_attention_mask)
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class CustomCLIPTextModel(CLIPTextModel):
|
| 17 |
+
""" Modification of CLIPTextModel to use our NeTI mapper for computing the embeddings of the concept. """
|
| 18 |
+
|
| 19 |
+
def __init__(self, config: CLIPTextConfig):
|
| 20 |
+
super().__init__(config)
|
| 21 |
+
self.text_model = CustomCLIPTextTransformer(config)
|
| 22 |
+
self.post_init()
|
| 23 |
+
|
| 24 |
+
def forward(self, input_ids: Optional[torch.Tensor] = None,
|
| 25 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 26 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 27 |
+
output_attentions: Optional[bool] = None,
|
| 28 |
+
output_hidden_states: Optional[bool] = None,
|
| 29 |
+
return_dict: Optional[bool] = None,
|
| 30 |
+
hidden_states: Optional[torch.Tensor] = None) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 31 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 32 |
+
|
| 33 |
+
return self.text_model.forward(
|
| 34 |
+
input_ids=input_ids,
|
| 35 |
+
attention_mask=attention_mask,
|
| 36 |
+
position_ids=position_ids,
|
| 37 |
+
output_attentions=output_attentions,
|
| 38 |
+
output_hidden_states=output_hidden_states,
|
| 39 |
+
return_dict=return_dict,
|
| 40 |
+
hidden_states=hidden_states
|
| 41 |
+
)
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
class CustomCLIPTextTransformer(CLIPTextTransformer):
|
| 45 |
+
def forward_embeddings(self, input_ids, position_ids, inputs_embeds):
|
| 46 |
+
input_shape = input_ids.size()
|
| 47 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
| 48 |
+
hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids, inputs_embeds=inputs_embeds)
|
| 49 |
+
return hidden_states
|
| 50 |
+
|
| 51 |
+
def forward_embeddings_with_mapper(self, input_ids, position_ids, mapper_outputs, placeholder_token_ids):
|
| 52 |
+
inputs_embeds = self.embeddings.token_embedding(input_ids)
|
| 53 |
+
dtype = inputs_embeds.dtype
|
| 54 |
+
|
| 55 |
+
offset = defaultdict(int)
|
| 56 |
+
for i, placeholder_token_id in enumerate(placeholder_token_ids):
|
| 57 |
+
# Overwrite the index of the placeholder token with the mapper output for each entry in the batch
|
| 58 |
+
# learnable_idxs = (input_ids == placeholder_token_id).nonzero(as_tuple=True)[1]
|
| 59 |
+
# inputs_embeds[torch.arange(input_ids.shape[0]), learnable_idxs] = mapper_outputs[:, i].to(dtype)
|
| 60 |
+
|
| 61 |
+
for bi in range(input_ids.shape[0]):
|
| 62 |
+
learnable_idx = (input_ids[bi] == placeholder_token_id).nonzero(as_tuple=True)[0]
|
| 63 |
+
|
| 64 |
+
if len(learnable_idx) != 0: # only assign if found
|
| 65 |
+
if len(learnable_idx) == 1:
|
| 66 |
+
offset_learnable_idx = learnable_idx
|
| 67 |
+
else: # if there is two and above.
|
| 68 |
+
start = offset[(bi, placeholder_token_id)]
|
| 69 |
+
offset_learnable_idx = learnable_idx[start:start + 1]
|
| 70 |
+
offset[(bi, placeholder_token_id)] += 1
|
| 71 |
+
|
| 72 |
+
# print(i, offset_learnable_idx)
|
| 73 |
+
|
| 74 |
+
# before = inputs_embeds[bi, learnable_idx]
|
| 75 |
+
inputs_embeds[bi, offset_learnable_idx] = mapper_outputs[bi, i].to(dtype)
|
| 76 |
+
# after = inputs_embeds[bi, learnable_idx]
|
| 77 |
+
|
| 78 |
+
return self.forward_embeddings(input_ids, position_ids, inputs_embeds)
|
| 79 |
+
|
| 80 |
+
def forward(
|
| 81 |
+
self,
|
| 82 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 83 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 84 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 85 |
+
output_attentions: Optional[bool] = None,
|
| 86 |
+
output_hidden_states: Optional[bool] = None,
|
| 87 |
+
return_dict: Optional[bool] = None,
|
| 88 |
+
hidden_states: Optional[torch.Tensor] = None
|
| 89 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 90 |
+
r"""
|
| 91 |
+
Returns:
|
| 92 |
+
|
| 93 |
+
"""
|
| 94 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 95 |
+
output_hidden_states = (
|
| 96 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 97 |
+
)
|
| 98 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 99 |
+
|
| 100 |
+
if input_ids is None:
|
| 101 |
+
raise ValueError("You have to specify either input_ids")
|
| 102 |
+
|
| 103 |
+
input_shape = input_ids.size()
|
| 104 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
| 105 |
+
if hidden_states is None:
|
| 106 |
+
hidden_states = self.embeddings(input_ids=input_ids, position_ids=position_ids)
|
| 107 |
+
|
| 108 |
+
# CLIP's text model uses causal mask, prepare it here.
|
| 109 |
+
# https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
|
| 110 |
+
causal_attention_mask = _create_4d_causal_attention_mask(
|
| 111 |
+
input_shape, hidden_states.dtype, device=hidden_states.device
|
| 112 |
+
)
|
| 113 |
+
# expand attention_mask
|
| 114 |
+
if attention_mask is not None:
|
| 115 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
| 116 |
+
attention_mask = _prepare_4d_attention_mask(attention_mask, hidden_states.dtype)
|
| 117 |
+
|
| 118 |
+
# # bsz, seq_len = input_shape
|
| 119 |
+
# # CLIP's text model uses causal mask, prepare it here.
|
| 120 |
+
# # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
|
| 121 |
+
# causal_attention_mask = _make_causal_mask(input_shape, hidden_states.dtype, device=hidden_states.device)
|
| 122 |
+
# # causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype).to(
|
| 123 |
+
# # hidden_states.device
|
| 124 |
+
# # )
|
| 125 |
+
#
|
| 126 |
+
# # expand attention_mask
|
| 127 |
+
# if attention_mask is not None:
|
| 128 |
+
# # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
| 129 |
+
# attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
|
| 130 |
+
|
| 131 |
+
encoder_outputs = self.encoder(
|
| 132 |
+
inputs_embeds=hidden_states,
|
| 133 |
+
attention_mask=attention_mask,
|
| 134 |
+
causal_attention_mask=causal_attention_mask,
|
| 135 |
+
output_attentions=output_attentions,
|
| 136 |
+
output_hidden_states=output_hidden_states,
|
| 137 |
+
return_dict=return_dict,
|
| 138 |
+
)
|
| 139 |
+
|
| 140 |
+
last_hidden_state = encoder_outputs[0]
|
| 141 |
+
last_hidden_state = self.final_layer_norm(last_hidden_state)
|
| 142 |
+
|
| 143 |
+
# text_embeds.shape = [batch_size, sequence_length, transformer.width]
|
| 144 |
+
# take features from the eot embedding (eot_token is the highest number in each sequence)
|
| 145 |
+
# casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
|
| 146 |
+
pooled_output = last_hidden_state[
|
| 147 |
+
torch.arange(last_hidden_state.shape[0], device=input_ids.device), input_ids.to(torch.int).argmax(dim=-1)
|
| 148 |
+
]
|
| 149 |
+
|
| 150 |
+
if not return_dict:
|
| 151 |
+
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
|
| 152 |
+
|
| 153 |
+
return BaseModelOutputWithPooling(
|
| 154 |
+
last_hidden_state=last_hidden_state,
|
| 155 |
+
pooler_output=pooled_output,
|
| 156 |
+
hidden_states=encoder_outputs.hidden_states,
|
| 157 |
+
attentions=encoder_outputs.attentions,
|
| 158 |
+
)
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
class CustomCLIPTextModelWithProjection(CLIPTextModelWithProjection):
|
| 162 |
+
""" Modification of CLIPTextModel to use our NeTI mapper for computing the embeddings of the concept. """
|
| 163 |
+
|
| 164 |
+
def __init__(self, config: CLIPTextConfig):
|
| 165 |
+
super().__init__(config)
|
| 166 |
+
self.text_model = CustomCLIPTextTransformer(config)
|
| 167 |
+
self.post_init()
|
| 168 |
+
|
| 169 |
+
def forward(self, input_ids: Optional[torch.Tensor] = None,
|
| 170 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 171 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 172 |
+
output_attentions: Optional[bool] = None,
|
| 173 |
+
output_hidden_states: Optional[bool] = None,
|
| 174 |
+
return_dict: Optional[bool] = None,
|
| 175 |
+
hidden_states: Optional[torch.Tensor] = None) -> Union[Tuple, CLIPTextModelOutput]:
|
| 176 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 177 |
+
|
| 178 |
+
text_outputs = self.text_model(
|
| 179 |
+
input_ids=input_ids,
|
| 180 |
+
attention_mask=attention_mask,
|
| 181 |
+
position_ids=position_ids,
|
| 182 |
+
output_attentions=output_attentions,
|
| 183 |
+
output_hidden_states=output_hidden_states,
|
| 184 |
+
return_dict=return_dict,
|
| 185 |
+
hidden_states=hidden_states
|
| 186 |
+
)
|
| 187 |
+
|
| 188 |
+
pooled_output = text_outputs[1]
|
| 189 |
+
|
| 190 |
+
text_embeds = self.text_projection(pooled_output)
|
| 191 |
+
|
| 192 |
+
if not return_dict:
|
| 193 |
+
outputs = (text_embeds, text_outputs[0]) + text_outputs[2:]
|
| 194 |
+
return tuple(output for output in outputs if output is not None)
|
| 195 |
+
|
| 196 |
+
return CLIPTextModelOutput(
|
| 197 |
+
text_embeds=text_embeds,
|
| 198 |
+
last_hidden_state=text_outputs.last_hidden_state,
|
| 199 |
+
hidden_states=text_outputs.hidden_states,
|
| 200 |
+
attentions=text_outputs.attentions,
|
| 201 |
+
)
|
dreamcreature/tokenizer.py
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
The main idea for this code is to provide a way for users to not need to bother with the hassle of multiple tokens for a concept by typing
|
| 3 |
+
a photo of <concept>_0 <concept>_1 ... and so on
|
| 4 |
+
and instead just do
|
| 5 |
+
a photo of <concept>
|
| 6 |
+
which gets translated to the above. This needs to work for both inference and training.
|
| 7 |
+
For inference,
|
| 8 |
+
the tokenizer encodes the text. So, we would want logic for our tokenizer to replace the placeholder token with
|
| 9 |
+
it's underlying vectors
|
| 10 |
+
For training,
|
| 11 |
+
we would want to abstract away some logic like
|
| 12 |
+
1. Adding tokens
|
| 13 |
+
2. Updating gradient mask
|
| 14 |
+
3. Saving embeddings
|
| 15 |
+
to our Util class here.
|
| 16 |
+
so
|
| 17 |
+
TODO:
|
| 18 |
+
1. have tokenizer keep track of concept, multiconcept pairs and replace during encode call x
|
| 19 |
+
2. have mechanism for adding tokens x
|
| 20 |
+
3. have mech for saving emebeddings x
|
| 21 |
+
4. get mask to update x
|
| 22 |
+
5. Loading tokens from embedding x
|
| 23 |
+
6. Integrate to training x
|
| 24 |
+
7. Test
|
| 25 |
+
"""
|
| 26 |
+
import copy
|
| 27 |
+
import random
|
| 28 |
+
|
| 29 |
+
from transformers import CLIPTokenizer
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
class MultiTokenCLIPTokenizer(CLIPTokenizer):
|
| 33 |
+
def __init__(self, *args, **kwargs):
|
| 34 |
+
super().__init__(*args, **kwargs)
|
| 35 |
+
self.token_map = {}
|
| 36 |
+
|
| 37 |
+
def try_adding_tokens(self, placeholder_token, *args, **kwargs):
|
| 38 |
+
num_added_tokens = super().add_tokens(placeholder_token, *args, **kwargs)
|
| 39 |
+
if num_added_tokens == 0:
|
| 40 |
+
raise ValueError(
|
| 41 |
+
f"The tokenizer already contains the token {placeholder_token}. Please pass a different"
|
| 42 |
+
" `placeholder_token` that is not already in the tokenizer."
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
def add_placeholder_tokens(self, placeholder_token, *args, num_vec_per_token=1, **kwargs):
|
| 46 |
+
output = []
|
| 47 |
+
# if num_vec_per_token == 1:
|
| 48 |
+
# self.try_adding_tokens(placeholder_token, *args, **kwargs)
|
| 49 |
+
# output.append(placeholder_token)
|
| 50 |
+
# else:
|
| 51 |
+
output = []
|
| 52 |
+
for i in range(num_vec_per_token):
|
| 53 |
+
ith_token = placeholder_token + f"_{i}"
|
| 54 |
+
self.try_adding_tokens(ith_token, *args, **kwargs)
|
| 55 |
+
output.append(ith_token)
|
| 56 |
+
# handle cases where there is a new placeholder token that contains the current placeholder token but is larger
|
| 57 |
+
for token in self.token_map:
|
| 58 |
+
if token in placeholder_token:
|
| 59 |
+
raise ValueError(
|
| 60 |
+
f"The tokenizer already has placeholder token {token} that can get confused with"
|
| 61 |
+
f" {placeholder_token}keep placeholder tokens independent"
|
| 62 |
+
)
|
| 63 |
+
self.token_map[placeholder_token] = output
|
| 64 |
+
|
| 65 |
+
def replace_placeholder_tokens_in_text(self, text, vector_shuffle=False, prop_tokens_to_load=1.0):
|
| 66 |
+
"""
|
| 67 |
+
Here, we replace the placeholder tokens in text recorded in token_map so that the text_encoder
|
| 68 |
+
can encode them
|
| 69 |
+
vector_shuffle was inspired by https://github.com/rinongal/textual_inversion/pull/119
|
| 70 |
+
where shuffling tokens were found to force the model to learn the concepts more descriptively.
|
| 71 |
+
"""
|
| 72 |
+
if isinstance(text, list):
|
| 73 |
+
output = []
|
| 74 |
+
for i in range(len(text)):
|
| 75 |
+
output.append(self.replace_placeholder_tokens_in_text(text[i], vector_shuffle=vector_shuffle))
|
| 76 |
+
return output
|
| 77 |
+
for placeholder_token in self.token_map:
|
| 78 |
+
if placeholder_token in text:
|
| 79 |
+
tokens = self.token_map[placeholder_token]
|
| 80 |
+
tokens = tokens[: 1 + int(len(tokens) * prop_tokens_to_load)]
|
| 81 |
+
if vector_shuffle:
|
| 82 |
+
tokens = copy.copy(tokens)
|
| 83 |
+
random.shuffle(tokens)
|
| 84 |
+
text = text.replace(placeholder_token, " ".join(tokens)) # <part>_0 <part>_1 -> <part>
|
| 85 |
+
return text
|
| 86 |
+
|
| 87 |
+
def __call__(self, text, *args, vector_shuffle=False, prop_tokens_to_load=1.0, replace_token=True, **kwargs):
|
| 88 |
+
if replace_token:
|
| 89 |
+
return super().__call__(
|
| 90 |
+
self.replace_placeholder_tokens_in_text(
|
| 91 |
+
text, vector_shuffle=vector_shuffle, prop_tokens_to_load=prop_tokens_to_load
|
| 92 |
+
),
|
| 93 |
+
*args,
|
| 94 |
+
**kwargs,
|
| 95 |
+
)
|
| 96 |
+
else:
|
| 97 |
+
return super().__call__(text, *args, **kwargs)
|
| 98 |
+
|
| 99 |
+
def encode(self, text, *args, vector_shuffle=False, prop_tokens_to_load=1.0, replace_token=True, **kwargs):
|
| 100 |
+
if replace_token:
|
| 101 |
+
return super().encode(
|
| 102 |
+
self.replace_placeholder_tokens_in_text(
|
| 103 |
+
text, vector_shuffle=vector_shuffle, prop_tokens_to_load=prop_tokens_to_load
|
| 104 |
+
),
|
| 105 |
+
*args,
|
| 106 |
+
**kwargs,
|
| 107 |
+
)
|
| 108 |
+
else:
|
| 109 |
+
return super().encoder(text, *args, **kwargs)
|
gradio_demo_cub200.py
ADDED
|
@@ -0,0 +1,166 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import gc
|
| 3 |
+
import os
|
| 4 |
+
import re
|
| 5 |
+
import shutil
|
| 6 |
+
|
| 7 |
+
import gradio as gr
|
| 8 |
+
import requests
|
| 9 |
+
import torch
|
| 10 |
+
|
| 11 |
+
from dreamcreature.pipeline import create_args, load_pipeline
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def download_file(url, local_path):
|
| 15 |
+
if os.path.exists(local_path):
|
| 16 |
+
return
|
| 17 |
+
|
| 18 |
+
with requests.get(url, stream=True) as r:
|
| 19 |
+
with open(local_path, 'wb') as f:
|
| 20 |
+
shutil.copyfileobj(r.raw, f)
|
| 21 |
+
|
| 22 |
+
# Example usage
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
parser = argparse.ArgumentParser()
|
| 26 |
+
parser.add_argument('--model_name', default='dreamcreature-sd1.5-cub200')
|
| 27 |
+
parser.add_argument('--checkpoint', default='checkpoint-74900')
|
| 28 |
+
opt = parser.parse_args()
|
| 29 |
+
|
| 30 |
+
model_name = opt.model_name
|
| 31 |
+
checkpoint_name = opt.checkpoint
|
| 32 |
+
|
| 33 |
+
repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
|
| 34 |
+
file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
|
| 35 |
+
local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
|
| 36 |
+
os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
|
| 37 |
+
download_file(file_url, local_path)
|
| 38 |
+
|
| 39 |
+
file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
|
| 40 |
+
local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
|
| 41 |
+
download_file(file_url, local_path)
|
| 42 |
+
|
| 43 |
+
OUTPUT_DIR = model_name
|
| 44 |
+
|
| 45 |
+
args = create_args(OUTPUT_DIR)
|
| 46 |
+
if 'dpo' in OUTPUT_DIR:
|
| 47 |
+
args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"
|
| 48 |
+
|
| 49 |
+
pipe = load_pipeline(args, torch.float16, 'cuda')
|
| 50 |
+
pipe = pipe.to(torch.float16)
|
| 51 |
+
|
| 52 |
+
pipe.verbose = True
|
| 53 |
+
pipe.v = 're'
|
| 54 |
+
pipe.num_k_per_part = 200
|
| 55 |
+
|
| 56 |
+
MAPPING = {
|
| 57 |
+
'body': 0,
|
| 58 |
+
'tail': 1,
|
| 59 |
+
'head': 2,
|
| 60 |
+
'wing': 4,
|
| 61 |
+
'leg': 6
|
| 62 |
+
}
|
| 63 |
+
|
| 64 |
+
ID2NAME = open('data/cub200_2011/class_names.txt').readlines()
|
| 65 |
+
ID2NAME = [line.strip() for line in ID2NAME]
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
def process_text(text):
|
| 69 |
+
pattern = r"<([^:>]+):(\d+)>"
|
| 70 |
+
result = text
|
| 71 |
+
offset = 0
|
| 72 |
+
|
| 73 |
+
part2id = []
|
| 74 |
+
|
| 75 |
+
for match in re.finditer(pattern, text):
|
| 76 |
+
key = match.group(1)
|
| 77 |
+
clsid = int(match.group(2))
|
| 78 |
+
clsid = min(max(clsid, 1), 200) # must be 1~200
|
| 79 |
+
|
| 80 |
+
replacement = f"<{MAPPING[key]}:{clsid - 1}>"
|
| 81 |
+
start, end = match.span()
|
| 82 |
+
|
| 83 |
+
# Adjust the start and end positions based on the offset from previous replacements
|
| 84 |
+
start += offset
|
| 85 |
+
end += offset
|
| 86 |
+
|
| 87 |
+
# Replace the matched text with the replacement
|
| 88 |
+
result = result[:start] + replacement + result[end:]
|
| 89 |
+
|
| 90 |
+
# Update the offset for the next replacement
|
| 91 |
+
offset += len(replacement) - (end - start)
|
| 92 |
+
|
| 93 |
+
part2id.append(f'{key}: {ID2NAME[clsid - 1]}')
|
| 94 |
+
|
| 95 |
+
return result, part2id
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
def generate_images(prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
|
| 99 |
+
generator = torch.Generator(device='cuda')
|
| 100 |
+
generator = generator.manual_seed(int(seed))
|
| 101 |
+
|
| 102 |
+
try:
|
| 103 |
+
prompt, part2id = process_text(prompt)
|
| 104 |
+
negative_prompt, _ = process_text(negative_prompt)
|
| 105 |
+
|
| 106 |
+
images = pipe(prompt,
|
| 107 |
+
negative_prompt=negative_prompt, generator=generator,
|
| 108 |
+
num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
|
| 109 |
+
num_images_per_prompt=num_images).images
|
| 110 |
+
except Exception as e:
|
| 111 |
+
raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
|
| 112 |
+
f"The error message: {e}")
|
| 113 |
+
finally:
|
| 114 |
+
gc.collect()
|
| 115 |
+
torch.cuda.empty_cache()
|
| 116 |
+
|
| 117 |
+
return images, '; '.join(part2id)
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
with gr.Blocks(title="DreamCreature") as demo:
|
| 121 |
+
with gr.Row():
|
| 122 |
+
gr.Markdown(
|
| 123 |
+
"""
|
| 124 |
+
# DreamCreature (CUB-200-2011)
|
| 125 |
+
To create your own creature, you can type:
|
| 126 |
+
|
| 127 |
+
`"a photo of a <head:id> <wing:id> bird"` where `id` ranges from 1~200 (200 classes corresponding to CUB-200-2011)
|
| 128 |
+
|
| 129 |
+
For instance `"a photo of a <head:17> <wing:18> bird"` using head of `cardinal (17)` and wing of `spotted catbird (18)`
|
| 130 |
+
|
| 131 |
+
Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/cub200_2011/class_names.txt
|
| 132 |
+
|
| 133 |
+
You can also try any prompt you like such as:
|
| 134 |
+
|
| 135 |
+
Sub-concept transfer: `"a photo of a <wing:17> cat"`
|
| 136 |
+
|
| 137 |
+
Inspiring design: `"a photo of a <head:101> <wing:191> teddy bear"`
|
| 138 |
+
|
| 139 |
+
(Experimental) You can also use two parts together such as:
|
| 140 |
+
|
| 141 |
+
`"a photo of a <head:17> <head:18> bird"` mixing head of `cardinal (17)` and `spotted catbird (18)`
|
| 142 |
+
|
| 143 |
+
The current available parts are: `head`, `body`, `wing`, `tail`, and `leg`
|
| 144 |
+
|
| 145 |
+
""")
|
| 146 |
+
with gr.Column():
|
| 147 |
+
with gr.Row():
|
| 148 |
+
with gr.Group():
|
| 149 |
+
prompt = gr.Textbox(label="Prompt", value="a photo of a <head:101> <wing:191> teddy bear")
|
| 150 |
+
negative_prompt = gr.Textbox(label="Negative Prompt",
|
| 151 |
+
value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
|
| 152 |
+
num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
|
| 153 |
+
guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
|
| 154 |
+
num_images = gr.Slider(minimum=1, maximum=4, step=1, value=4, label="Number of Images")
|
| 155 |
+
seed = gr.Number(label="Seed", value=777881414)
|
| 156 |
+
button = gr.Button()
|
| 157 |
+
|
| 158 |
+
with gr.Column():
|
| 159 |
+
output_images = gr.Gallery(columns=4, label='Output')
|
| 160 |
+
markdown_labels = gr.Markdown("")
|
| 161 |
+
|
| 162 |
+
button.click(fn=generate_images,
|
| 163 |
+
inputs=[prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
|
| 164 |
+
seed], outputs=[output_images, markdown_labels], show_progress=True)
|
| 165 |
+
|
| 166 |
+
demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')
|
gradio_demo_dog.py
ADDED
|
@@ -0,0 +1,166 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import gc
|
| 3 |
+
import os
|
| 4 |
+
import re
|
| 5 |
+
import shutil
|
| 6 |
+
|
| 7 |
+
import gradio as gr
|
| 8 |
+
import requests
|
| 9 |
+
import torch
|
| 10 |
+
|
| 11 |
+
from dreamcreature.pipeline import create_args, load_pipeline
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def download_file(url, local_path):
|
| 15 |
+
if os.path.exists(local_path):
|
| 16 |
+
return
|
| 17 |
+
|
| 18 |
+
with requests.get(url, stream=True) as r:
|
| 19 |
+
with open(local_path, 'wb') as f:
|
| 20 |
+
shutil.copyfileobj(r.raw, f)
|
| 21 |
+
|
| 22 |
+
# Example usage
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
parser = argparse.ArgumentParser()
|
| 26 |
+
parser.add_argument('--model_name', default='dreamcreature-sd1.5-dog')
|
| 27 |
+
parser.add_argument('--checkpoint', default='checkpoint-150000')
|
| 28 |
+
opt = parser.parse_args()
|
| 29 |
+
|
| 30 |
+
model_name = opt.model_name
|
| 31 |
+
checkpoint_name = opt.checkpoint
|
| 32 |
+
|
| 33 |
+
repo_url = f"https://huggingface.co/kamwoh/{model_name}/resolve/main"
|
| 34 |
+
file_url = repo_url + f"/{checkpoint_name}/pytorch_model.bin"
|
| 35 |
+
local_path = f"{model_name}/{checkpoint_name}/pytorch_model.bin"
|
| 36 |
+
os.makedirs(f"{model_name}/{checkpoint_name}", exist_ok=True)
|
| 37 |
+
download_file(file_url, local_path)
|
| 38 |
+
|
| 39 |
+
file_url = repo_url + f"/{checkpoint_name}/pytorch_model_1.bin"
|
| 40 |
+
local_path = f"{model_name}/{checkpoint_name}/pytorch_model_1.bin"
|
| 41 |
+
download_file(file_url, local_path)
|
| 42 |
+
|
| 43 |
+
OUTPUT_DIR = model_name
|
| 44 |
+
|
| 45 |
+
args = create_args(OUTPUT_DIR)
|
| 46 |
+
if 'dpo' in OUTPUT_DIR:
|
| 47 |
+
args.unet_path = "mhdang/dpo-sd1.5-text2image-v1"
|
| 48 |
+
|
| 49 |
+
pipe = load_pipeline(args, torch.float16, 'cuda')
|
| 50 |
+
pipe = pipe.to(torch.float16)
|
| 51 |
+
|
| 52 |
+
pipe.verbose = True
|
| 53 |
+
pipe.v = 're'
|
| 54 |
+
pipe.num_k_per_part = 120
|
| 55 |
+
|
| 56 |
+
MAPPING = {
|
| 57 |
+
'eye': 0,
|
| 58 |
+
'neck': 2,
|
| 59 |
+
'ear': 3,
|
| 60 |
+
'body': 4,
|
| 61 |
+
'leg': 5,
|
| 62 |
+
'nose': 6,
|
| 63 |
+
'forehead': 7
|
| 64 |
+
}
|
| 65 |
+
|
| 66 |
+
ID2NAME = open('data/dogs/class_names.txt').readlines()
|
| 67 |
+
ID2NAME = [line.strip() for line in ID2NAME]
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def process_text(text):
|
| 71 |
+
pattern = r"<([^:>]+):(\d+)>"
|
| 72 |
+
result = text
|
| 73 |
+
offset = 0
|
| 74 |
+
|
| 75 |
+
part2id = []
|
| 76 |
+
|
| 77 |
+
for match in re.finditer(pattern, text):
|
| 78 |
+
key = match.group(1)
|
| 79 |
+
clsid = int(match.group(2))
|
| 80 |
+
clsid = min(max(clsid, 1), 200) # must be 1~200
|
| 81 |
+
|
| 82 |
+
replacement = f"<{MAPPING[key]}:{clsid - 1}>"
|
| 83 |
+
start, end = match.span()
|
| 84 |
+
|
| 85 |
+
# Adjust the start and end positions based on the offset from previous replacements
|
| 86 |
+
start += offset
|
| 87 |
+
end += offset
|
| 88 |
+
|
| 89 |
+
# Replace the matched text with the replacement
|
| 90 |
+
result = result[:start] + replacement + result[end:]
|
| 91 |
+
|
| 92 |
+
# Update the offset for the next replacement
|
| 93 |
+
offset += len(replacement) - (end - start)
|
| 94 |
+
|
| 95 |
+
part2id.append(f'{key}: {ID2NAME[clsid - 1]}')
|
| 96 |
+
|
| 97 |
+
return result, part2id
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
def generate_images(prompt, negative_prompt, num_inference_steps, guidance_scale, num_images, seed):
|
| 101 |
+
generator = torch.Generator(device='cuda')
|
| 102 |
+
generator = generator.manual_seed(int(seed))
|
| 103 |
+
|
| 104 |
+
try:
|
| 105 |
+
prompt, part2id = process_text(prompt)
|
| 106 |
+
negative_prompt, _ = process_text(negative_prompt)
|
| 107 |
+
|
| 108 |
+
images = pipe(prompt,
|
| 109 |
+
negative_prompt=negative_prompt, generator=generator,
|
| 110 |
+
num_inference_steps=int(num_inference_steps), guidance_scale=guidance_scale,
|
| 111 |
+
num_images_per_prompt=num_images).images
|
| 112 |
+
except Exception as e:
|
| 113 |
+
raise gr.Error(f"Probably due to the prompt have invalid input, please follow the instruction. "
|
| 114 |
+
f"The error message: {e}")
|
| 115 |
+
finally:
|
| 116 |
+
gc.collect()
|
| 117 |
+
torch.cuda.empty_cache()
|
| 118 |
+
|
| 119 |
+
return images, '; '.join(part2id)
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
with gr.Blocks(title="DreamCreature") as demo:
|
| 123 |
+
with gr.Row():
|
| 124 |
+
gr.Markdown(
|
| 125 |
+
"""
|
| 126 |
+
# DreamCreature (Stanford Dogs)
|
| 127 |
+
To create your own creature, you can type:
|
| 128 |
+
|
| 129 |
+
`"a photo of a <nose:id> <ear:id> dog"` where `id` ranges from 0~119 (120 classes corresponding to Stanford Dogs)
|
| 130 |
+
|
| 131 |
+
For instance `"a photo of a <nose:2> <ear:112> dog"` using head of `maltese dog (2)` and wing of `cardigan (112)`
|
| 132 |
+
|
| 133 |
+
Please see `id` in https://github.com/kamwoh/dreamcreature/blob/master/src/data/dogs/class_names.txt
|
| 134 |
+
|
| 135 |
+
Sub-concept transfer: `"a photo of a <ear:112> cat"`
|
| 136 |
+
|
| 137 |
+
Inspiring design: `"a photo of a <eye:38> <body:38> teddy bear"`
|
| 138 |
+
|
| 139 |
+
(Experimental) You can also use two parts together such as:
|
| 140 |
+
|
| 141 |
+
`"a photo of a <nose:1> <nose:112> dog"` mixing head of `maltese dog (2)` and `spotted cardigan (112)`
|
| 142 |
+
|
| 143 |
+
The current available parts are: `eye`, `neck`, `ear`, `body`, `leg`, `nose` and `forehead`
|
| 144 |
+
|
| 145 |
+
""")
|
| 146 |
+
with gr.Column():
|
| 147 |
+
with gr.Row():
|
| 148 |
+
with gr.Group():
|
| 149 |
+
prompt = gr.Textbox(label="Prompt", value="a photo of a <eye:37> <body:37> teddy bear")
|
| 150 |
+
negative_prompt = gr.Textbox(label="Negative Prompt",
|
| 151 |
+
value="blurry, ugly, duplicate, poorly drawn, deformed, mosaic")
|
| 152 |
+
num_inference_steps = gr.Slider(minimum=10, maximum=100, step=1, value=30, label="Num Inference Steps")
|
| 153 |
+
guidance_scale = gr.Slider(minimum=2, maximum=20, step=0.1, value=7.5, label="Guidance Scale")
|
| 154 |
+
num_images = gr.Slider(minimum=1, maximum=4, step=1, value=1, label="Number of Images")
|
| 155 |
+
seed = gr.Number(label="Seed", value=777881414)
|
| 156 |
+
button = gr.Button()
|
| 157 |
+
|
| 158 |
+
with gr.Column():
|
| 159 |
+
output_images = gr.Gallery(columns=4, label='Output')
|
| 160 |
+
markdown_labels = gr.Markdown("")
|
| 161 |
+
|
| 162 |
+
button.click(fn=generate_images,
|
| 163 |
+
inputs=[prompt, negative_prompt, num_inference_steps, guidance_scale, num_images,
|
| 164 |
+
seed], outputs=[output_images, markdown_labels], show_progress=True)
|
| 165 |
+
|
| 166 |
+
demo.queue().launch(inline=False, share=True, debug=True, server_name='0.0.0.0')
|
requirements.txt
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
torch
|
| 2 |
+
torchvision
|
| 3 |
+
git+https://github.com/huggingface/diffusers
|
| 4 |
+
transformers
|
| 5 |
+
torchpq
|
| 6 |
+
omegaconf
|
| 7 |
+
scikit-learn
|
| 8 |
+
faiss-cpu
|
| 9 |
+
tqdm
|
| 10 |
+
accelerate
|
| 11 |
+
gradio
|
| 12 |
+
huggingface_hub
|
run_sd_sup.sh
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
python train_dreamcreature_sd.py \
|
| 2 |
+
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
|
| 3 |
+
--train_data_dir=data/cub200_2011 \
|
| 4 |
+
--resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
|
| 5 |
+
--num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
|
| 6 |
+
--lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sd15-cub200-sup" \
|
| 7 |
+
--validation_prompt="a photo of a 0:16 1:16 2:16 4:16 6:16" \
|
| 8 |
+
--num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
|
| 9 |
+
--code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
|
| 10 |
+
--use_templates --vector_shuffle --snr_gamma=5 \
|
| 11 |
+
--attn_loss=0.01 --use_gt_label --bg_code=7 \
|
| 12 |
+
--resume_from_checkpoint="latest" --mixed_precision="fp16"
|
run_sd_unsup.sh
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
python train_dreamcreature_sd.py \
|
| 2 |
+
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" \
|
| 3 |
+
--train_data_dir=data/cub200_2011 \
|
| 4 |
+
--resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
|
| 5 |
+
--num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
|
| 6 |
+
--lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sd15-cub200-unsup" \
|
| 7 |
+
--validation_prompt="a photo of a 0:16 1:16 2:16 4:16 6:16" \
|
| 8 |
+
--num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
|
| 9 |
+
--code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
|
| 10 |
+
--use_templates --vector_shuffle --snr_gamma=5 \
|
| 11 |
+
--attn_loss=0.01 \
|
| 12 |
+
--resume_from_checkpoint="latest"
|
run_sdxl_sup.sh
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
python train_dreamcreature_sdxl.py \
|
| 2 |
+
--pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
|
| 3 |
+
--scheduler_steps 1000 \
|
| 4 |
+
--train_data_dir=data/cub200_2011 \
|
| 5 |
+
--resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
|
| 6 |
+
--num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
|
| 7 |
+
--lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sdxlbase-cub200-sup" \
|
| 8 |
+
--validation_prompt="a photo of a 0:1 2:1 3:1 4:1 5:1 6:1 7:1" \
|
| 9 |
+
--num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
|
| 10 |
+
--code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
|
| 11 |
+
--use_templates --vector_shuffle --snr_gamma=5 \
|
| 12 |
+
--attn_loss=0.1 --use_gt_label --bg_code=7 \
|
| 13 |
+
--resume_from_checkpoint="latest"
|
run_sdxl_unsup.sh
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
python train_dreamcreature_sdxl.py \
|
| 2 |
+
--pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
|
| 3 |
+
--scheduler_steps 1000 \
|
| 4 |
+
--train_data_dir=data/cub200_2011 \
|
| 5 |
+
--resolution=512 --random_flip --train_batch_size=2 --gradient_accumulation_steps=4 \
|
| 6 |
+
--num_train_epochs=100 --checkpointing_steps=749 --learning_rate=0.0001 \
|
| 7 |
+
--lr_scheduler="constant" --lr_warmup_steps=0 --seed=42 --output_dir="sdxlbase-cub200-unsup" \
|
| 8 |
+
--validation_prompt="a photo of a 0:1 2:1 3:1 4:1 5:1 6:1 7:1" \
|
| 9 |
+
--num_validation_images 8 --num_parts 8 --num_k_per_part 256 --filename="train.txt" \
|
| 10 |
+
--code_filename="train_caps_better_m8_k256.txt" --projection_nlayers=1 \
|
| 11 |
+
--use_templates --vector_shuffle --snr_gamma=5 \
|
| 12 |
+
--attn_loss=0.1 \
|
| 13 |
+
--resume_from_checkpoint="latest"
|
train_dreamcreature_sd.py
ADDED
|
@@ -0,0 +1,1122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""Fine-tuning script for Stable Diffusion for text2image with support for LoRA."""
|
| 16 |
+
|
| 17 |
+
import argparse
|
| 18 |
+
import copy
|
| 19 |
+
import logging
|
| 20 |
+
import math
|
| 21 |
+
import os
|
| 22 |
+
import random
|
| 23 |
+
import shutil
|
| 24 |
+
from pathlib import Path
|
| 25 |
+
|
| 26 |
+
import datasets
|
| 27 |
+
import diffusers
|
| 28 |
+
import numpy as np
|
| 29 |
+
import torch
|
| 30 |
+
import torch.nn.functional as F
|
| 31 |
+
import torch.utils.checkpoint
|
| 32 |
+
import transformers
|
| 33 |
+
from accelerate import Accelerator
|
| 34 |
+
from accelerate.logging import get_logger
|
| 35 |
+
from accelerate.utils import ProjectConfiguration, set_seed
|
| 36 |
+
from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
|
| 37 |
+
from diffusers.loaders import AttnProcsLayers
|
| 38 |
+
from diffusers.optimization import get_scheduler
|
| 39 |
+
from diffusers.training_utils import compute_snr
|
| 40 |
+
from diffusers.utils import check_min_version, is_wandb_available
|
| 41 |
+
from diffusers.utils.import_utils import is_xformers_available
|
| 42 |
+
from huggingface_hub import create_repo, upload_folder
|
| 43 |
+
from packaging import version
|
| 44 |
+
from torchvision import transforms
|
| 45 |
+
from torchvision.transforms import InterpolationMode
|
| 46 |
+
from tqdm.auto import tqdm
|
| 47 |
+
|
| 48 |
+
from dreamcreature.attn_processor import LoRAAttnProcessorCustom
|
| 49 |
+
from dreamcreature.dataset import DreamCreatureDataset
|
| 50 |
+
from dreamcreature.dino import DINO
|
| 51 |
+
from dreamcreature.kmeans_segmentation import KMeansSegmentation
|
| 52 |
+
from dreamcreature.loss import dreamcreature_loss
|
| 53 |
+
from dreamcreature.mapper import TokenMapper
|
| 54 |
+
from dreamcreature.pipeline import DreamCreatureSDPipeline
|
| 55 |
+
from dreamcreature.text_encoder import CustomCLIPTextModel
|
| 56 |
+
from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
|
| 57 |
+
from utils import add_tokens, tokenize_prompt, get_attn_processors
|
| 58 |
+
|
| 59 |
+
imagenet_templates = [
|
| 60 |
+
"a photo of a {}",
|
| 61 |
+
"a rendering of a {}",
|
| 62 |
+
"a cropped photo of the {}",
|
| 63 |
+
"the photo of a {}",
|
| 64 |
+
"a photo of a clean {}",
|
| 65 |
+
"a photo of a dirty {}",
|
| 66 |
+
"a dark photo of the {}",
|
| 67 |
+
"a photo of my {}",
|
| 68 |
+
"a photo of the cool {}",
|
| 69 |
+
"a close-up photo of a {}",
|
| 70 |
+
"a bright photo of the {}",
|
| 71 |
+
"a cropped photo of a {}",
|
| 72 |
+
"a photo of the {}",
|
| 73 |
+
"a good photo of the {}",
|
| 74 |
+
"a photo of one {}",
|
| 75 |
+
"a close-up photo of the {}",
|
| 76 |
+
"a rendition of the {}",
|
| 77 |
+
"a photo of the clean {}",
|
| 78 |
+
"a rendition of a {}",
|
| 79 |
+
"a photo of a nice {}",
|
| 80 |
+
"a good photo of a {}",
|
| 81 |
+
"a photo of the nice {}",
|
| 82 |
+
"a photo of the small {}",
|
| 83 |
+
"a photo of the weird {}",
|
| 84 |
+
"a photo of the large {}",
|
| 85 |
+
"a photo of a cool {}",
|
| 86 |
+
"a photo of a small {}",
|
| 87 |
+
]
|
| 88 |
+
|
| 89 |
+
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
| 90 |
+
check_min_version("0.21.0.dev0")
|
| 91 |
+
|
| 92 |
+
logger = get_logger(__name__, log_level="INFO")
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None):
|
| 96 |
+
img_str = ""
|
| 97 |
+
for i, image in enumerate(images):
|
| 98 |
+
image.save(os.path.join(repo_folder, f"image_{i}.png"))
|
| 99 |
+
img_str += f"\n"
|
| 100 |
+
|
| 101 |
+
yaml = f"""
|
| 102 |
+
---
|
| 103 |
+
license: creativeml-openrail-m
|
| 104 |
+
base_model: {base_model}
|
| 105 |
+
tags:
|
| 106 |
+
- stable-diffusion
|
| 107 |
+
- stable-diffusion-diffusers
|
| 108 |
+
- text-to-image
|
| 109 |
+
- diffusers
|
| 110 |
+
- lora
|
| 111 |
+
inference: true
|
| 112 |
+
---
|
| 113 |
+
"""
|
| 114 |
+
model_card = f"""
|
| 115 |
+
# LoRA text2image fine-tuning - {repo_id}
|
| 116 |
+
These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
|
| 117 |
+
{img_str}
|
| 118 |
+
"""
|
| 119 |
+
with open(os.path.join(repo_folder, "README.md"), "w") as f:
|
| 120 |
+
f.write(yaml + model_card)
|
| 121 |
+
|
| 122 |
+
|
| 123 |
+
def parse_args():
|
| 124 |
+
parser = argparse.ArgumentParser(description="Simple example of a training script.")
|
| 125 |
+
parser.add_argument(
|
| 126 |
+
"--pretrained_model_name_or_path",
|
| 127 |
+
type=str,
|
| 128 |
+
default=None,
|
| 129 |
+
required=True,
|
| 130 |
+
help="Path to pretrained model or model identifier from huggingface.co/models.",
|
| 131 |
+
)
|
| 132 |
+
parser.add_argument(
|
| 133 |
+
"--revision",
|
| 134 |
+
type=str,
|
| 135 |
+
default=None,
|
| 136 |
+
required=False,
|
| 137 |
+
help="Revision of pretrained model identifier from huggingface.co/models.",
|
| 138 |
+
)
|
| 139 |
+
parser.add_argument(
|
| 140 |
+
"--dataset_name",
|
| 141 |
+
type=str,
|
| 142 |
+
default=None,
|
| 143 |
+
help=(
|
| 144 |
+
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
|
| 145 |
+
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
|
| 146 |
+
" or to a folder containing files that 🤗 Datasets can understand."
|
| 147 |
+
),
|
| 148 |
+
)
|
| 149 |
+
parser.add_argument(
|
| 150 |
+
"--dataset_config_name",
|
| 151 |
+
type=str,
|
| 152 |
+
default=None,
|
| 153 |
+
help="The config of the Dataset, leave as None if there's only one config.",
|
| 154 |
+
)
|
| 155 |
+
parser.add_argument(
|
| 156 |
+
"--train_data_dir",
|
| 157 |
+
type=str,
|
| 158 |
+
default=None,
|
| 159 |
+
help=(
|
| 160 |
+
"A folder containing the training data. Folder contents must follow the structure described in"
|
| 161 |
+
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
|
| 162 |
+
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
|
| 163 |
+
),
|
| 164 |
+
)
|
| 165 |
+
parser.add_argument(
|
| 166 |
+
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
|
| 167 |
+
)
|
| 168 |
+
parser.add_argument(
|
| 169 |
+
"--caption_column",
|
| 170 |
+
type=str,
|
| 171 |
+
default="text",
|
| 172 |
+
help="The column of the dataset containing a caption or a list of captions.",
|
| 173 |
+
)
|
| 174 |
+
parser.add_argument(
|
| 175 |
+
"--validation_prompt", type=str, default=None, help="A prompt that is sampled during training for inference."
|
| 176 |
+
)
|
| 177 |
+
parser.add_argument(
|
| 178 |
+
"--num_validation_images",
|
| 179 |
+
type=int,
|
| 180 |
+
default=4,
|
| 181 |
+
help="Number of images that should be generated during validation with `validation_prompt`.",
|
| 182 |
+
)
|
| 183 |
+
parser.add_argument(
|
| 184 |
+
"--validation_epochs",
|
| 185 |
+
type=int,
|
| 186 |
+
default=1,
|
| 187 |
+
help=(
|
| 188 |
+
"Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
|
| 189 |
+
" `args.validation_prompt` multiple times: `args.num_validation_images`."
|
| 190 |
+
),
|
| 191 |
+
)
|
| 192 |
+
parser.add_argument(
|
| 193 |
+
"--max_train_samples",
|
| 194 |
+
type=int,
|
| 195 |
+
default=None,
|
| 196 |
+
help=(
|
| 197 |
+
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
| 198 |
+
"value if set."
|
| 199 |
+
),
|
| 200 |
+
)
|
| 201 |
+
parser.add_argument(
|
| 202 |
+
"--output_dir",
|
| 203 |
+
type=str,
|
| 204 |
+
default="sd-model-finetuned-lora",
|
| 205 |
+
help="The output directory where the model predictions and checkpoints will be written.",
|
| 206 |
+
)
|
| 207 |
+
parser.add_argument(
|
| 208 |
+
"--cache_dir",
|
| 209 |
+
type=str,
|
| 210 |
+
default=None,
|
| 211 |
+
help="The directory where the downloaded models and datasets will be stored.",
|
| 212 |
+
)
|
| 213 |
+
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
|
| 214 |
+
parser.add_argument(
|
| 215 |
+
"--resolution",
|
| 216 |
+
type=int,
|
| 217 |
+
default=512,
|
| 218 |
+
help=(
|
| 219 |
+
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
|
| 220 |
+
" resolution"
|
| 221 |
+
),
|
| 222 |
+
)
|
| 223 |
+
parser.add_argument(
|
| 224 |
+
"--center_crop",
|
| 225 |
+
default=False,
|
| 226 |
+
action="store_true",
|
| 227 |
+
help=(
|
| 228 |
+
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
|
| 229 |
+
" cropped. The images will be resized to the resolution first before cropping."
|
| 230 |
+
),
|
| 231 |
+
)
|
| 232 |
+
parser.add_argument(
|
| 233 |
+
"--random_flip",
|
| 234 |
+
action="store_true",
|
| 235 |
+
help="whether to randomly flip images horizontally",
|
| 236 |
+
)
|
| 237 |
+
parser.add_argument(
|
| 238 |
+
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
|
| 239 |
+
)
|
| 240 |
+
parser.add_argument("--num_train_epochs", type=int, default=100)
|
| 241 |
+
parser.add_argument(
|
| 242 |
+
"--max_train_steps",
|
| 243 |
+
type=int,
|
| 244 |
+
default=None,
|
| 245 |
+
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
|
| 246 |
+
)
|
| 247 |
+
parser.add_argument(
|
| 248 |
+
"--gradient_accumulation_steps",
|
| 249 |
+
type=int,
|
| 250 |
+
default=1,
|
| 251 |
+
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
| 252 |
+
)
|
| 253 |
+
parser.add_argument(
|
| 254 |
+
"--gradient_checkpointing",
|
| 255 |
+
action="store_true",
|
| 256 |
+
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
|
| 257 |
+
)
|
| 258 |
+
parser.add_argument(
|
| 259 |
+
"--learning_rate",
|
| 260 |
+
type=float,
|
| 261 |
+
default=1e-4,
|
| 262 |
+
help="Initial learning rate (after the potential warmup period) to use.",
|
| 263 |
+
)
|
| 264 |
+
parser.add_argument(
|
| 265 |
+
"--scale_lr",
|
| 266 |
+
action="store_true",
|
| 267 |
+
default=False,
|
| 268 |
+
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
|
| 269 |
+
)
|
| 270 |
+
parser.add_argument(
|
| 271 |
+
"--lr_scheduler",
|
| 272 |
+
type=str,
|
| 273 |
+
default="constant",
|
| 274 |
+
help=(
|
| 275 |
+
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
|
| 276 |
+
' "constant", "constant_with_warmup"]'
|
| 277 |
+
),
|
| 278 |
+
)
|
| 279 |
+
parser.add_argument(
|
| 280 |
+
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
|
| 281 |
+
)
|
| 282 |
+
parser.add_argument(
|
| 283 |
+
"--snr_gamma",
|
| 284 |
+
type=float,
|
| 285 |
+
default=None,
|
| 286 |
+
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
|
| 287 |
+
"More details here: https://arxiv.org/abs/2303.09556.",
|
| 288 |
+
)
|
| 289 |
+
parser.add_argument(
|
| 290 |
+
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
|
| 291 |
+
)
|
| 292 |
+
parser.add_argument(
|
| 293 |
+
"--allow_tf32",
|
| 294 |
+
action="store_true",
|
| 295 |
+
help=(
|
| 296 |
+
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
|
| 297 |
+
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
|
| 298 |
+
),
|
| 299 |
+
)
|
| 300 |
+
parser.add_argument(
|
| 301 |
+
"--dataloader_num_workers",
|
| 302 |
+
type=int,
|
| 303 |
+
default=0,
|
| 304 |
+
help=(
|
| 305 |
+
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
| 306 |
+
),
|
| 307 |
+
)
|
| 308 |
+
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
|
| 309 |
+
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
|
| 310 |
+
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
|
| 311 |
+
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
|
| 312 |
+
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
|
| 313 |
+
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
|
| 314 |
+
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
|
| 315 |
+
parser.add_argument(
|
| 316 |
+
"--prediction_type",
|
| 317 |
+
type=str,
|
| 318 |
+
default=None,
|
| 319 |
+
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
|
| 320 |
+
)
|
| 321 |
+
parser.add_argument(
|
| 322 |
+
"--hub_model_id",
|
| 323 |
+
type=str,
|
| 324 |
+
default=None,
|
| 325 |
+
help="The name of the repository to keep in sync with the local `output_dir`.",
|
| 326 |
+
)
|
| 327 |
+
parser.add_argument(
|
| 328 |
+
"--logging_dir",
|
| 329 |
+
type=str,
|
| 330 |
+
default="logs",
|
| 331 |
+
help=(
|
| 332 |
+
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
|
| 333 |
+
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
|
| 334 |
+
),
|
| 335 |
+
)
|
| 336 |
+
parser.add_argument(
|
| 337 |
+
"--mixed_precision",
|
| 338 |
+
type=str,
|
| 339 |
+
default=None,
|
| 340 |
+
choices=["no", "fp16", "bf16"],
|
| 341 |
+
help=(
|
| 342 |
+
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
| 343 |
+
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
|
| 344 |
+
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
|
| 345 |
+
),
|
| 346 |
+
)
|
| 347 |
+
parser.add_argument(
|
| 348 |
+
"--report_to",
|
| 349 |
+
type=str,
|
| 350 |
+
default="tensorboard",
|
| 351 |
+
help=(
|
| 352 |
+
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
|
| 353 |
+
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
|
| 354 |
+
),
|
| 355 |
+
)
|
| 356 |
+
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
| 357 |
+
parser.add_argument(
|
| 358 |
+
"--checkpointing_steps",
|
| 359 |
+
type=int,
|
| 360 |
+
default=500,
|
| 361 |
+
help=(
|
| 362 |
+
"Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
|
| 363 |
+
" training using `--resume_from_checkpoint`."
|
| 364 |
+
),
|
| 365 |
+
)
|
| 366 |
+
parser.add_argument(
|
| 367 |
+
"--checkpoints_total_limit",
|
| 368 |
+
type=int,
|
| 369 |
+
default=None,
|
| 370 |
+
help=("Max number of checkpoints to store."),
|
| 371 |
+
)
|
| 372 |
+
parser.add_argument(
|
| 373 |
+
"--resume_from_checkpoint",
|
| 374 |
+
type=str,
|
| 375 |
+
default=None,
|
| 376 |
+
help=(
|
| 377 |
+
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
|
| 378 |
+
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
|
| 379 |
+
),
|
| 380 |
+
)
|
| 381 |
+
parser.add_argument(
|
| 382 |
+
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
|
| 383 |
+
)
|
| 384 |
+
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
|
| 385 |
+
parser.add_argument(
|
| 386 |
+
"--rank",
|
| 387 |
+
type=int,
|
| 388 |
+
default=4,
|
| 389 |
+
help=("The dimension of the LoRA update matrices."),
|
| 390 |
+
)
|
| 391 |
+
|
| 392 |
+
parser.add_argument('--filename', default='train.txt')
|
| 393 |
+
parser.add_argument('--code_filename', default='train_caps_better_m8_k256.txt')
|
| 394 |
+
parser.add_argument('--repeat', default=1, type=int)
|
| 395 |
+
|
| 396 |
+
parser.add_argument('--scheduler_steps', default=1000, type=int, help='scheduler step, if turbo, set to 4')
|
| 397 |
+
parser.add_argument('--num_parts', type=int, default=4, help="Number of parts")
|
| 398 |
+
parser.add_argument('--num_k_per_part', type=int, default=256, help='Number of k')
|
| 399 |
+
|
| 400 |
+
parser.add_argument('--mapper_lr_scale', default=1, type=float)
|
| 401 |
+
parser.add_argument('--mapper_lr', default=0.0001, type=float)
|
| 402 |
+
parser.add_argument('--attn_loss', default=0, type=float)
|
| 403 |
+
parser.add_argument('--projection_nlayers', default=3, type=int)
|
| 404 |
+
|
| 405 |
+
parser.add_argument('--masked_training', action='store_true')
|
| 406 |
+
parser.add_argument('--drop_tokens', action='store_true')
|
| 407 |
+
parser.add_argument('--drop_rate', type=float, default=0.5)
|
| 408 |
+
parser.add_argument('--drop_counts', default='half')
|
| 409 |
+
|
| 410 |
+
parser.add_argument('--class_name', default='')
|
| 411 |
+
parser.add_argument('--no_pe', action='store_true')
|
| 412 |
+
parser.add_argument('--vector_shuffle', action='store_true')
|
| 413 |
+
parser.add_argument('--use_templates', action='store_true')
|
| 414 |
+
|
| 415 |
+
parser.add_argument('--use_gt_label', action='store_true')
|
| 416 |
+
parser.add_argument('--bg_code', default=7, type=int) # for gt_label
|
| 417 |
+
parser.add_argument('--fg_idx', default=0, type=int) # for gt_label
|
| 418 |
+
|
| 419 |
+
parser.add_argument('--filter_class', default=None, type=int, help='debugging purpose')
|
| 420 |
+
|
| 421 |
+
parser.add_argument('--unet_path', default=None)
|
| 422 |
+
|
| 423 |
+
args = parser.parse_args()
|
| 424 |
+
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
|
| 425 |
+
if env_local_rank != -1 and env_local_rank != args.local_rank:
|
| 426 |
+
args.local_rank = env_local_rank
|
| 427 |
+
|
| 428 |
+
# Sanity checks
|
| 429 |
+
if args.dataset_name is None and args.train_data_dir is None:
|
| 430 |
+
raise ValueError("Need either a dataset name or a training folder.")
|
| 431 |
+
|
| 432 |
+
return args
|
| 433 |
+
|
| 434 |
+
|
| 435 |
+
def collate_fn(args, tokenizer, placeholder_token):
|
| 436 |
+
train_resizecrop = transforms.Compose([
|
| 437 |
+
transforms.Resize(int(args.resolution), InterpolationMode.BILINEAR),
|
| 438 |
+
transforms.RandomCrop(args.resolution),
|
| 439 |
+
])
|
| 440 |
+
|
| 441 |
+
train_transforms = transforms.Compose(
|
| 442 |
+
[
|
| 443 |
+
transforms.ToTensor(),
|
| 444 |
+
transforms.Normalize([0.5], [0.5]),
|
| 445 |
+
]
|
| 446 |
+
)
|
| 447 |
+
|
| 448 |
+
def f(examples):
|
| 449 |
+
raw_images = [train_resizecrop(example["pixel_values"]) for example in examples]
|
| 450 |
+
|
| 451 |
+
pixel_values = torch.stack([train_transforms(image) for image in raw_images])
|
| 452 |
+
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
|
| 453 |
+
|
| 454 |
+
captions = []
|
| 455 |
+
appeared_tokens = []
|
| 456 |
+
|
| 457 |
+
for i in range(len(examples)):
|
| 458 |
+
if args.use_templates and random.random() <= 0.5: # 50% using templates
|
| 459 |
+
if args.class_name != '':
|
| 460 |
+
caption = random.choice(imagenet_templates).format(f'{placeholder_token} {args.class_name}')
|
| 461 |
+
else:
|
| 462 |
+
caption = random.choice(imagenet_templates).format(placeholder_token)
|
| 463 |
+
else:
|
| 464 |
+
if args.class_name != '':
|
| 465 |
+
caption = f'{placeholder_token} {args.class_name}'
|
| 466 |
+
else:
|
| 467 |
+
caption = placeholder_token
|
| 468 |
+
|
| 469 |
+
tokens = tokenizer.token_map[placeholder_token][:args.num_parts]
|
| 470 |
+
tokens = [tokens[a] for a in examples[i]['appeared']]
|
| 471 |
+
|
| 472 |
+
if args.vector_shuffle or args.drop_tokens:
|
| 473 |
+
tokens = copy.copy(tokens)
|
| 474 |
+
random.shuffle(tokens)
|
| 475 |
+
|
| 476 |
+
if args.drop_tokens and random.random() < args.drop_rate and len(tokens) >= 2:
|
| 477 |
+
# randomly drop half of the tokens
|
| 478 |
+
if args.drop_counts == 'half':
|
| 479 |
+
tokens = tokens[:len(tokens) // 2]
|
| 480 |
+
else:
|
| 481 |
+
tokens = tokens[:int(args.drop_counts)]
|
| 482 |
+
|
| 483 |
+
appeared = [int(t.split('_')[1]) for t in tokens] # <part>_i
|
| 484 |
+
appeared_tokens.append(appeared)
|
| 485 |
+
|
| 486 |
+
caption = caption.replace(placeholder_token, ' '.join(tokens))
|
| 487 |
+
captions.append(caption)
|
| 488 |
+
|
| 489 |
+
input_ids = tokenize_prompt(tokenizer, captions)
|
| 490 |
+
# input_ids = inputs.input_ids.repeat(len(examples), 1) # (1, 77) -> (B, 77)
|
| 491 |
+
|
| 492 |
+
codes = torch.stack([example["codes"] for example in examples])
|
| 493 |
+
|
| 494 |
+
return {"pixel_values": pixel_values,
|
| 495 |
+
"raw_images": raw_images,
|
| 496 |
+
"appeared_tokens": appeared_tokens,
|
| 497 |
+
"input_ids": input_ids,
|
| 498 |
+
"codes": codes}
|
| 499 |
+
|
| 500 |
+
return f
|
| 501 |
+
|
| 502 |
+
|
| 503 |
+
def setup_attn_processor(unet, **kwargs):
|
| 504 |
+
lora_attn_procs = {}
|
| 505 |
+
for name in unet.attn_processors.keys():
|
| 506 |
+
cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
|
| 507 |
+
if name.startswith("mid_block"):
|
| 508 |
+
hidden_size = unet.config.block_out_channels[-1]
|
| 509 |
+
elif name.startswith("up_blocks"):
|
| 510 |
+
block_id = int(name[len("up_blocks.")])
|
| 511 |
+
hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
|
| 512 |
+
elif name.startswith("down_blocks"):
|
| 513 |
+
block_id = int(name[len("down_blocks.")])
|
| 514 |
+
hidden_size = unet.config.block_out_channels[block_id]
|
| 515 |
+
|
| 516 |
+
lora_attn_procs[name] = LoRAAttnProcessorCustom(
|
| 517 |
+
hidden_size=hidden_size,
|
| 518 |
+
cross_attention_dim=cross_attention_dim,
|
| 519 |
+
rank=kwargs['rank'],
|
| 520 |
+
)
|
| 521 |
+
|
| 522 |
+
unet.set_attn_processor(lora_attn_procs)
|
| 523 |
+
|
| 524 |
+
|
| 525 |
+
def load_attn_processor(unet, filename):
|
| 526 |
+
logger.info(f'Load attn processors from {filename}')
|
| 527 |
+
lora_layers = AttnProcsLayers(get_attn_processors(unet))
|
| 528 |
+
lora_layers.load_state_dict(torch.load(filename))
|
| 529 |
+
|
| 530 |
+
|
| 531 |
+
def main():
|
| 532 |
+
args = parse_args()
|
| 533 |
+
logging_dir = Path(args.output_dir, args.logging_dir)
|
| 534 |
+
|
| 535 |
+
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
|
| 536 |
+
|
| 537 |
+
accelerator = Accelerator(
|
| 538 |
+
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
| 539 |
+
mixed_precision=args.mixed_precision,
|
| 540 |
+
log_with=args.report_to,
|
| 541 |
+
project_config=accelerator_project_config,
|
| 542 |
+
)
|
| 543 |
+
if args.report_to == "wandb":
|
| 544 |
+
if not is_wandb_available():
|
| 545 |
+
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
|
| 546 |
+
import wandb
|
| 547 |
+
|
| 548 |
+
# Make one log on every process with the configuration for debugging.
|
| 549 |
+
logging.basicConfig(
|
| 550 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
| 551 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
| 552 |
+
level=logging.INFO,
|
| 553 |
+
)
|
| 554 |
+
logger.info(accelerator.state, main_process_only=False)
|
| 555 |
+
if accelerator.is_local_main_process:
|
| 556 |
+
datasets.utils.logging.set_verbosity_warning()
|
| 557 |
+
transformers.utils.logging.set_verbosity_warning()
|
| 558 |
+
diffusers.utils.logging.set_verbosity_info()
|
| 559 |
+
else:
|
| 560 |
+
datasets.utils.logging.set_verbosity_error()
|
| 561 |
+
transformers.utils.logging.set_verbosity_error()
|
| 562 |
+
diffusers.utils.logging.set_verbosity_error()
|
| 563 |
+
|
| 564 |
+
# If passed along, set the training seed now.
|
| 565 |
+
if args.seed is not None:
|
| 566 |
+
set_seed(args.seed)
|
| 567 |
+
|
| 568 |
+
# Handle the repository creation
|
| 569 |
+
if accelerator.is_main_process:
|
| 570 |
+
if args.output_dir is not None:
|
| 571 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
| 572 |
+
|
| 573 |
+
if args.push_to_hub:
|
| 574 |
+
repo_id = create_repo(
|
| 575 |
+
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
|
| 576 |
+
).repo_id
|
| 577 |
+
# Load scheduler, tokenizer and models.
|
| 578 |
+
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
|
| 579 |
+
tokenizer = MultiTokenCLIPTokenizer.from_pretrained(
|
| 580 |
+
args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
|
| 581 |
+
)
|
| 582 |
+
|
| 583 |
+
OUT_DIMS = 1024 if 'stabilityai/stable-diffusion-2-1' in args.pretrained_model_name_or_path else 768
|
| 584 |
+
|
| 585 |
+
text_encoder = CustomCLIPTextModel.from_pretrained(
|
| 586 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
|
| 587 |
+
)
|
| 588 |
+
vae = AutoencoderKL.from_pretrained(args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision)
|
| 589 |
+
|
| 590 |
+
unet_path = args.unet_path if args.unet_path is not None else args.pretrained_model_name_or_path
|
| 591 |
+
unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
|
| 592 |
+
unet_path, subfolder="unet", revision=args.revision
|
| 593 |
+
)
|
| 594 |
+
|
| 595 |
+
dino = DINO()
|
| 596 |
+
seg = KMeansSegmentation(args.train_data_dir + '/pretrained_kmeans.pth',
|
| 597 |
+
args.fg_idx,
|
| 598 |
+
args.bg_code,
|
| 599 |
+
args.num_parts,
|
| 600 |
+
args.num_k_per_part)
|
| 601 |
+
|
| 602 |
+
simple_mapper = TokenMapper(args.num_parts,
|
| 603 |
+
args.num_k_per_part,
|
| 604 |
+
OUT_DIMS,
|
| 605 |
+
args.projection_nlayers)
|
| 606 |
+
# initialize placeholder token
|
| 607 |
+
placeholder_token = "<part>"
|
| 608 |
+
initializer_token = None
|
| 609 |
+
placeholder_token_ids = add_tokens(tokenizer,
|
| 610 |
+
text_encoder,
|
| 611 |
+
placeholder_token,
|
| 612 |
+
args.num_parts,
|
| 613 |
+
initializer_token)
|
| 614 |
+
|
| 615 |
+
# freeze parameters of models to save more memory
|
| 616 |
+
unet.requires_grad_(False)
|
| 617 |
+
vae.requires_grad_(False)
|
| 618 |
+
text_encoder.requires_grad_(False)
|
| 619 |
+
dino.requires_grad_(False)
|
| 620 |
+
|
| 621 |
+
# For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
|
| 622 |
+
# as these weights are only used for inference, keeping weights in full precision is not required.
|
| 623 |
+
weight_dtype = torch.float32
|
| 624 |
+
if accelerator.mixed_precision == "fp16":
|
| 625 |
+
weight_dtype = torch.float16
|
| 626 |
+
elif accelerator.mixed_precision == "bf16":
|
| 627 |
+
weight_dtype = torch.bfloat16
|
| 628 |
+
|
| 629 |
+
# Move unet, vae and text_encoder to device and cast to weight_dtype
|
| 630 |
+
unet.to(accelerator.device, dtype=weight_dtype)
|
| 631 |
+
vae.to(accelerator.device, dtype=weight_dtype)
|
| 632 |
+
text_encoder.to(accelerator.device, dtype=weight_dtype)
|
| 633 |
+
|
| 634 |
+
# now we will add new LoRA weights to the attention layers
|
| 635 |
+
# It's important to realize here how many attention weights will be added and of which sizes
|
| 636 |
+
# The sizes of the attention layers consist only of two different variables:
|
| 637 |
+
# 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
|
| 638 |
+
# 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.
|
| 639 |
+
|
| 640 |
+
# Let's first see how many attention processors we will have to set.
|
| 641 |
+
# For Stable Diffusion, it should be equal to:
|
| 642 |
+
# - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
|
| 643 |
+
# - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
|
| 644 |
+
# - up blocks (2x attention layers) * (3x transformer layers) * (3x down blocks) = 18
|
| 645 |
+
# => 32 layers
|
| 646 |
+
|
| 647 |
+
# Set correct lora layers
|
| 648 |
+
setup_attn_processor(unet, rank=args.rank)
|
| 649 |
+
|
| 650 |
+
if args.enable_xformers_memory_efficient_attention:
|
| 651 |
+
if is_xformers_available():
|
| 652 |
+
import xformers
|
| 653 |
+
|
| 654 |
+
xformers_version = version.parse(xformers.__version__)
|
| 655 |
+
if xformers_version == version.parse("0.0.16"):
|
| 656 |
+
logger.warn(
|
| 657 |
+
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
|
| 658 |
+
)
|
| 659 |
+
unet.enable_xformers_memory_efficient_attention()
|
| 660 |
+
else:
|
| 661 |
+
raise ValueError("xformers is not available. Make sure it is installed correctly")
|
| 662 |
+
|
| 663 |
+
lora_layers = AttnProcsLayers(get_attn_processors(unet))
|
| 664 |
+
|
| 665 |
+
# Enable TF32 for faster training on Ampere GPUs,
|
| 666 |
+
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
| 667 |
+
if args.allow_tf32:
|
| 668 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 669 |
+
|
| 670 |
+
if args.scale_lr:
|
| 671 |
+
args.learning_rate = (
|
| 672 |
+
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
| 673 |
+
)
|
| 674 |
+
|
| 675 |
+
# Initialize the optimizer
|
| 676 |
+
if args.use_8bit_adam:
|
| 677 |
+
try:
|
| 678 |
+
import bitsandbytes as bnb
|
| 679 |
+
except ImportError:
|
| 680 |
+
raise ImportError(
|
| 681 |
+
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
|
| 682 |
+
)
|
| 683 |
+
|
| 684 |
+
optimizer_cls = bnb.optim.AdamW8bit
|
| 685 |
+
else:
|
| 686 |
+
optimizer_cls = torch.optim.AdamW
|
| 687 |
+
|
| 688 |
+
extra_params = list(simple_mapper.parameters())
|
| 689 |
+
mapper_lr = args.learning_rate * args.mapper_lr_scale if args.learning_rate != 0 else args.mapper_lr
|
| 690 |
+
|
| 691 |
+
optimizer = optimizer_cls(
|
| 692 |
+
[{'params': lora_layers.parameters()},
|
| 693 |
+
{'params': extra_params, 'lr': mapper_lr}],
|
| 694 |
+
lr=args.learning_rate,
|
| 695 |
+
betas=(args.adam_beta1, args.adam_beta2),
|
| 696 |
+
weight_decay=args.adam_weight_decay,
|
| 697 |
+
eps=args.adam_epsilon,
|
| 698 |
+
)
|
| 699 |
+
|
| 700 |
+
train_dataset = DreamCreatureDataset(args.train_data_dir,
|
| 701 |
+
args.filename,
|
| 702 |
+
code_filename=args.code_filename,
|
| 703 |
+
num_parts=args.num_parts,
|
| 704 |
+
num_k_per_part=args.num_k_per_part,
|
| 705 |
+
use_gt_label=args.use_gt_label,
|
| 706 |
+
bg_code=args.bg_code,
|
| 707 |
+
repeat=args.repeat)
|
| 708 |
+
|
| 709 |
+
with accelerator.main_process_first():
|
| 710 |
+
if args.max_train_samples is not None:
|
| 711 |
+
train_dataset.set_max_samples(args.max_train_samples, args.seed)
|
| 712 |
+
|
| 713 |
+
# DataLoaders creation:
|
| 714 |
+
train_dataloader = torch.utils.data.DataLoader(
|
| 715 |
+
train_dataset,
|
| 716 |
+
shuffle=True,
|
| 717 |
+
collate_fn=collate_fn(args, tokenizer, placeholder_token),
|
| 718 |
+
batch_size=args.train_batch_size,
|
| 719 |
+
num_workers=args.dataloader_num_workers,
|
| 720 |
+
)
|
| 721 |
+
|
| 722 |
+
# Scheduler and math around the number of training steps.
|
| 723 |
+
overrode_max_train_steps = False
|
| 724 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 725 |
+
if args.max_train_steps is None:
|
| 726 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 727 |
+
overrode_max_train_steps = True
|
| 728 |
+
|
| 729 |
+
lr_scheduler = get_scheduler(
|
| 730 |
+
args.lr_scheduler,
|
| 731 |
+
optimizer=optimizer,
|
| 732 |
+
num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
|
| 733 |
+
num_training_steps=args.max_train_steps * accelerator.num_processes,
|
| 734 |
+
)
|
| 735 |
+
|
| 736 |
+
# Prepare everything with our `accelerator`.
|
| 737 |
+
lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
| 738 |
+
lora_layers, optimizer, train_dataloader, lr_scheduler
|
| 739 |
+
)
|
| 740 |
+
simple_mapper = accelerator.prepare(simple_mapper)
|
| 741 |
+
|
| 742 |
+
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
|
| 743 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 744 |
+
if overrode_max_train_steps:
|
| 745 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 746 |
+
# Afterwards we recalculate our number of training epochs
|
| 747 |
+
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
| 748 |
+
|
| 749 |
+
# We need to initialize the trackers we use, and also store our configuration.
|
| 750 |
+
# The trackers initializes automatically on the main process.
|
| 751 |
+
if accelerator.is_main_process:
|
| 752 |
+
accelerator.init_trackers("text2image-fine-tune", config=vars(args))
|
| 753 |
+
|
| 754 |
+
# Train!
|
| 755 |
+
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
| 756 |
+
|
| 757 |
+
logger.info("***** Running training *****")
|
| 758 |
+
logger.info(f" Num examples = {len(train_dataset)}")
|
| 759 |
+
logger.info(f" Num Epochs = {args.num_train_epochs}")
|
| 760 |
+
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
|
| 761 |
+
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
| 762 |
+
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
|
| 763 |
+
logger.info(f" Total optimization steps = {args.max_train_steps}")
|
| 764 |
+
global_step = 0
|
| 765 |
+
first_epoch = 0
|
| 766 |
+
|
| 767 |
+
# Potentially load in the weights and states from a previous save
|
| 768 |
+
if args.resume_from_checkpoint:
|
| 769 |
+
if args.resume_from_checkpoint != "latest":
|
| 770 |
+
path = os.path.basename(args.resume_from_checkpoint)
|
| 771 |
+
else:
|
| 772 |
+
# Get the most recent checkpoint
|
| 773 |
+
dirs = os.listdir(args.output_dir)
|
| 774 |
+
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
| 775 |
+
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
| 776 |
+
path = dirs[-1] if len(dirs) > 0 else None
|
| 777 |
+
|
| 778 |
+
if path is None:
|
| 779 |
+
accelerator.print(
|
| 780 |
+
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
|
| 781 |
+
)
|
| 782 |
+
args.resume_from_checkpoint = None
|
| 783 |
+
else:
|
| 784 |
+
accelerator.print(f"Resuming from checkpoint {path}")
|
| 785 |
+
accelerator.load_state(os.path.join(args.output_dir, path))
|
| 786 |
+
global_step = int(path.split("-")[1])
|
| 787 |
+
|
| 788 |
+
resume_global_step = global_step * args.gradient_accumulation_steps
|
| 789 |
+
first_epoch = global_step // num_update_steps_per_epoch
|
| 790 |
+
resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
|
| 791 |
+
|
| 792 |
+
# Only show the progress bar once on each machine.
|
| 793 |
+
progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process,
|
| 794 |
+
bar_format="{l_bar}{bar:10}{r_bar}{bar:-10b}")
|
| 795 |
+
progress_bar.set_description("Steps")
|
| 796 |
+
|
| 797 |
+
print(simple_mapper)
|
| 798 |
+
|
| 799 |
+
for epoch in range(first_epoch, args.num_train_epochs):
|
| 800 |
+
unet.train()
|
| 801 |
+
train_loss = 0.0
|
| 802 |
+
train_attn_loss = 0.0
|
| 803 |
+
train_diff_loss = 0.0
|
| 804 |
+
for step, batch in enumerate(train_dataloader):
|
| 805 |
+
# Skip steps until we reach the resumed step
|
| 806 |
+
if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
|
| 807 |
+
if step % args.gradient_accumulation_steps == 0:
|
| 808 |
+
progress_bar.update(1)
|
| 809 |
+
continue
|
| 810 |
+
|
| 811 |
+
with accelerator.accumulate(unet, simple_mapper):
|
| 812 |
+
# Convert images to latent space
|
| 813 |
+
latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
|
| 814 |
+
latents = latents * vae.config.scaling_factor
|
| 815 |
+
|
| 816 |
+
# Sample noise that we'll add to the latents
|
| 817 |
+
noise = torch.randn_like(latents)
|
| 818 |
+
if args.noise_offset:
|
| 819 |
+
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
|
| 820 |
+
noise += args.noise_offset * torch.randn(
|
| 821 |
+
(latents.shape[0], latents.shape[1], 1, 1), device=latents.device
|
| 822 |
+
)
|
| 823 |
+
|
| 824 |
+
bsz = latents.shape[0]
|
| 825 |
+
# Sample a random timestep for each image
|
| 826 |
+
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
|
| 827 |
+
timesteps = timesteps.long()
|
| 828 |
+
|
| 829 |
+
# Add noise to the latents according to the noise magnitude at each timestep
|
| 830 |
+
# (this is the forward diffusion process)
|
| 831 |
+
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
|
| 832 |
+
|
| 833 |
+
# Get the text embedding for conditioning
|
| 834 |
+
mapper_outputs = simple_mapper(batch['codes'])
|
| 835 |
+
# print(mapper_outputs.size(), batch["input_ids"].size())
|
| 836 |
+
modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(batch["input_ids"],
|
| 837 |
+
None,
|
| 838 |
+
mapper_outputs,
|
| 839 |
+
placeholder_token_ids)
|
| 840 |
+
# print(modified_hs.size())
|
| 841 |
+
encoder_hidden_states = text_encoder(batch["input_ids"], hidden_states=modified_hs)[0]
|
| 842 |
+
|
| 843 |
+
# Get the target for loss depending on the prediction type
|
| 844 |
+
if args.prediction_type is not None:
|
| 845 |
+
# set prediction_type of scheduler if defined
|
| 846 |
+
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
|
| 847 |
+
|
| 848 |
+
if noise_scheduler.config.prediction_type == "epsilon":
|
| 849 |
+
target = noise
|
| 850 |
+
elif noise_scheduler.config.prediction_type == "v_prediction":
|
| 851 |
+
target = noise_scheduler.get_velocity(latents, noise, timesteps)
|
| 852 |
+
else:
|
| 853 |
+
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
|
| 854 |
+
|
| 855 |
+
# Predict the noise residual and compute loss
|
| 856 |
+
model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
|
| 857 |
+
|
| 858 |
+
if args.snr_gamma is None:
|
| 859 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
|
| 860 |
+
attn_loss, max_attn = dreamcreature_loss(batch,
|
| 861 |
+
unet,
|
| 862 |
+
dino,
|
| 863 |
+
seg,
|
| 864 |
+
placeholder_token_ids,
|
| 865 |
+
accelerator)
|
| 866 |
+
if args.masked_training:
|
| 867 |
+
masks = batch['masks'].unsqueeze(1).to(accelerator.device)
|
| 868 |
+
loss_image_mask = F.interpolate(masks.float(),
|
| 869 |
+
size=target.shape[-2:],
|
| 870 |
+
mode='bilinear') * torch.ones_like(target)
|
| 871 |
+
loss = loss * loss_image_mask
|
| 872 |
+
loss = loss.sum() / loss_image_mask.sum()
|
| 873 |
+
else:
|
| 874 |
+
loss = loss.mean()
|
| 875 |
+
else:
|
| 876 |
+
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
|
| 877 |
+
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
|
| 878 |
+
# This is discussed in Section 4.2 of the same paper.
|
| 879 |
+
snr = compute_snr(noise_scheduler, timesteps)
|
| 880 |
+
if noise_scheduler.config.prediction_type == "v_prediction":
|
| 881 |
+
# Velocity objective requires that we add one to SNR values before we divide by them.
|
| 882 |
+
snr = snr + 1
|
| 883 |
+
mse_loss_weights = (
|
| 884 |
+
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
|
| 885 |
+
)
|
| 886 |
+
|
| 887 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
|
| 888 |
+
attn_loss, max_attn = dreamcreature_loss(batch,
|
| 889 |
+
unet,
|
| 890 |
+
dino,
|
| 891 |
+
seg,
|
| 892 |
+
placeholder_token_ids,
|
| 893 |
+
accelerator)
|
| 894 |
+
if args.masked_training:
|
| 895 |
+
masks = batch['masks'].unsqueeze(1).to(accelerator.device)
|
| 896 |
+
loss_image_mask = F.interpolate(masks.float(),
|
| 897 |
+
size=target.shape[-2:],
|
| 898 |
+
mode='bilinear') * torch.ones_like(target)
|
| 899 |
+
loss = loss * loss_image_mask
|
| 900 |
+
loss = loss.sum(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
|
| 901 |
+
loss = loss.sum() / loss_image_mask.sum()
|
| 902 |
+
else:
|
| 903 |
+
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
|
| 904 |
+
loss = loss.mean()
|
| 905 |
+
|
| 906 |
+
diff_loss = loss.clone().detach()
|
| 907 |
+
avg_diff_loss = accelerator.gather(diff_loss.repeat(args.train_batch_size)).mean()
|
| 908 |
+
train_diff_loss += avg_diff_loss.item() / args.gradient_accumulation_steps
|
| 909 |
+
|
| 910 |
+
avg_attn_loss = accelerator.gather(attn_loss.repeat(args.train_batch_size)).mean()
|
| 911 |
+
train_attn_loss += avg_attn_loss.item() / args.gradient_accumulation_steps
|
| 912 |
+
|
| 913 |
+
loss += args.attn_loss * attn_loss
|
| 914 |
+
|
| 915 |
+
# Gather the losses across all processes for logging (if we use distributed training).
|
| 916 |
+
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
|
| 917 |
+
train_loss += avg_loss.item() / args.gradient_accumulation_steps
|
| 918 |
+
|
| 919 |
+
# Backpropagate
|
| 920 |
+
accelerator.backward(loss)
|
| 921 |
+
if accelerator.sync_gradients:
|
| 922 |
+
params_to_clip = list(lora_layers.parameters()) + list(simple_mapper.parameters())
|
| 923 |
+
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
|
| 924 |
+
|
| 925 |
+
optimizer.step()
|
| 926 |
+
lr_scheduler.step()
|
| 927 |
+
optimizer.zero_grad()
|
| 928 |
+
|
| 929 |
+
# Checks if the accelerator has performed an optimization step behind the scenes
|
| 930 |
+
if accelerator.sync_gradients:
|
| 931 |
+
progress_bar.update(1)
|
| 932 |
+
global_step += 1
|
| 933 |
+
accelerator.log({"train_loss": train_loss,
|
| 934 |
+
"diff_loss": train_diff_loss,
|
| 935 |
+
"attn_loss": train_attn_loss,
|
| 936 |
+
"mapper_norm": mapper_outputs.detach().norm().item(),
|
| 937 |
+
"max_attn": max_attn.item()
|
| 938 |
+
}, step=global_step)
|
| 939 |
+
train_loss = 0.0
|
| 940 |
+
train_attn_loss = 0.0
|
| 941 |
+
train_diff_loss = 0.0
|
| 942 |
+
|
| 943 |
+
if global_step % args.checkpointing_steps == 0:
|
| 944 |
+
if accelerator.is_main_process:
|
| 945 |
+
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
|
| 946 |
+
if args.checkpoints_total_limit is not None:
|
| 947 |
+
checkpoints = os.listdir(args.output_dir)
|
| 948 |
+
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
|
| 949 |
+
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
|
| 950 |
+
|
| 951 |
+
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
|
| 952 |
+
if len(checkpoints) >= args.checkpoints_total_limit:
|
| 953 |
+
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
|
| 954 |
+
removing_checkpoints = checkpoints[0:num_to_remove]
|
| 955 |
+
|
| 956 |
+
logger.info(
|
| 957 |
+
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
|
| 958 |
+
)
|
| 959 |
+
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
|
| 960 |
+
|
| 961 |
+
for removing_checkpoint in removing_checkpoints:
|
| 962 |
+
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
|
| 963 |
+
shutil.rmtree(removing_checkpoint)
|
| 964 |
+
|
| 965 |
+
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
| 966 |
+
accelerator.save_state(save_path)
|
| 967 |
+
logger.info(f"Saved state to {save_path}")
|
| 968 |
+
|
| 969 |
+
logs = {"step_loss": diff_loss.detach().item(),
|
| 970 |
+
"attn_loss": attn_loss.detach().item(),
|
| 971 |
+
"lr": lr_scheduler.get_last_lr()[0]}
|
| 972 |
+
progress_bar.set_postfix(**logs)
|
| 973 |
+
|
| 974 |
+
if global_step >= args.max_train_steps:
|
| 975 |
+
break
|
| 976 |
+
|
| 977 |
+
if accelerator.is_main_process:
|
| 978 |
+
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
|
| 979 |
+
logger.info(
|
| 980 |
+
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
|
| 981 |
+
f" {args.validation_prompt}."
|
| 982 |
+
)
|
| 983 |
+
pipeline = DreamCreatureSDPipeline.from_pretrained(
|
| 984 |
+
args.pretrained_model_name_or_path,
|
| 985 |
+
unet=accelerator.unwrap_model(unet),
|
| 986 |
+
text_encoder=accelerator.unwrap_model(text_encoder),
|
| 987 |
+
tokenizer=tokenizer,
|
| 988 |
+
revision=args.revision,
|
| 989 |
+
torch_dtype=weight_dtype,
|
| 990 |
+
)
|
| 991 |
+
pipeline.placeholder_token_ids = placeholder_token_ids
|
| 992 |
+
pipeline.simple_mapper = accelerator.unwrap_model(simple_mapper)
|
| 993 |
+
pipeline.replace_token = False
|
| 994 |
+
|
| 995 |
+
pipeline = pipeline.to(accelerator.device)
|
| 996 |
+
pipeline.set_progress_bar_config(disable=True)
|
| 997 |
+
|
| 998 |
+
# run inference
|
| 999 |
+
generator = torch.Generator(device=accelerator.device)
|
| 1000 |
+
if args.seed is not None:
|
| 1001 |
+
generator = generator.manual_seed(args.seed)
|
| 1002 |
+
images = []
|
| 1003 |
+
for _ in range(args.num_validation_images):
|
| 1004 |
+
images.append(
|
| 1005 |
+
pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]
|
| 1006 |
+
)
|
| 1007 |
+
|
| 1008 |
+
for tracker in accelerator.trackers:
|
| 1009 |
+
if tracker.name == "tensorboard":
|
| 1010 |
+
np_images = np.stack([np.asarray(img) for img in images])
|
| 1011 |
+
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
|
| 1012 |
+
if tracker.name == "wandb":
|
| 1013 |
+
tracker.log(
|
| 1014 |
+
{
|
| 1015 |
+
"validation": [
|
| 1016 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 1017 |
+
for i, image in enumerate(images)
|
| 1018 |
+
]
|
| 1019 |
+
}
|
| 1020 |
+
)
|
| 1021 |
+
|
| 1022 |
+
del pipeline
|
| 1023 |
+
torch.cuda.empty_cache()
|
| 1024 |
+
|
| 1025 |
+
# Save the lora layers
|
| 1026 |
+
accelerator.wait_for_everyone()
|
| 1027 |
+
if accelerator.is_main_process:
|
| 1028 |
+
# unet = unet.to(torch.float32)
|
| 1029 |
+
# unet.save_attn_procs(args.output_dir, safe_serialization=not args.custom_diffusion)
|
| 1030 |
+
|
| 1031 |
+
torch.save(lora_layers.to(torch.float32).state_dict(), args.output_dir + '/lora_layers.pth')
|
| 1032 |
+
torch.save(simple_mapper.to(torch.float32).state_dict(), args.output_dir + '/hash_mapper.pth')
|
| 1033 |
+
|
| 1034 |
+
if args.push_to_hub:
|
| 1035 |
+
save_model_card(
|
| 1036 |
+
repo_id,
|
| 1037 |
+
images=images,
|
| 1038 |
+
base_model=args.pretrained_model_name_or_path,
|
| 1039 |
+
dataset_name=args.dataset_name,
|
| 1040 |
+
repo_folder=args.output_dir,
|
| 1041 |
+
)
|
| 1042 |
+
upload_folder(
|
| 1043 |
+
repo_id=repo_id,
|
| 1044 |
+
folder_path=args.output_dir,
|
| 1045 |
+
commit_message="End of training",
|
| 1046 |
+
ignore_patterns=["step_*", "epoch_*"],
|
| 1047 |
+
)
|
| 1048 |
+
|
| 1049 |
+
del unet
|
| 1050 |
+
|
| 1051 |
+
# Final inference
|
| 1052 |
+
# Load previous pipeline
|
| 1053 |
+
tokenizer = MultiTokenCLIPTokenizer.from_pretrained(
|
| 1054 |
+
args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
|
| 1055 |
+
)
|
| 1056 |
+
text_encoder = CustomCLIPTextModel.from_pretrained(
|
| 1057 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
|
| 1058 |
+
)
|
| 1059 |
+
unet_path = args.unet_path if args.unet_path is not None else args.pretrained_model_name_or_path
|
| 1060 |
+
unet: UNet2DConditionModel = UNet2DConditionModel.from_pretrained(
|
| 1061 |
+
unet_path, subfolder="unet", revision=args.revision
|
| 1062 |
+
)
|
| 1063 |
+
pipeline = DreamCreatureSDPipeline.from_pretrained(
|
| 1064 |
+
args.pretrained_model_name_or_path,
|
| 1065 |
+
unet=unet,
|
| 1066 |
+
text_encoder=text_encoder,
|
| 1067 |
+
tokenizer=tokenizer,
|
| 1068 |
+
revision=args.revision,
|
| 1069 |
+
torch_dtype=weight_dtype,
|
| 1070 |
+
)
|
| 1071 |
+
placeholder_token = "<part>"
|
| 1072 |
+
initializer_token = None
|
| 1073 |
+
placeholder_token_ids = add_tokens(tokenizer,
|
| 1074 |
+
text_encoder,
|
| 1075 |
+
placeholder_token,
|
| 1076 |
+
args.num_parts,
|
| 1077 |
+
initializer_token)
|
| 1078 |
+
pipeline.placeholder_token_ids = placeholder_token_ids
|
| 1079 |
+
pipeline.simple_mapper = TokenMapper(args.num_parts,
|
| 1080 |
+
args.num_k_per_part,
|
| 1081 |
+
OUT_DIMS,
|
| 1082 |
+
args.projection_nlayers)
|
| 1083 |
+
pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + '/hash_mapper.pth', map_location='cpu'))
|
| 1084 |
+
pipeline.simple_mapper.to(accelerator.device)
|
| 1085 |
+
|
| 1086 |
+
pipeline = pipeline.to(accelerator.device)
|
| 1087 |
+
|
| 1088 |
+
# load attention processors
|
| 1089 |
+
# pipeline.unet.load_attn_procs(args.output_dir, use_safetensors=not args.custom_diffusion)
|
| 1090 |
+
setup_attn_processor(pipeline.unet, rank=args.rank)
|
| 1091 |
+
load_attn_processor(pipeline.unet, args.output_dir + '/lora_layers.pth')
|
| 1092 |
+
|
| 1093 |
+
# run inference
|
| 1094 |
+
pipeline.replace_token = False
|
| 1095 |
+
generator = torch.Generator(device=accelerator.device)
|
| 1096 |
+
if args.seed is not None:
|
| 1097 |
+
generator = generator.manual_seed(args.seed)
|
| 1098 |
+
images = []
|
| 1099 |
+
for _ in range(args.num_validation_images):
|
| 1100 |
+
images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0])
|
| 1101 |
+
|
| 1102 |
+
if accelerator.is_main_process:
|
| 1103 |
+
for tracker in accelerator.trackers:
|
| 1104 |
+
if len(images) != 0:
|
| 1105 |
+
if tracker.name == "tensorboard":
|
| 1106 |
+
np_images = np.stack([np.asarray(img) for img in images])
|
| 1107 |
+
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
|
| 1108 |
+
if tracker.name == "wandb":
|
| 1109 |
+
tracker.log(
|
| 1110 |
+
{
|
| 1111 |
+
"test": [
|
| 1112 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 1113 |
+
for i, image in enumerate(images)
|
| 1114 |
+
]
|
| 1115 |
+
}
|
| 1116 |
+
)
|
| 1117 |
+
|
| 1118 |
+
accelerator.end_training()
|
| 1119 |
+
|
| 1120 |
+
|
| 1121 |
+
if __name__ == "__main__":
|
| 1122 |
+
main()
|
train_dreamcreature_sdxl.py
ADDED
|
@@ -0,0 +1,1539 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
# coding=utf-8
|
| 3 |
+
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
+
# you may not use this file except in compliance with the License.
|
| 7 |
+
# You may obtain a copy of the License at
|
| 8 |
+
#
|
| 9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
+
#
|
| 11 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
+
# See the License for the specific language governing permissions and
|
| 15 |
+
# limitations under the License.
|
| 16 |
+
"""Fine-tuning script for Stable Diffusion XL for text2image with support for LoRA."""
|
| 17 |
+
|
| 18 |
+
import argparse
|
| 19 |
+
import copy
|
| 20 |
+
import itertools
|
| 21 |
+
import logging
|
| 22 |
+
import math
|
| 23 |
+
import os
|
| 24 |
+
import random
|
| 25 |
+
import shutil
|
| 26 |
+
from pathlib import Path
|
| 27 |
+
from typing import Dict
|
| 28 |
+
|
| 29 |
+
import datasets
|
| 30 |
+
import diffusers
|
| 31 |
+
import numpy as np
|
| 32 |
+
import torch
|
| 33 |
+
import torch.nn.functional as F
|
| 34 |
+
import torch.utils.checkpoint
|
| 35 |
+
import transformers
|
| 36 |
+
from accelerate import Accelerator
|
| 37 |
+
from accelerate.logging import get_logger
|
| 38 |
+
from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
|
| 39 |
+
from diffusers import (
|
| 40 |
+
AutoencoderKL,
|
| 41 |
+
DDPMScheduler,
|
| 42 |
+
StableDiffusionXLPipeline,
|
| 43 |
+
UNet2DConditionModel,
|
| 44 |
+
)
|
| 45 |
+
from diffusers.loaders import LoraLoaderMixin
|
| 46 |
+
from diffusers.models.lora import LoRALinearLayer
|
| 47 |
+
from diffusers.optimization import get_scheduler
|
| 48 |
+
from diffusers.training_utils import compute_snr
|
| 49 |
+
from diffusers.utils import check_min_version, is_wandb_available
|
| 50 |
+
from diffusers.utils.import_utils import is_xformers_available
|
| 51 |
+
from huggingface_hub import create_repo, upload_folder
|
| 52 |
+
from packaging import version
|
| 53 |
+
from torchvision import transforms
|
| 54 |
+
from torchvision.transforms.functional import crop
|
| 55 |
+
from tqdm.auto import tqdm
|
| 56 |
+
from transformers import PretrainedConfig
|
| 57 |
+
|
| 58 |
+
from dreamcreature.attn_processor import AttnProcessorCustom
|
| 59 |
+
from dreamcreature.dataset import DreamCreatureDataset
|
| 60 |
+
from dreamcreature.dino import DINO
|
| 61 |
+
from dreamcreature.kmeans_segmentation import KMeansSegmentation
|
| 62 |
+
from dreamcreature.loss import dreamcreature_loss
|
| 63 |
+
from dreamcreature.mapper import TokenMapper
|
| 64 |
+
from dreamcreature.pipeline_xl import DreamCreatureSDXLPipeline
|
| 65 |
+
from dreamcreature.text_encoder import CustomCLIPTextModel, CustomCLIPTextModelWithProjection
|
| 66 |
+
from dreamcreature.tokenizer import MultiTokenCLIPTokenizer
|
| 67 |
+
from utils import add_tokens, tokenize_prompt, get_attn_processors
|
| 68 |
+
|
| 69 |
+
IMAGENET_TEMPLATES = [
|
| 70 |
+
"a photo of a {}",
|
| 71 |
+
"a rendering of a {}",
|
| 72 |
+
"a cropped photo of the {}",
|
| 73 |
+
"the photo of a {}",
|
| 74 |
+
"a photo of a clean {}",
|
| 75 |
+
"a photo of a dirty {}",
|
| 76 |
+
"a dark photo of the {}",
|
| 77 |
+
"a photo of my {}",
|
| 78 |
+
"a photo of the cool {}",
|
| 79 |
+
"a close-up photo of a {}",
|
| 80 |
+
"a bright photo of the {}",
|
| 81 |
+
"a cropped photo of a {}",
|
| 82 |
+
"a photo of the {}",
|
| 83 |
+
"a good photo of the {}",
|
| 84 |
+
"a photo of one {}",
|
| 85 |
+
"a close-up photo of the {}",
|
| 86 |
+
"a rendition of the {}",
|
| 87 |
+
"a photo of the clean {}",
|
| 88 |
+
"a rendition of a {}",
|
| 89 |
+
"a photo of a nice {}",
|
| 90 |
+
"a good photo of a {}",
|
| 91 |
+
"a photo of the nice {}",
|
| 92 |
+
"a photo of the small {}",
|
| 93 |
+
"a photo of the weird {}",
|
| 94 |
+
"a photo of the large {}",
|
| 95 |
+
"a photo of a cool {}",
|
| 96 |
+
"a photo of a small {}",
|
| 97 |
+
]
|
| 98 |
+
|
| 99 |
+
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
| 100 |
+
check_min_version("0.25.0.dev0")
|
| 101 |
+
|
| 102 |
+
logger = get_logger(__name__)
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
# TODO: This function should be removed once training scripts are rewritten in PEFT
|
| 106 |
+
def text_encoder_lora_state_dict(text_encoder):
|
| 107 |
+
state_dict = {}
|
| 108 |
+
|
| 109 |
+
def text_encoder_attn_modules(text_encoder):
|
| 110 |
+
from transformers import CLIPTextModel, CLIPTextModelWithProjection
|
| 111 |
+
|
| 112 |
+
attn_modules = []
|
| 113 |
+
|
| 114 |
+
if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
|
| 115 |
+
for i, layer in enumerate(text_encoder.text_model.encoder.layers):
|
| 116 |
+
name = f"text_model.encoder.layers.{i}.self_attn"
|
| 117 |
+
mod = layer.self_attn
|
| 118 |
+
attn_modules.append((name, mod))
|
| 119 |
+
|
| 120 |
+
return attn_modules
|
| 121 |
+
|
| 122 |
+
for name, module in text_encoder_attn_modules(text_encoder):
|
| 123 |
+
for k, v in module.q_proj.lora_linear_layer.state_dict().items():
|
| 124 |
+
state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
|
| 125 |
+
|
| 126 |
+
for k, v in module.k_proj.lora_linear_layer.state_dict().items():
|
| 127 |
+
state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
|
| 128 |
+
|
| 129 |
+
for k, v in module.v_proj.lora_linear_layer.state_dict().items():
|
| 130 |
+
state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
|
| 131 |
+
|
| 132 |
+
for k, v in module.out_proj.lora_linear_layer.state_dict().items():
|
| 133 |
+
state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
|
| 134 |
+
|
| 135 |
+
return state_dict
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
def save_model_card(
|
| 139 |
+
repo_id: str,
|
| 140 |
+
images=None,
|
| 141 |
+
base_model=str,
|
| 142 |
+
dataset_name=str,
|
| 143 |
+
train_text_encoder=False,
|
| 144 |
+
repo_folder=None,
|
| 145 |
+
vae_path=None,
|
| 146 |
+
):
|
| 147 |
+
img_str = ""
|
| 148 |
+
for i, image in enumerate(images):
|
| 149 |
+
image.save(os.path.join(repo_folder, f"image_{i}.png"))
|
| 150 |
+
img_str += f"\n"
|
| 151 |
+
|
| 152 |
+
yaml = f"""
|
| 153 |
+
---
|
| 154 |
+
license: creativeml-openrail-m
|
| 155 |
+
base_model: {base_model}
|
| 156 |
+
dataset: {dataset_name}
|
| 157 |
+
tags:
|
| 158 |
+
- stable-diffusion-xl
|
| 159 |
+
- stable-diffusion-xl-diffusers
|
| 160 |
+
- text-to-image
|
| 161 |
+
- diffusers
|
| 162 |
+
- lora
|
| 163 |
+
inference: true
|
| 164 |
+
---
|
| 165 |
+
"""
|
| 166 |
+
model_card = f"""
|
| 167 |
+
# LoRA text2image fine-tuning - {repo_id}
|
| 168 |
+
|
| 169 |
+
These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n
|
| 170 |
+
{img_str}
|
| 171 |
+
|
| 172 |
+
LoRA for the text encoder was enabled: {train_text_encoder}.
|
| 173 |
+
|
| 174 |
+
Special VAE used for training: {vae_path}.
|
| 175 |
+
"""
|
| 176 |
+
with open(os.path.join(repo_folder, "README.md"), "w") as f:
|
| 177 |
+
f.write(yaml + model_card)
|
| 178 |
+
|
| 179 |
+
|
| 180 |
+
def import_model_class_from_model_name_or_path(
|
| 181 |
+
pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
|
| 182 |
+
):
|
| 183 |
+
text_encoder_config = PretrainedConfig.from_pretrained(
|
| 184 |
+
pretrained_model_name_or_path, subfolder=subfolder, revision=revision
|
| 185 |
+
)
|
| 186 |
+
model_class = text_encoder_config.architectures[0]
|
| 187 |
+
|
| 188 |
+
if model_class == "CLIPTextModel":
|
| 189 |
+
from transformers import CLIPTextModel
|
| 190 |
+
|
| 191 |
+
return CLIPTextModel
|
| 192 |
+
elif model_class == "CLIPTextModelWithProjection":
|
| 193 |
+
from transformers import CLIPTextModelWithProjection
|
| 194 |
+
|
| 195 |
+
return CLIPTextModelWithProjection
|
| 196 |
+
else:
|
| 197 |
+
raise ValueError(f"{model_class} is not supported.")
|
| 198 |
+
|
| 199 |
+
|
| 200 |
+
def parse_args(input_args=None):
|
| 201 |
+
parser = argparse.ArgumentParser(description="Simple example of a training script.")
|
| 202 |
+
parser.add_argument(
|
| 203 |
+
"--pretrained_model_name_or_path",
|
| 204 |
+
type=str,
|
| 205 |
+
default=None,
|
| 206 |
+
required=True,
|
| 207 |
+
help="Path to pretrained model or model identifier from huggingface.co/models.",
|
| 208 |
+
)
|
| 209 |
+
parser.add_argument(
|
| 210 |
+
"--pretrained_vae_model_name_or_path",
|
| 211 |
+
type=str,
|
| 212 |
+
default=None,
|
| 213 |
+
help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
|
| 214 |
+
)
|
| 215 |
+
parser.add_argument(
|
| 216 |
+
"--revision",
|
| 217 |
+
type=str,
|
| 218 |
+
default=None,
|
| 219 |
+
required=False,
|
| 220 |
+
help="Revision of pretrained model identifier from huggingface.co/models.",
|
| 221 |
+
)
|
| 222 |
+
parser.add_argument(
|
| 223 |
+
"--variant",
|
| 224 |
+
type=str,
|
| 225 |
+
default=None,
|
| 226 |
+
help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
|
| 227 |
+
)
|
| 228 |
+
parser.add_argument(
|
| 229 |
+
"--dataset_name",
|
| 230 |
+
type=str,
|
| 231 |
+
default=None,
|
| 232 |
+
help=(
|
| 233 |
+
"The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
|
| 234 |
+
" dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
|
| 235 |
+
" or to a folder containing files that 🤗 Datasets can understand."
|
| 236 |
+
),
|
| 237 |
+
)
|
| 238 |
+
parser.add_argument(
|
| 239 |
+
"--dataset_config_name",
|
| 240 |
+
type=str,
|
| 241 |
+
default=None,
|
| 242 |
+
help="The config of the Dataset, leave as None if there's only one config.",
|
| 243 |
+
)
|
| 244 |
+
parser.add_argument(
|
| 245 |
+
"--train_data_dir",
|
| 246 |
+
type=str,
|
| 247 |
+
default=None,
|
| 248 |
+
help=(
|
| 249 |
+
"A folder containing the training data. Folder contents must follow the structure described in"
|
| 250 |
+
" https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
|
| 251 |
+
" must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
|
| 252 |
+
),
|
| 253 |
+
)
|
| 254 |
+
parser.add_argument(
|
| 255 |
+
"--image_column", type=str, default="image", help="The column of the dataset containing an image."
|
| 256 |
+
)
|
| 257 |
+
parser.add_argument(
|
| 258 |
+
"--caption_column",
|
| 259 |
+
type=str,
|
| 260 |
+
default="text",
|
| 261 |
+
help="The column of the dataset containing a caption or a list of captions.",
|
| 262 |
+
)
|
| 263 |
+
parser.add_argument(
|
| 264 |
+
"--validation_prompt",
|
| 265 |
+
type=str,
|
| 266 |
+
default=None,
|
| 267 |
+
help="A prompt that is used during validation to verify that the model is learning.",
|
| 268 |
+
)
|
| 269 |
+
parser.add_argument(
|
| 270 |
+
"--num_validation_images",
|
| 271 |
+
type=int,
|
| 272 |
+
default=4,
|
| 273 |
+
help="Number of images that should be generated during validation with `validation_prompt`.",
|
| 274 |
+
)
|
| 275 |
+
parser.add_argument(
|
| 276 |
+
"--validation_epochs",
|
| 277 |
+
type=int,
|
| 278 |
+
default=1,
|
| 279 |
+
help=(
|
| 280 |
+
"Run fine-tuning validation every X epochs. The validation process consists of running the prompt"
|
| 281 |
+
" `args.validation_prompt` multiple times: `args.num_validation_images`."
|
| 282 |
+
),
|
| 283 |
+
)
|
| 284 |
+
parser.add_argument(
|
| 285 |
+
"--max_train_samples",
|
| 286 |
+
type=int,
|
| 287 |
+
default=None,
|
| 288 |
+
help=(
|
| 289 |
+
"For debugging purposes or quicker training, truncate the number of training examples to this "
|
| 290 |
+
"value if set."
|
| 291 |
+
),
|
| 292 |
+
)
|
| 293 |
+
parser.add_argument(
|
| 294 |
+
"--output_dir",
|
| 295 |
+
type=str,
|
| 296 |
+
default="sd-model-finetuned-lora",
|
| 297 |
+
help="The output directory where the model predictions and checkpoints will be written.",
|
| 298 |
+
)
|
| 299 |
+
parser.add_argument(
|
| 300 |
+
"--cache_dir",
|
| 301 |
+
type=str,
|
| 302 |
+
default=None,
|
| 303 |
+
help="The directory where the downloaded models and datasets will be stored.",
|
| 304 |
+
)
|
| 305 |
+
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
|
| 306 |
+
parser.add_argument(
|
| 307 |
+
"--resolution",
|
| 308 |
+
type=int,
|
| 309 |
+
default=1024,
|
| 310 |
+
help=(
|
| 311 |
+
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
|
| 312 |
+
" resolution"
|
| 313 |
+
),
|
| 314 |
+
)
|
| 315 |
+
parser.add_argument(
|
| 316 |
+
"--center_crop",
|
| 317 |
+
default=False,
|
| 318 |
+
action="store_true",
|
| 319 |
+
help=(
|
| 320 |
+
"Whether to center crop the input images to the resolution. If not set, the images will be randomly"
|
| 321 |
+
" cropped. The images will be resized to the resolution first before cropping."
|
| 322 |
+
),
|
| 323 |
+
)
|
| 324 |
+
parser.add_argument(
|
| 325 |
+
"--random_flip",
|
| 326 |
+
action="store_true",
|
| 327 |
+
help="whether to randomly flip images horizontally",
|
| 328 |
+
)
|
| 329 |
+
parser.add_argument(
|
| 330 |
+
"--train_text_encoder",
|
| 331 |
+
action="store_true",
|
| 332 |
+
help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
|
| 333 |
+
)
|
| 334 |
+
parser.add_argument(
|
| 335 |
+
"--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
|
| 336 |
+
)
|
| 337 |
+
parser.add_argument("--num_train_epochs", type=int, default=100)
|
| 338 |
+
parser.add_argument(
|
| 339 |
+
"--max_train_steps",
|
| 340 |
+
type=int,
|
| 341 |
+
default=None,
|
| 342 |
+
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
|
| 343 |
+
)
|
| 344 |
+
parser.add_argument(
|
| 345 |
+
"--checkpointing_steps",
|
| 346 |
+
type=int,
|
| 347 |
+
default=500,
|
| 348 |
+
help=(
|
| 349 |
+
"Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
|
| 350 |
+
" checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
|
| 351 |
+
" training using `--resume_from_checkpoint`."
|
| 352 |
+
),
|
| 353 |
+
)
|
| 354 |
+
parser.add_argument(
|
| 355 |
+
"--checkpoints_total_limit",
|
| 356 |
+
type=int,
|
| 357 |
+
default=None,
|
| 358 |
+
help=("Max number of checkpoints to store."),
|
| 359 |
+
)
|
| 360 |
+
parser.add_argument(
|
| 361 |
+
"--resume_from_checkpoint",
|
| 362 |
+
type=str,
|
| 363 |
+
default=None,
|
| 364 |
+
help=(
|
| 365 |
+
"Whether training should be resumed from a previous checkpoint. Use a path saved by"
|
| 366 |
+
' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
|
| 367 |
+
),
|
| 368 |
+
)
|
| 369 |
+
parser.add_argument(
|
| 370 |
+
"--gradient_accumulation_steps",
|
| 371 |
+
type=int,
|
| 372 |
+
default=1,
|
| 373 |
+
help="Number of updates steps to accumulate before performing a backward/update pass.",
|
| 374 |
+
)
|
| 375 |
+
parser.add_argument(
|
| 376 |
+
"--gradient_checkpointing",
|
| 377 |
+
action="store_true",
|
| 378 |
+
help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
|
| 379 |
+
)
|
| 380 |
+
parser.add_argument(
|
| 381 |
+
"--learning_rate",
|
| 382 |
+
type=float,
|
| 383 |
+
default=1e-4,
|
| 384 |
+
help="Initial learning rate (after the potential warmup period) to use.",
|
| 385 |
+
)
|
| 386 |
+
parser.add_argument(
|
| 387 |
+
"--scale_lr",
|
| 388 |
+
action="store_true",
|
| 389 |
+
default=False,
|
| 390 |
+
help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
|
| 391 |
+
)
|
| 392 |
+
parser.add_argument(
|
| 393 |
+
"--lr_scheduler",
|
| 394 |
+
type=str,
|
| 395 |
+
default="constant",
|
| 396 |
+
help=(
|
| 397 |
+
'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
|
| 398 |
+
' "constant", "constant_with_warmup"]'
|
| 399 |
+
),
|
| 400 |
+
)
|
| 401 |
+
parser.add_argument(
|
| 402 |
+
"--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
|
| 403 |
+
)
|
| 404 |
+
parser.add_argument(
|
| 405 |
+
"--snr_gamma",
|
| 406 |
+
type=float,
|
| 407 |
+
default=None,
|
| 408 |
+
help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
|
| 409 |
+
"More details here: https://arxiv.org/abs/2303.09556.",
|
| 410 |
+
)
|
| 411 |
+
parser.add_argument(
|
| 412 |
+
"--allow_tf32",
|
| 413 |
+
action="store_true",
|
| 414 |
+
help=(
|
| 415 |
+
"Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
|
| 416 |
+
" https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
|
| 417 |
+
),
|
| 418 |
+
)
|
| 419 |
+
parser.add_argument(
|
| 420 |
+
"--dataloader_num_workers",
|
| 421 |
+
type=int,
|
| 422 |
+
default=0,
|
| 423 |
+
help=(
|
| 424 |
+
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
| 425 |
+
),
|
| 426 |
+
)
|
| 427 |
+
parser.add_argument(
|
| 428 |
+
"--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
|
| 429 |
+
)
|
| 430 |
+
parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
|
| 431 |
+
parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
|
| 432 |
+
parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
|
| 433 |
+
parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
|
| 434 |
+
parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
|
| 435 |
+
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
|
| 436 |
+
parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
|
| 437 |
+
parser.add_argument(
|
| 438 |
+
"--prediction_type",
|
| 439 |
+
type=str,
|
| 440 |
+
default=None,
|
| 441 |
+
help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediciton_type` is chosen.",
|
| 442 |
+
)
|
| 443 |
+
parser.add_argument(
|
| 444 |
+
"--hub_model_id",
|
| 445 |
+
type=str,
|
| 446 |
+
default=None,
|
| 447 |
+
help="The name of the repository to keep in sync with the local `output_dir`.",
|
| 448 |
+
)
|
| 449 |
+
parser.add_argument(
|
| 450 |
+
"--logging_dir",
|
| 451 |
+
type=str,
|
| 452 |
+
default="logs",
|
| 453 |
+
help=(
|
| 454 |
+
"[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
|
| 455 |
+
" *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
|
| 456 |
+
),
|
| 457 |
+
)
|
| 458 |
+
parser.add_argument(
|
| 459 |
+
"--report_to",
|
| 460 |
+
type=str,
|
| 461 |
+
default="tensorboard",
|
| 462 |
+
help=(
|
| 463 |
+
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
|
| 464 |
+
' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
|
| 465 |
+
),
|
| 466 |
+
)
|
| 467 |
+
parser.add_argument(
|
| 468 |
+
"--mixed_precision",
|
| 469 |
+
type=str,
|
| 470 |
+
default=None,
|
| 471 |
+
choices=["no", "fp16", "bf16"],
|
| 472 |
+
help=(
|
| 473 |
+
"Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
|
| 474 |
+
" 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the"
|
| 475 |
+
" flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
|
| 476 |
+
),
|
| 477 |
+
)
|
| 478 |
+
parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
|
| 479 |
+
parser.add_argument(
|
| 480 |
+
"--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
|
| 481 |
+
)
|
| 482 |
+
parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
|
| 483 |
+
parser.add_argument(
|
| 484 |
+
"--rank",
|
| 485 |
+
type=int,
|
| 486 |
+
default=4,
|
| 487 |
+
help=("The dimension of the LoRA update matrices."),
|
| 488 |
+
)
|
| 489 |
+
|
| 490 |
+
parser.add_argument('--filename', default='train.txt')
|
| 491 |
+
parser.add_argument('--code_filename', default='train_caps_better_m8_k256.txt')
|
| 492 |
+
parser.add_argument('--repeat', default=1, type=int)
|
| 493 |
+
|
| 494 |
+
parser.add_argument('--scheduler_steps', default=1000, type=int, help='scheduler step, if turbo, set to 4')
|
| 495 |
+
parser.add_argument('--num_parts', type=int, default=4, help="Number of parts")
|
| 496 |
+
parser.add_argument('--num_k_per_part', type=int, default=256, help='Number of k')
|
| 497 |
+
|
| 498 |
+
parser.add_argument('--mapper_lr_scale', default=1, type=float)
|
| 499 |
+
parser.add_argument('--mapper_lr', default=0.0001, type=float)
|
| 500 |
+
parser.add_argument('--attn_loss', default=0, type=float)
|
| 501 |
+
parser.add_argument('--projection_nlayers', default=3, type=int)
|
| 502 |
+
|
| 503 |
+
parser.add_argument('--masked_training', action='store_true')
|
| 504 |
+
parser.add_argument('--drop_tokens', action='store_true')
|
| 505 |
+
parser.add_argument('--drop_rate', type=float, default=0.5)
|
| 506 |
+
parser.add_argument('--drop_counts', default='half')
|
| 507 |
+
|
| 508 |
+
parser.add_argument('--class_name', default='')
|
| 509 |
+
parser.add_argument('--no_pe', action='store_true')
|
| 510 |
+
parser.add_argument('--vector_shuffle', action='store_true')
|
| 511 |
+
|
| 512 |
+
parser.add_argument('--use_gt_label', action='store_true')
|
| 513 |
+
parser.add_argument('--bg_code', default=7, type=int) # for gt_label
|
| 514 |
+
parser.add_argument('--fg_idx', default=0, type=int)
|
| 515 |
+
parser.add_argument('--use_templates', action='store_true')
|
| 516 |
+
|
| 517 |
+
parser.add_argument('--filter_class', default=None, type=int, help='debugging purpose')
|
| 518 |
+
|
| 519 |
+
if input_args is not None:
|
| 520 |
+
args = parser.parse_args(input_args)
|
| 521 |
+
else:
|
| 522 |
+
args = parser.parse_args()
|
| 523 |
+
|
| 524 |
+
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
|
| 525 |
+
if env_local_rank != -1 and env_local_rank != args.local_rank:
|
| 526 |
+
args.local_rank = env_local_rank
|
| 527 |
+
|
| 528 |
+
# Sanity checks
|
| 529 |
+
if args.dataset_name is None and args.train_data_dir is None:
|
| 530 |
+
raise ValueError("Need either a dataset name or a training folder.")
|
| 531 |
+
|
| 532 |
+
return args
|
| 533 |
+
|
| 534 |
+
|
| 535 |
+
def unet_attn_processors_state_dict(unet) -> Dict[str, torch.tensor]:
|
| 536 |
+
"""
|
| 537 |
+
Returns:
|
| 538 |
+
a state dict containing just the attention processor parameters.
|
| 539 |
+
"""
|
| 540 |
+
attn_processors = get_attn_processors(unet)
|
| 541 |
+
|
| 542 |
+
attn_processors_state_dict = {}
|
| 543 |
+
|
| 544 |
+
for attn_processor_key, attn_processor in attn_processors.items():
|
| 545 |
+
for parameter_key, parameter in attn_processor.state_dict().items():
|
| 546 |
+
attn_processors_state_dict[f"{attn_processor_key}.{parameter_key}"] = parameter
|
| 547 |
+
|
| 548 |
+
return attn_processors_state_dict
|
| 549 |
+
|
| 550 |
+
|
| 551 |
+
def encode_prompt(text_encoders, text_input_ids_list, placeholder_token_ids, mapper_outputs):
|
| 552 |
+
prompt_embeds_list = []
|
| 553 |
+
|
| 554 |
+
for i, text_encoder in enumerate(text_encoders):
|
| 555 |
+
text_input_ids = text_input_ids_list[i]
|
| 556 |
+
|
| 557 |
+
modified_hs = text_encoder.text_model.forward_embeddings_with_mapper(text_input_ids,
|
| 558 |
+
None,
|
| 559 |
+
mapper_outputs[i],
|
| 560 |
+
placeholder_token_ids)
|
| 561 |
+
|
| 562 |
+
prompt_embeds = text_encoder(text_input_ids,
|
| 563 |
+
hidden_states=modified_hs,
|
| 564 |
+
output_hidden_states=True)
|
| 565 |
+
# prompt_embeds = text_encoder(
|
| 566 |
+
# text_input_ids.to(text_encoder.device),
|
| 567 |
+
# output_hidden_states=True,
|
| 568 |
+
# )
|
| 569 |
+
|
| 570 |
+
# We are only ALWAYS interested in the pooled output of the final text encoder
|
| 571 |
+
pooled_prompt_embeds = prompt_embeds[0]
|
| 572 |
+
prompt_embeds = prompt_embeds.hidden_states[-2]
|
| 573 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
| 574 |
+
prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
|
| 575 |
+
prompt_embeds_list.append(prompt_embeds)
|
| 576 |
+
|
| 577 |
+
prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
|
| 578 |
+
pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
|
| 579 |
+
return prompt_embeds, pooled_prompt_embeds
|
| 580 |
+
|
| 581 |
+
|
| 582 |
+
def collate_fn(args, tokenizer_one, tokenizer_two, placeholder_token):
|
| 583 |
+
# Preprocessing the datasets.
|
| 584 |
+
train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR)
|
| 585 |
+
train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
|
| 586 |
+
train_flip = transforms.RandomHorizontalFlip(p=1.0)
|
| 587 |
+
train_transforms = transforms.Compose(
|
| 588 |
+
[
|
| 589 |
+
transforms.ToTensor(),
|
| 590 |
+
transforms.Normalize([0.5], [0.5]),
|
| 591 |
+
]
|
| 592 |
+
)
|
| 593 |
+
|
| 594 |
+
def f(examples):
|
| 595 |
+
# image aug
|
| 596 |
+
original_sizes = []
|
| 597 |
+
all_images = []
|
| 598 |
+
crop_top_lefts = []
|
| 599 |
+
captions = []
|
| 600 |
+
raw_images = []
|
| 601 |
+
appeared_tokens = []
|
| 602 |
+
codes = []
|
| 603 |
+
for i in range(len(examples)):
|
| 604 |
+
##### original sdxl process #####
|
| 605 |
+
image = examples[i]['pixel_values'].convert('RGB')
|
| 606 |
+
original_sizes.append((image.height, image.width))
|
| 607 |
+
image = train_resize(image)
|
| 608 |
+
if args.center_crop:
|
| 609 |
+
y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
|
| 610 |
+
x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
|
| 611 |
+
image = train_crop(image)
|
| 612 |
+
else:
|
| 613 |
+
y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
|
| 614 |
+
image = crop(image, y1, x1, h, w)
|
| 615 |
+
if args.random_flip and random.random() < 0.5:
|
| 616 |
+
# flip
|
| 617 |
+
x1 = image.width - x1
|
| 618 |
+
image = train_flip(image)
|
| 619 |
+
crop_top_left = (y1, x1)
|
| 620 |
+
crop_top_lefts.append(crop_top_left)
|
| 621 |
+
raw_images.append(image)
|
| 622 |
+
image = train_transforms(image)
|
| 623 |
+
all_images.append(image)
|
| 624 |
+
|
| 625 |
+
##### dreamcreature caption #####
|
| 626 |
+
if args.use_templates and random.random() <= 0.5: # 50% using templates
|
| 627 |
+
if args.class_name != '':
|
| 628 |
+
caption = random.choice(IMAGENET_TEMPLATES).format(f'{placeholder_token} {args.class_name}')
|
| 629 |
+
else:
|
| 630 |
+
caption = random.choice(IMAGENET_TEMPLATES).format(placeholder_token)
|
| 631 |
+
else:
|
| 632 |
+
if args.class_name != '':
|
| 633 |
+
caption = f'{placeholder_token} {args.class_name}'
|
| 634 |
+
else:
|
| 635 |
+
caption = placeholder_token
|
| 636 |
+
|
| 637 |
+
tokens = tokenizer_one.token_map[placeholder_token][:args.num_parts]
|
| 638 |
+
tokens = [tokens[a] for a in examples[i]['appeared']]
|
| 639 |
+
|
| 640 |
+
if args.vector_shuffle or args.drop_tokens:
|
| 641 |
+
tokens = copy.copy(tokens)
|
| 642 |
+
random.shuffle(tokens)
|
| 643 |
+
|
| 644 |
+
if args.drop_tokens and random.random() < args.drop_rate and len(tokens) >= 2:
|
| 645 |
+
# randomly drop half of the tokens
|
| 646 |
+
if args.drop_counts == 'half':
|
| 647 |
+
tokens = tokens[:len(tokens) // 2]
|
| 648 |
+
else:
|
| 649 |
+
tokens = tokens[:int(args.drop_counts)]
|
| 650 |
+
|
| 651 |
+
caption = caption.replace(placeholder_token, ' '.join(tokens))
|
| 652 |
+
captions.append(caption)
|
| 653 |
+
|
| 654 |
+
appeared = [int(t.split('_')[1]) for t in tokens] # <part>_i
|
| 655 |
+
# examples[i]['appeared'] = appeared
|
| 656 |
+
|
| 657 |
+
appeared_tokens.append(appeared)
|
| 658 |
+
|
| 659 |
+
code = examples[i]['codes']
|
| 660 |
+
codes.append(code)
|
| 661 |
+
|
| 662 |
+
tokens_one = tokenize_prompt(tokenizer_one, captions)
|
| 663 |
+
tokens_two = tokenize_prompt(tokenizer_two, captions)
|
| 664 |
+
|
| 665 |
+
##### start stacking #####
|
| 666 |
+
pixel_values = torch.stack([image for image in all_images])
|
| 667 |
+
pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
|
| 668 |
+
original_sizes = [s for s in original_sizes]
|
| 669 |
+
crop_top_lefts = [c for c in crop_top_lefts]
|
| 670 |
+
input_ids_one = torch.stack([t for t in tokens_one])
|
| 671 |
+
input_ids_two = torch.stack([t for t in tokens_two])
|
| 672 |
+
|
| 673 |
+
codes = torch.stack(codes, dim=0)
|
| 674 |
+
|
| 675 |
+
collate_output = {
|
| 676 |
+
"original_sizes": original_sizes,
|
| 677 |
+
"crop_top_lefts": crop_top_lefts,
|
| 678 |
+
"pixel_values": pixel_values,
|
| 679 |
+
"input_ids_one": input_ids_one,
|
| 680 |
+
"input_ids_two": input_ids_two,
|
| 681 |
+
"raw_images": raw_images,
|
| 682 |
+
"appeared_tokens": appeared_tokens,
|
| 683 |
+
"codes": codes
|
| 684 |
+
}
|
| 685 |
+
|
| 686 |
+
return collate_output
|
| 687 |
+
|
| 688 |
+
return f
|
| 689 |
+
|
| 690 |
+
|
| 691 |
+
def setup_attn_processors(unet, args):
|
| 692 |
+
attn_size = args.resolution // 32
|
| 693 |
+
attn_procs = {}
|
| 694 |
+
for name in unet.attn_processors.keys():
|
| 695 |
+
attn_procs[name] = AttnProcessorCustom(attn_size)
|
| 696 |
+
unet.set_attn_processor(attn_procs)
|
| 697 |
+
|
| 698 |
+
|
| 699 |
+
def init_for_pipeline(args):
|
| 700 |
+
tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
|
| 701 |
+
args.pretrained_model_name_or_path,
|
| 702 |
+
subfolder="tokenizer",
|
| 703 |
+
revision=args.revision,
|
| 704 |
+
use_fast=False,
|
| 705 |
+
)
|
| 706 |
+
tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
|
| 707 |
+
args.pretrained_model_name_or_path,
|
| 708 |
+
subfolder="tokenizer_2",
|
| 709 |
+
revision=args.revision,
|
| 710 |
+
use_fast=False,
|
| 711 |
+
)
|
| 712 |
+
text_encoder_cls_one = CustomCLIPTextModel
|
| 713 |
+
text_encoder_cls_two = CustomCLIPTextModelWithProjection
|
| 714 |
+
text_encoder_one = text_encoder_cls_one.from_pretrained(
|
| 715 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
|
| 716 |
+
)
|
| 717 |
+
text_encoder_two = text_encoder_cls_two.from_pretrained(
|
| 718 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
|
| 719 |
+
)
|
| 720 |
+
|
| 721 |
+
OUT_DIMS = 768 + 1280 # 2048
|
| 722 |
+
simple_mapper = TokenMapper(args.num_parts,
|
| 723 |
+
args.num_k_per_part,
|
| 724 |
+
OUT_DIMS,
|
| 725 |
+
args.projection_nlayers)
|
| 726 |
+
return text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper
|
| 727 |
+
|
| 728 |
+
|
| 729 |
+
def main(args):
|
| 730 |
+
logging_dir = Path(args.output_dir, args.logging_dir)
|
| 731 |
+
|
| 732 |
+
accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
|
| 733 |
+
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
|
| 734 |
+
accelerator = Accelerator(
|
| 735 |
+
gradient_accumulation_steps=args.gradient_accumulation_steps,
|
| 736 |
+
mixed_precision=args.mixed_precision,
|
| 737 |
+
log_with=args.report_to,
|
| 738 |
+
project_config=accelerator_project_config,
|
| 739 |
+
kwargs_handlers=[kwargs],
|
| 740 |
+
)
|
| 741 |
+
|
| 742 |
+
if args.report_to == "wandb":
|
| 743 |
+
if not is_wandb_available():
|
| 744 |
+
raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
|
| 745 |
+
import wandb
|
| 746 |
+
|
| 747 |
+
# Make one log on every process with the configuration for debugging.
|
| 748 |
+
logging.basicConfig(
|
| 749 |
+
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
| 750 |
+
datefmt="%m/%d/%Y %H:%M:%S",
|
| 751 |
+
level=logging.INFO,
|
| 752 |
+
)
|
| 753 |
+
logger.info(accelerator.state, main_process_only=False)
|
| 754 |
+
if accelerator.is_local_main_process:
|
| 755 |
+
datasets.utils.logging.set_verbosity_warning()
|
| 756 |
+
transformers.utils.logging.set_verbosity_warning()
|
| 757 |
+
diffusers.utils.logging.set_verbosity_info()
|
| 758 |
+
else:
|
| 759 |
+
datasets.utils.logging.set_verbosity_error()
|
| 760 |
+
transformers.utils.logging.set_verbosity_error()
|
| 761 |
+
diffusers.utils.logging.set_verbosity_error()
|
| 762 |
+
|
| 763 |
+
# If passed along, set the training seed now.
|
| 764 |
+
if args.seed is not None:
|
| 765 |
+
set_seed(args.seed)
|
| 766 |
+
|
| 767 |
+
# Handle the repository creation
|
| 768 |
+
if accelerator.is_main_process:
|
| 769 |
+
if args.output_dir is not None:
|
| 770 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
| 771 |
+
|
| 772 |
+
if args.push_to_hub:
|
| 773 |
+
repo_id = create_repo(
|
| 774 |
+
repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
|
| 775 |
+
).repo_id
|
| 776 |
+
|
| 777 |
+
# Load the tokenizers (replace AutoTokenizer with the custom MultiTokenCLIPTokenizer)
|
| 778 |
+
tokenizer_one = MultiTokenCLIPTokenizer.from_pretrained(
|
| 779 |
+
args.pretrained_model_name_or_path,
|
| 780 |
+
subfolder="tokenizer",
|
| 781 |
+
revision=args.revision,
|
| 782 |
+
use_fast=False,
|
| 783 |
+
)
|
| 784 |
+
tokenizer_two = MultiTokenCLIPTokenizer.from_pretrained(
|
| 785 |
+
args.pretrained_model_name_or_path,
|
| 786 |
+
subfolder="tokenizer_2",
|
| 787 |
+
revision=args.revision,
|
| 788 |
+
use_fast=False,
|
| 789 |
+
)
|
| 790 |
+
# import correct text encoder classes
|
| 791 |
+
# text_encoder_cls_one = import_model_class_from_model_name_or_path(
|
| 792 |
+
# args.pretrained_model_name_or_path, args.revision
|
| 793 |
+
# )
|
| 794 |
+
# text_encoder_cls_two = import_model_class_from_model_name_or_path(
|
| 795 |
+
# args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
|
| 796 |
+
# )
|
| 797 |
+
text_encoder_cls_one = CustomCLIPTextModel
|
| 798 |
+
text_encoder_cls_two = CustomCLIPTextModelWithProjection
|
| 799 |
+
|
| 800 |
+
# Load scheduler and models
|
| 801 |
+
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path,
|
| 802 |
+
subfolder="scheduler",
|
| 803 |
+
num_train_steps=args.scheduler_steps)
|
| 804 |
+
text_encoder_one = text_encoder_cls_one.from_pretrained(
|
| 805 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
|
| 806 |
+
)
|
| 807 |
+
text_encoder_two = text_encoder_cls_two.from_pretrained(
|
| 808 |
+
args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
|
| 809 |
+
)
|
| 810 |
+
vae_path = (
|
| 811 |
+
args.pretrained_model_name_or_path
|
| 812 |
+
if args.pretrained_vae_model_name_or_path is None
|
| 813 |
+
else args.pretrained_vae_model_name_or_path
|
| 814 |
+
)
|
| 815 |
+
vae = AutoencoderKL.from_pretrained(
|
| 816 |
+
vae_path,
|
| 817 |
+
subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
|
| 818 |
+
revision=args.revision,
|
| 819 |
+
variant=args.variant,
|
| 820 |
+
)
|
| 821 |
+
unet = UNet2DConditionModel.from_pretrained(
|
| 822 |
+
args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
|
| 823 |
+
)
|
| 824 |
+
|
| 825 |
+
##### dreamcreature init #####
|
| 826 |
+
OUT_DIMS = 768 + 1280 # 2048
|
| 827 |
+
|
| 828 |
+
dino = DINO()
|
| 829 |
+
seg = KMeansSegmentation(args.train_data_dir + '/pretrained_kmeans.pth',
|
| 830 |
+
args.fg_idx,
|
| 831 |
+
args.bg_code,
|
| 832 |
+
args.num_parts,
|
| 833 |
+
args.num_k_per_part)
|
| 834 |
+
|
| 835 |
+
simple_mapper = TokenMapper(args.num_parts,
|
| 836 |
+
args.num_k_per_part,
|
| 837 |
+
OUT_DIMS,
|
| 838 |
+
args.projection_nlayers)
|
| 839 |
+
|
| 840 |
+
# We only train the additional adapter LoRA layers
|
| 841 |
+
vae.requires_grad_(False)
|
| 842 |
+
text_encoder_one.requires_grad_(False)
|
| 843 |
+
text_encoder_two.requires_grad_(False)
|
| 844 |
+
unet.requires_grad_(False)
|
| 845 |
+
dino.requires_grad_(False)
|
| 846 |
+
|
| 847 |
+
##### dreamcreature, add sub-concepts token ids ####
|
| 848 |
+
placeholder_token = "<part>"
|
| 849 |
+
initializer_token = None
|
| 850 |
+
placeholder_token_ids_one = add_tokens(tokenizer_one,
|
| 851 |
+
text_encoder_one,
|
| 852 |
+
placeholder_token,
|
| 853 |
+
args.num_parts,
|
| 854 |
+
initializer_token)
|
| 855 |
+
placeholder_token_ids_two = add_tokens(tokenizer_two,
|
| 856 |
+
text_encoder_two,
|
| 857 |
+
placeholder_token,
|
| 858 |
+
args.num_parts,
|
| 859 |
+
initializer_token)
|
| 860 |
+
|
| 861 |
+
# For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
|
| 862 |
+
# as these weights are only used for inference, keeping weights in full precision is not required.
|
| 863 |
+
weight_dtype = torch.float32
|
| 864 |
+
if accelerator.mixed_precision == "fp16":
|
| 865 |
+
weight_dtype = torch.float16
|
| 866 |
+
elif accelerator.mixed_precision == "bf16":
|
| 867 |
+
weight_dtype = torch.bfloat16
|
| 868 |
+
|
| 869 |
+
# Move unet, vae and text_encoder to device and cast to weight_dtype
|
| 870 |
+
# The VAE is in float32 to avoid NaN losses.
|
| 871 |
+
unet.to(accelerator.device, dtype=weight_dtype)
|
| 872 |
+
if args.pretrained_vae_model_name_or_path is None:
|
| 873 |
+
vae.to(accelerator.device, dtype=torch.float32)
|
| 874 |
+
else:
|
| 875 |
+
vae.to(accelerator.device, dtype=weight_dtype)
|
| 876 |
+
text_encoder_one.to(accelerator.device, dtype=weight_dtype)
|
| 877 |
+
text_encoder_two.to(accelerator.device, dtype=weight_dtype)
|
| 878 |
+
simple_mapper.to(accelerator.device)
|
| 879 |
+
|
| 880 |
+
if args.enable_xformers_memory_efficient_attention:
|
| 881 |
+
if is_xformers_available():
|
| 882 |
+
import xformers
|
| 883 |
+
|
| 884 |
+
xformers_version = version.parse(xformers.__version__)
|
| 885 |
+
if xformers_version == version.parse("0.0.16"):
|
| 886 |
+
logger.warn(
|
| 887 |
+
"xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
|
| 888 |
+
)
|
| 889 |
+
unet.enable_xformers_memory_efficient_attention()
|
| 890 |
+
else:
|
| 891 |
+
raise ValueError("xformers is not available. Make sure it is installed correctly")
|
| 892 |
+
|
| 893 |
+
# now we will add new LoRA weights to the attention layers
|
| 894 |
+
# Set correct lora layers
|
| 895 |
+
unet_lora_parameters = []
|
| 896 |
+
for attn_processor_name, attn_processor in unet.attn_processors.items():
|
| 897 |
+
# Parse the attention module.
|
| 898 |
+
attn_module = unet
|
| 899 |
+
for n in attn_processor_name.split(".")[:-1]:
|
| 900 |
+
attn_module = getattr(attn_module, n)
|
| 901 |
+
|
| 902 |
+
# Set the `lora_layer` attribute of the attention-related matrices.
|
| 903 |
+
attn_module.to_q.set_lora_layer(
|
| 904 |
+
LoRALinearLayer(
|
| 905 |
+
in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
|
| 906 |
+
)
|
| 907 |
+
)
|
| 908 |
+
attn_module.to_k.set_lora_layer(
|
| 909 |
+
LoRALinearLayer(
|
| 910 |
+
in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
|
| 911 |
+
)
|
| 912 |
+
)
|
| 913 |
+
attn_module.to_v.set_lora_layer(
|
| 914 |
+
LoRALinearLayer(
|
| 915 |
+
in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
|
| 916 |
+
)
|
| 917 |
+
)
|
| 918 |
+
attn_module.to_out[0].set_lora_layer(
|
| 919 |
+
LoRALinearLayer(
|
| 920 |
+
in_features=attn_module.to_out[0].in_features,
|
| 921 |
+
out_features=attn_module.to_out[0].out_features,
|
| 922 |
+
rank=args.rank,
|
| 923 |
+
)
|
| 924 |
+
)
|
| 925 |
+
|
| 926 |
+
# Accumulate the LoRA params to optimize.
|
| 927 |
+
unet_lora_parameters.extend(attn_module.to_q.lora_layer.parameters())
|
| 928 |
+
unet_lora_parameters.extend(attn_module.to_k.lora_layer.parameters())
|
| 929 |
+
unet_lora_parameters.extend(attn_module.to_v.lora_layer.parameters())
|
| 930 |
+
unet_lora_parameters.extend(attn_module.to_out[0].lora_layer.parameters())
|
| 931 |
+
|
| 932 |
+
setup_attn_processors(unet, args)
|
| 933 |
+
|
| 934 |
+
# The text encoder comes from 🤗 transformers, so we cannot directly modify it.
|
| 935 |
+
# So, instead, we monkey-patch the forward calls of its attention-blocks.
|
| 936 |
+
if args.train_text_encoder:
|
| 937 |
+
# ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
|
| 938 |
+
text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
|
| 939 |
+
text_encoder_one, dtype=torch.float32, rank=args.rank
|
| 940 |
+
)
|
| 941 |
+
text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
|
| 942 |
+
text_encoder_two, dtype=torch.float32, rank=args.rank
|
| 943 |
+
)
|
| 944 |
+
|
| 945 |
+
# create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
|
| 946 |
+
def save_model_hook(models, weights, output_dir):
|
| 947 |
+
if accelerator.is_main_process:
|
| 948 |
+
# there are only two options here. Either are just the unet attn processor layers
|
| 949 |
+
# or there are the unet and text encoder atten layers
|
| 950 |
+
unet_lora_layers_to_save = None
|
| 951 |
+
text_encoder_one_lora_layers_to_save = None
|
| 952 |
+
text_encoder_two_lora_layers_to_save = None
|
| 953 |
+
mapper_to_save = None
|
| 954 |
+
|
| 955 |
+
for model in models:
|
| 956 |
+
if isinstance(model, type(accelerator.unwrap_model(unet))):
|
| 957 |
+
unet_lora_layers_to_save = unet_attn_processors_state_dict(model)
|
| 958 |
+
elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
|
| 959 |
+
text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model)
|
| 960 |
+
elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
|
| 961 |
+
text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model)
|
| 962 |
+
elif isinstance(model, TokenMapper):
|
| 963 |
+
mapper_to_save = model.state_dict()
|
| 964 |
+
else:
|
| 965 |
+
raise ValueError(f"unexpected save model: {model.__class__}")
|
| 966 |
+
|
| 967 |
+
# make sure to pop weight so that corresponding model is not saved again
|
| 968 |
+
weights.pop()
|
| 969 |
+
|
| 970 |
+
StableDiffusionXLPipeline.save_lora_weights(
|
| 971 |
+
output_dir,
|
| 972 |
+
unet_lora_layers=unet_lora_layers_to_save,
|
| 973 |
+
text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
|
| 974 |
+
text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
|
| 975 |
+
)
|
| 976 |
+
torch.save(mapper_to_save, output_dir + '/hash_mapper.pth')
|
| 977 |
+
|
| 978 |
+
def load_model_hook(models, input_dir):
|
| 979 |
+
unet_ = None
|
| 980 |
+
text_encoder_one_ = None
|
| 981 |
+
text_encoder_two_ = None
|
| 982 |
+
mapper_ = None
|
| 983 |
+
|
| 984 |
+
while len(models) > 0:
|
| 985 |
+
model = models.pop()
|
| 986 |
+
|
| 987 |
+
if isinstance(model, type(accelerator.unwrap_model(unet))):
|
| 988 |
+
unet_ = model
|
| 989 |
+
elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))):
|
| 990 |
+
text_encoder_one_ = model
|
| 991 |
+
elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))):
|
| 992 |
+
text_encoder_two_ = model
|
| 993 |
+
elif isinstance(model, TokenMapper):
|
| 994 |
+
mapper_ = model
|
| 995 |
+
else:
|
| 996 |
+
raise ValueError(f"unexpected save model: {model.__class__}")
|
| 997 |
+
|
| 998 |
+
lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
|
| 999 |
+
LoraLoaderMixin.load_lora_into_unet(lora_state_dict, network_alphas=network_alphas, unet=unet_)
|
| 1000 |
+
|
| 1001 |
+
text_encoder_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder." in k}
|
| 1002 |
+
LoraLoaderMixin.load_lora_into_text_encoder(
|
| 1003 |
+
text_encoder_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_one_
|
| 1004 |
+
)
|
| 1005 |
+
|
| 1006 |
+
text_encoder_2_state_dict = {k: v for k, v in lora_state_dict.items() if "text_encoder_2." in k}
|
| 1007 |
+
LoraLoaderMixin.load_lora_into_text_encoder(
|
| 1008 |
+
text_encoder_2_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_two_
|
| 1009 |
+
)
|
| 1010 |
+
mapper_.load_state_dict(torch.load(input_dir + '/hash_mapper.pth'))
|
| 1011 |
+
|
| 1012 |
+
accelerator.register_save_state_pre_hook(save_model_hook)
|
| 1013 |
+
accelerator.register_load_state_pre_hook(load_model_hook)
|
| 1014 |
+
|
| 1015 |
+
# Enable TF32 for faster training on Ampere GPUs,
|
| 1016 |
+
# cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
|
| 1017 |
+
if args.allow_tf32:
|
| 1018 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 1019 |
+
|
| 1020 |
+
if args.scale_lr:
|
| 1021 |
+
args.learning_rate = (
|
| 1022 |
+
args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
|
| 1023 |
+
)
|
| 1024 |
+
|
| 1025 |
+
# Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
|
| 1026 |
+
if args.use_8bit_adam:
|
| 1027 |
+
try:
|
| 1028 |
+
import bitsandbytes as bnb
|
| 1029 |
+
except ImportError:
|
| 1030 |
+
raise ImportError(
|
| 1031 |
+
"To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
|
| 1032 |
+
)
|
| 1033 |
+
|
| 1034 |
+
optimizer_class = bnb.optim.AdamW8bit
|
| 1035 |
+
else:
|
| 1036 |
+
optimizer_class = torch.optim.AdamW
|
| 1037 |
+
|
| 1038 |
+
extra_params = list(simple_mapper.parameters())
|
| 1039 |
+
mapper_lr = args.learning_rate * args.mapper_lr_scale if args.learning_rate != 0 else args.mapper_lr
|
| 1040 |
+
|
| 1041 |
+
# Optimizer creation
|
| 1042 |
+
params_to_optimize = (
|
| 1043 |
+
itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
|
| 1044 |
+
if args.train_text_encoder
|
| 1045 |
+
else unet_lora_parameters
|
| 1046 |
+
)
|
| 1047 |
+
optimizer = optimizer_class(
|
| 1048 |
+
[{'params': params_to_optimize},
|
| 1049 |
+
{'params': extra_params, 'lr': mapper_lr}],
|
| 1050 |
+
lr=args.learning_rate,
|
| 1051 |
+
betas=(args.adam_beta1, args.adam_beta2),
|
| 1052 |
+
weight_decay=args.adam_weight_decay,
|
| 1053 |
+
eps=args.adam_epsilon,
|
| 1054 |
+
)
|
| 1055 |
+
|
| 1056 |
+
# create
|
| 1057 |
+
train_dataset = DreamCreatureDataset(args.train_data_dir,
|
| 1058 |
+
args.filename,
|
| 1059 |
+
code_filename=args.code_filename,
|
| 1060 |
+
num_parts=args.num_parts,
|
| 1061 |
+
num_k_per_part=args.num_k_per_part,
|
| 1062 |
+
repeat=args.repeat,
|
| 1063 |
+
use_gt_label=args.use_gt_label,
|
| 1064 |
+
bg_code=args.bg_code)
|
| 1065 |
+
|
| 1066 |
+
with accelerator.main_process_first():
|
| 1067 |
+
if args.filter_class is not None:
|
| 1068 |
+
train_dataset.filter_by_class(args.filter_class)
|
| 1069 |
+
print('selected', len(train_dataset))
|
| 1070 |
+
if args.max_train_samples is not None:
|
| 1071 |
+
train_dataset.set_max_samples(args.max_train_samples, args.seed)
|
| 1072 |
+
|
| 1073 |
+
# DataLoaders creation:
|
| 1074 |
+
train_dataloader = torch.utils.data.DataLoader(
|
| 1075 |
+
train_dataset,
|
| 1076 |
+
shuffle=True,
|
| 1077 |
+
collate_fn=collate_fn(args, tokenizer_one, tokenizer_two, placeholder_token),
|
| 1078 |
+
batch_size=args.train_batch_size,
|
| 1079 |
+
num_workers=args.dataloader_num_workers,
|
| 1080 |
+
)
|
| 1081 |
+
|
| 1082 |
+
# Scheduler and math around the number of training steps.
|
| 1083 |
+
overrode_max_train_steps = False
|
| 1084 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 1085 |
+
if args.max_train_steps is None:
|
| 1086 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 1087 |
+
overrode_max_train_steps = True
|
| 1088 |
+
|
| 1089 |
+
lr_scheduler = get_scheduler(
|
| 1090 |
+
args.lr_scheduler,
|
| 1091 |
+
optimizer=optimizer,
|
| 1092 |
+
num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
|
| 1093 |
+
num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
|
| 1094 |
+
)
|
| 1095 |
+
|
| 1096 |
+
# Prepare everything with our `accelerator`.
|
| 1097 |
+
if args.train_text_encoder:
|
| 1098 |
+
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
| 1099 |
+
unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler
|
| 1100 |
+
)
|
| 1101 |
+
else:
|
| 1102 |
+
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
| 1103 |
+
unet, optimizer, train_dataloader, lr_scheduler
|
| 1104 |
+
)
|
| 1105 |
+
simple_mapper = accelerator.prepare(simple_mapper)
|
| 1106 |
+
|
| 1107 |
+
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
|
| 1108 |
+
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
|
| 1109 |
+
if overrode_max_train_steps:
|
| 1110 |
+
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
|
| 1111 |
+
# Afterwards we recalculate our number of training epochs
|
| 1112 |
+
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
|
| 1113 |
+
|
| 1114 |
+
# We need to initialize the trackers we use, and also store our configuration.
|
| 1115 |
+
# The trackers initializes automatically on the main process.
|
| 1116 |
+
if accelerator.is_main_process:
|
| 1117 |
+
accelerator.init_trackers("text2image-fine-tune", config=vars(args))
|
| 1118 |
+
|
| 1119 |
+
# Train!
|
| 1120 |
+
total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
|
| 1121 |
+
|
| 1122 |
+
logger.info("***** Running training *****")
|
| 1123 |
+
logger.info(f" Num examples = {len(train_dataset)}")
|
| 1124 |
+
logger.info(f" Num Epochs = {args.num_train_epochs}")
|
| 1125 |
+
logger.info(f" Instantaneous batch size per device = {args.train_batch_size}")
|
| 1126 |
+
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
| 1127 |
+
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
|
| 1128 |
+
logger.info(f" Total optimization steps = {args.max_train_steps}")
|
| 1129 |
+
global_step = 0
|
| 1130 |
+
first_epoch = 0
|
| 1131 |
+
|
| 1132 |
+
# Potentially load in the weights and states from a previous save
|
| 1133 |
+
if args.resume_from_checkpoint:
|
| 1134 |
+
if args.resume_from_checkpoint != "latest":
|
| 1135 |
+
path = os.path.basename(args.resume_from_checkpoint)
|
| 1136 |
+
else:
|
| 1137 |
+
# Get the most recent checkpoint
|
| 1138 |
+
dirs = os.listdir(args.output_dir)
|
| 1139 |
+
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
| 1140 |
+
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
| 1141 |
+
path = dirs[-1] if len(dirs) > 0 else None
|
| 1142 |
+
|
| 1143 |
+
if path is None:
|
| 1144 |
+
accelerator.print(
|
| 1145 |
+
f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
|
| 1146 |
+
)
|
| 1147 |
+
args.resume_from_checkpoint = None
|
| 1148 |
+
initial_global_step = 0
|
| 1149 |
+
else:
|
| 1150 |
+
accelerator.print(f"Resuming from checkpoint {path}")
|
| 1151 |
+
accelerator.load_state(os.path.join(args.output_dir, path))
|
| 1152 |
+
global_step = int(path.split("-")[1])
|
| 1153 |
+
|
| 1154 |
+
initial_global_step = global_step
|
| 1155 |
+
first_epoch = global_step // num_update_steps_per_epoch
|
| 1156 |
+
|
| 1157 |
+
else:
|
| 1158 |
+
initial_global_step = 0
|
| 1159 |
+
|
| 1160 |
+
progress_bar = tqdm(
|
| 1161 |
+
range(0, args.max_train_steps),
|
| 1162 |
+
initial=initial_global_step,
|
| 1163 |
+
desc="Steps",
|
| 1164 |
+
# Only show the progress bar once on each machine.
|
| 1165 |
+
disable=not accelerator.is_local_main_process,
|
| 1166 |
+
)
|
| 1167 |
+
|
| 1168 |
+
for epoch in range(first_epoch, args.num_train_epochs):
|
| 1169 |
+
unet.train()
|
| 1170 |
+
if args.train_text_encoder:
|
| 1171 |
+
text_encoder_one.train()
|
| 1172 |
+
text_encoder_two.train()
|
| 1173 |
+
train_loss = 0.0
|
| 1174 |
+
train_diff_loss = 0.0
|
| 1175 |
+
train_attn_loss = 0.0
|
| 1176 |
+
for step, batch in enumerate(train_dataloader):
|
| 1177 |
+
with accelerator.accumulate(unet, simple_mapper):
|
| 1178 |
+
# Convert images to latent space
|
| 1179 |
+
if args.pretrained_vae_model_name_or_path is not None:
|
| 1180 |
+
pixel_values = batch["pixel_values"].to(dtype=weight_dtype)
|
| 1181 |
+
else:
|
| 1182 |
+
pixel_values = batch["pixel_values"]
|
| 1183 |
+
|
| 1184 |
+
model_input = vae.encode(pixel_values).latent_dist.sample()
|
| 1185 |
+
model_input = model_input * vae.config.scaling_factor
|
| 1186 |
+
if args.pretrained_vae_model_name_or_path is None:
|
| 1187 |
+
model_input = model_input.to(weight_dtype)
|
| 1188 |
+
|
| 1189 |
+
# Sample noise that we'll add to the latents
|
| 1190 |
+
noise = torch.randn_like(model_input)
|
| 1191 |
+
if args.noise_offset:
|
| 1192 |
+
# https://www.crosslabs.org//blog/diffusion-with-offset-noise
|
| 1193 |
+
noise += args.noise_offset * torch.randn(
|
| 1194 |
+
(model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device
|
| 1195 |
+
)
|
| 1196 |
+
|
| 1197 |
+
bsz = model_input.shape[0]
|
| 1198 |
+
# Sample a random timestep for each image
|
| 1199 |
+
timesteps = torch.randint(
|
| 1200 |
+
0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
|
| 1201 |
+
)
|
| 1202 |
+
timesteps = timesteps.long()
|
| 1203 |
+
|
| 1204 |
+
# Add noise to the model input according to the noise magnitude at each timestep
|
| 1205 |
+
# (this is the forward diffusion process)
|
| 1206 |
+
noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)
|
| 1207 |
+
|
| 1208 |
+
# time ids
|
| 1209 |
+
def compute_time_ids(original_size, crops_coords_top_left):
|
| 1210 |
+
# Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
|
| 1211 |
+
target_size = (args.resolution, args.resolution)
|
| 1212 |
+
add_time_ids = list(original_size + crops_coords_top_left + target_size)
|
| 1213 |
+
add_time_ids = torch.tensor([add_time_ids])
|
| 1214 |
+
add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
|
| 1215 |
+
return add_time_ids
|
| 1216 |
+
|
| 1217 |
+
add_time_ids = torch.cat(
|
| 1218 |
+
[compute_time_ids(s, c) for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])]
|
| 1219 |
+
)
|
| 1220 |
+
|
| 1221 |
+
# Predict the noise residual
|
| 1222 |
+
unet_added_conditions = {"time_ids": add_time_ids}
|
| 1223 |
+
# prompt_embeds, pooled_prompt_embeds = encode_prompt(
|
| 1224 |
+
# text_encoders=[text_encoder_one, text_encoder_two],
|
| 1225 |
+
# tokenizers=None,
|
| 1226 |
+
# prompt=None,
|
| 1227 |
+
# text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
|
| 1228 |
+
# )
|
| 1229 |
+
mapper_outputs = simple_mapper(batch['codes'])
|
| 1230 |
+
prompt_embeds, pooled_prompt_embeds = encode_prompt(
|
| 1231 |
+
text_encoders=[text_encoder_one, text_encoder_two],
|
| 1232 |
+
text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]],
|
| 1233 |
+
placeholder_token_ids=placeholder_token_ids_one,
|
| 1234 |
+
mapper_outputs=[mapper_outputs[..., :768], mapper_outputs[..., 768:]]
|
| 1235 |
+
)
|
| 1236 |
+
|
| 1237 |
+
unet_added_conditions.update({"text_embeds": pooled_prompt_embeds})
|
| 1238 |
+
model_pred = unet(
|
| 1239 |
+
noisy_model_input, timesteps, prompt_embeds, added_cond_kwargs=unet_added_conditions
|
| 1240 |
+
).sample
|
| 1241 |
+
|
| 1242 |
+
# Get the target for loss depending on the prediction type
|
| 1243 |
+
if args.prediction_type is not None:
|
| 1244 |
+
# set prediction_type of scheduler if defined
|
| 1245 |
+
noise_scheduler.register_to_config(prediction_type=args.prediction_type)
|
| 1246 |
+
|
| 1247 |
+
if noise_scheduler.config.prediction_type == "epsilon":
|
| 1248 |
+
target = noise
|
| 1249 |
+
elif noise_scheduler.config.prediction_type == "v_prediction":
|
| 1250 |
+
target = noise_scheduler.get_velocity(model_input, noise, timesteps)
|
| 1251 |
+
else:
|
| 1252 |
+
raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
|
| 1253 |
+
|
| 1254 |
+
if args.snr_gamma is None:
|
| 1255 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
|
| 1256 |
+
attn_loss, max_attn = dreamcreature_loss(batch,
|
| 1257 |
+
unet,
|
| 1258 |
+
dino,
|
| 1259 |
+
seg,
|
| 1260 |
+
placeholder_token_ids_one,
|
| 1261 |
+
accelerator)
|
| 1262 |
+
if args.masked_training:
|
| 1263 |
+
masks = batch['masks'].unsqueeze(1).to(accelerator.device)
|
| 1264 |
+
loss_image_mask = F.interpolate(masks.float(),
|
| 1265 |
+
size=target.shape[-2:],
|
| 1266 |
+
mode='bilinear') * torch.ones_like(target)
|
| 1267 |
+
loss = loss * loss_image_mask
|
| 1268 |
+
loss = loss.sum() / loss_image_mask.sum()
|
| 1269 |
+
else:
|
| 1270 |
+
loss = loss.mean()
|
| 1271 |
+
else:
|
| 1272 |
+
# Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
|
| 1273 |
+
# Since we predict the noise instead of x_0, the original formulation is slightly changed.
|
| 1274 |
+
# This is discussed in Section 4.2 of the same paper.
|
| 1275 |
+
snr = compute_snr(noise_scheduler, timesteps)
|
| 1276 |
+
if noise_scheduler.config.prediction_type == "v_prediction":
|
| 1277 |
+
# Velocity objective requires that we add one to SNR values before we divide by them.
|
| 1278 |
+
snr = snr + 1
|
| 1279 |
+
mse_loss_weights = (
|
| 1280 |
+
torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
|
| 1281 |
+
)
|
| 1282 |
+
|
| 1283 |
+
loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
|
| 1284 |
+
attn_loss, max_attn = dreamcreature_loss(batch,
|
| 1285 |
+
unet,
|
| 1286 |
+
dino,
|
| 1287 |
+
seg,
|
| 1288 |
+
placeholder_token_ids_one,
|
| 1289 |
+
accelerator)
|
| 1290 |
+
if args.masked_training:
|
| 1291 |
+
masks = batch['masks'].unsqueeze(1).to(accelerator.device)
|
| 1292 |
+
loss_image_mask = F.interpolate(masks.float(),
|
| 1293 |
+
size=target.shape[-2:],
|
| 1294 |
+
mode='bilinear') * torch.ones_like(target)
|
| 1295 |
+
loss = loss * loss_image_mask
|
| 1296 |
+
loss = loss.sum(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
|
| 1297 |
+
loss = loss.sum() / loss_image_mask.sum()
|
| 1298 |
+
else:
|
| 1299 |
+
loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
|
| 1300 |
+
loss = loss.mean()
|
| 1301 |
+
|
| 1302 |
+
diff_loss = loss.clone().detach()
|
| 1303 |
+
avg_diff_loss = accelerator.gather(diff_loss.repeat(args.train_batch_size)).mean()
|
| 1304 |
+
train_diff_loss += avg_diff_loss.item() / args.gradient_accumulation_steps
|
| 1305 |
+
|
| 1306 |
+
avg_attn_loss = accelerator.gather(attn_loss.repeat(args.train_batch_size)).mean()
|
| 1307 |
+
train_attn_loss += avg_attn_loss.item() / args.gradient_accumulation_steps
|
| 1308 |
+
|
| 1309 |
+
loss += args.attn_loss * attn_loss
|
| 1310 |
+
|
| 1311 |
+
# Gather the losses across all processes for logging (if we use distributed training).
|
| 1312 |
+
avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
|
| 1313 |
+
train_loss += avg_loss.item() / args.gradient_accumulation_steps
|
| 1314 |
+
|
| 1315 |
+
# Backpropagate
|
| 1316 |
+
accelerator.backward(loss)
|
| 1317 |
+
if accelerator.sync_gradients:
|
| 1318 |
+
params_to_clip = (
|
| 1319 |
+
itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two)
|
| 1320 |
+
if args.train_text_encoder
|
| 1321 |
+
else unet_lora_parameters
|
| 1322 |
+
)
|
| 1323 |
+
params_to_clip = list(params_to_clip) + extra_params
|
| 1324 |
+
accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
|
| 1325 |
+
optimizer.step()
|
| 1326 |
+
lr_scheduler.step()
|
| 1327 |
+
optimizer.zero_grad()
|
| 1328 |
+
|
| 1329 |
+
# Checks if the accelerator has performed an optimization step behind the scenes
|
| 1330 |
+
if accelerator.sync_gradients:
|
| 1331 |
+
progress_bar.update(1)
|
| 1332 |
+
global_step += 1
|
| 1333 |
+
accelerator.log({"train_loss": train_loss,
|
| 1334 |
+
"diff_loss": train_diff_loss,
|
| 1335 |
+
"attn_loss": train_attn_loss,
|
| 1336 |
+
"max_attn": max_attn.item()
|
| 1337 |
+
}, step=global_step)
|
| 1338 |
+
train_loss = 0.0
|
| 1339 |
+
train_attn_loss = 0.0
|
| 1340 |
+
train_diff_loss = 0.0
|
| 1341 |
+
|
| 1342 |
+
if accelerator.is_main_process:
|
| 1343 |
+
if global_step % args.checkpointing_steps == 0:
|
| 1344 |
+
# _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
|
| 1345 |
+
if args.checkpoints_total_limit is not None:
|
| 1346 |
+
checkpoints = os.listdir(args.output_dir)
|
| 1347 |
+
checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
|
| 1348 |
+
checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
|
| 1349 |
+
|
| 1350 |
+
# before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
|
| 1351 |
+
if len(checkpoints) >= args.checkpoints_total_limit:
|
| 1352 |
+
num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
|
| 1353 |
+
removing_checkpoints = checkpoints[0:num_to_remove]
|
| 1354 |
+
|
| 1355 |
+
logger.info(
|
| 1356 |
+
f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
|
| 1357 |
+
)
|
| 1358 |
+
logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
|
| 1359 |
+
|
| 1360 |
+
for removing_checkpoint in removing_checkpoints:
|
| 1361 |
+
removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
|
| 1362 |
+
shutil.rmtree(removing_checkpoint)
|
| 1363 |
+
|
| 1364 |
+
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
|
| 1365 |
+
accelerator.save_state(save_path)
|
| 1366 |
+
logger.info(f"Saved state to {save_path}")
|
| 1367 |
+
|
| 1368 |
+
logs = {"step_loss": diff_loss.detach().item(),
|
| 1369 |
+
"attn_loss": attn_loss.detach().item(),
|
| 1370 |
+
"lr": lr_scheduler.get_last_lr()[0]}
|
| 1371 |
+
progress_bar.set_postfix(**logs)
|
| 1372 |
+
|
| 1373 |
+
if global_step >= args.max_train_steps:
|
| 1374 |
+
break
|
| 1375 |
+
|
| 1376 |
+
if accelerator.is_main_process:
|
| 1377 |
+
# todo: change pipeline
|
| 1378 |
+
if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
|
| 1379 |
+
logger.info(
|
| 1380 |
+
f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
|
| 1381 |
+
f" {args.validation_prompt}."
|
| 1382 |
+
)
|
| 1383 |
+
# create pipeline
|
| 1384 |
+
pipeline = DreamCreatureSDXLPipeline.from_pretrained(
|
| 1385 |
+
args.pretrained_model_name_or_path,
|
| 1386 |
+
vae=vae,
|
| 1387 |
+
tokenizer=tokenizer_one,
|
| 1388 |
+
tokenizer_2=tokenizer_two,
|
| 1389 |
+
text_encoder=accelerator.unwrap_model(text_encoder_one),
|
| 1390 |
+
text_encoder_2=accelerator.unwrap_model(text_encoder_two),
|
| 1391 |
+
unet=accelerator.unwrap_model(unet),
|
| 1392 |
+
revision=args.revision,
|
| 1393 |
+
variant=args.variant,
|
| 1394 |
+
torch_dtype=weight_dtype,
|
| 1395 |
+
)
|
| 1396 |
+
pipeline.placeholder_token_ids = placeholder_token_ids_one
|
| 1397 |
+
pipeline.simple_mapper = accelerator.unwrap_model(simple_mapper)
|
| 1398 |
+
pipeline.replace_token = False
|
| 1399 |
+
|
| 1400 |
+
pipeline = pipeline.to(accelerator.device)
|
| 1401 |
+
pipeline.set_progress_bar_config(disable=True)
|
| 1402 |
+
|
| 1403 |
+
# run inference
|
| 1404 |
+
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
|
| 1405 |
+
pipeline_args = {"prompt": args.validation_prompt}
|
| 1406 |
+
|
| 1407 |
+
num_steps = 4 if 'turbo' in args.pretrained_model_name_or_path else 25
|
| 1408 |
+
gs = 0 if 'turbo' in args.pretrained_model_name_or_path else 5.0
|
| 1409 |
+
|
| 1410 |
+
images = [
|
| 1411 |
+
pipeline(**pipeline_args, num_inference_steps=num_steps, guidance_scale=gs,
|
| 1412 |
+
generator=generator, height=args.resolution, width=args.resolution).images[0]
|
| 1413 |
+
for _ in range(args.num_validation_images)
|
| 1414 |
+
]
|
| 1415 |
+
|
| 1416 |
+
for tracker in accelerator.trackers:
|
| 1417 |
+
if tracker.name == "tensorboard":
|
| 1418 |
+
np_images = np.stack([np.asarray(img) for img in images])
|
| 1419 |
+
tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
|
| 1420 |
+
if tracker.name == "wandb":
|
| 1421 |
+
tracker.log(
|
| 1422 |
+
{
|
| 1423 |
+
"validation": [
|
| 1424 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 1425 |
+
for i, image in enumerate(images)
|
| 1426 |
+
]
|
| 1427 |
+
}
|
| 1428 |
+
)
|
| 1429 |
+
|
| 1430 |
+
del pipeline
|
| 1431 |
+
torch.cuda.empty_cache()
|
| 1432 |
+
|
| 1433 |
+
# Save the lora layers
|
| 1434 |
+
accelerator.wait_for_everyone()
|
| 1435 |
+
if accelerator.is_main_process:
|
| 1436 |
+
unet = accelerator.unwrap_model(unet)
|
| 1437 |
+
unet_lora_layers = unet_attn_processors_state_dict(unet)
|
| 1438 |
+
|
| 1439 |
+
if args.train_text_encoder:
|
| 1440 |
+
text_encoder_one = accelerator.unwrap_model(text_encoder_one)
|
| 1441 |
+
text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one)
|
| 1442 |
+
text_encoder_two = accelerator.unwrap_model(text_encoder_two)
|
| 1443 |
+
text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two)
|
| 1444 |
+
else:
|
| 1445 |
+
text_encoder_lora_layers = None
|
| 1446 |
+
text_encoder_2_lora_layers = None
|
| 1447 |
+
|
| 1448 |
+
StableDiffusionXLPipeline.save_lora_weights(
|
| 1449 |
+
save_directory=args.output_dir,
|
| 1450 |
+
unet_lora_layers=unet_lora_layers,
|
| 1451 |
+
text_encoder_lora_layers=text_encoder_lora_layers,
|
| 1452 |
+
text_encoder_2_lora_layers=text_encoder_2_lora_layers,
|
| 1453 |
+
)
|
| 1454 |
+
torch.save(simple_mapper.to(torch.float32).state_dict(), args.output_dir + '/hash_mapper.pth')
|
| 1455 |
+
|
| 1456 |
+
del unet
|
| 1457 |
+
del text_encoder_one
|
| 1458 |
+
del text_encoder_two
|
| 1459 |
+
del text_encoder_lora_layers
|
| 1460 |
+
del text_encoder_2_lora_layers
|
| 1461 |
+
del simple_mapper
|
| 1462 |
+
torch.cuda.empty_cache()
|
| 1463 |
+
|
| 1464 |
+
# Final inference
|
| 1465 |
+
# Load previous pipeline
|
| 1466 |
+
text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, simple_mapper = init_for_pipeline(args)
|
| 1467 |
+
pipeline = DreamCreatureSDXLPipeline.from_pretrained(
|
| 1468 |
+
args.pretrained_model_name_or_path,
|
| 1469 |
+
vae=vae,
|
| 1470 |
+
tokenizer=tokenizer_one,
|
| 1471 |
+
tokenizer_2=tokenizer_two,
|
| 1472 |
+
text_encoder=text_encoder_one,
|
| 1473 |
+
text_encoder_2=text_encoder_two,
|
| 1474 |
+
revision=args.revision,
|
| 1475 |
+
variant=args.variant,
|
| 1476 |
+
torch_dtype=weight_dtype,
|
| 1477 |
+
)
|
| 1478 |
+
pipeline.placeholder_token_ids = placeholder_token_ids_one
|
| 1479 |
+
pipeline.replace_token = False
|
| 1480 |
+
pipeline.simple_mapper = simple_mapper
|
| 1481 |
+
pipeline.simple_mapper.load_state_dict(torch.load(args.output_dir + '/hash_mapper.pth', map_location='cpu'))
|
| 1482 |
+
pipeline.simple_mapper.to(accelerator.device)
|
| 1483 |
+
setup_attn_processors(pipeline.unet, args)
|
| 1484 |
+
|
| 1485 |
+
pipeline = pipeline.to(accelerator.device)
|
| 1486 |
+
|
| 1487 |
+
# load attention processors
|
| 1488 |
+
pipeline.load_lora_weights(args.output_dir)
|
| 1489 |
+
|
| 1490 |
+
# run inference
|
| 1491 |
+
images = []
|
| 1492 |
+
if args.validation_prompt and args.num_validation_images > 0:
|
| 1493 |
+
num_steps = 4 if 'turbo' in args.pretrained_model_name_or_path else 25
|
| 1494 |
+
gs = 0 if 'turbo' in args.pretrained_model_name_or_path else 5.0
|
| 1495 |
+
generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
|
| 1496 |
+
images = [
|
| 1497 |
+
pipeline(args.validation_prompt, num_inference_steps=num_steps,
|
| 1498 |
+
guidance_scale=gs, generator=generator, height=args.resolution,
|
| 1499 |
+
width=args.resolution).images[0]
|
| 1500 |
+
for _ in range(args.num_validation_images)
|
| 1501 |
+
]
|
| 1502 |
+
|
| 1503 |
+
for tracker in accelerator.trackers:
|
| 1504 |
+
if tracker.name == "tensorboard":
|
| 1505 |
+
np_images = np.stack([np.asarray(img) for img in images])
|
| 1506 |
+
tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC")
|
| 1507 |
+
if tracker.name == "wandb":
|
| 1508 |
+
tracker.log(
|
| 1509 |
+
{
|
| 1510 |
+
"test": [
|
| 1511 |
+
wandb.Image(image, caption=f"{i}: {args.validation_prompt}")
|
| 1512 |
+
for i, image in enumerate(images)
|
| 1513 |
+
]
|
| 1514 |
+
}
|
| 1515 |
+
)
|
| 1516 |
+
|
| 1517 |
+
if args.push_to_hub:
|
| 1518 |
+
save_model_card(
|
| 1519 |
+
repo_id,
|
| 1520 |
+
images=images,
|
| 1521 |
+
base_model=args.pretrained_model_name_or_path,
|
| 1522 |
+
dataset_name=args.dataset_name,
|
| 1523 |
+
train_text_encoder=args.train_text_encoder,
|
| 1524 |
+
repo_folder=args.output_dir,
|
| 1525 |
+
vae_path=args.pretrained_vae_model_name_or_path,
|
| 1526 |
+
)
|
| 1527 |
+
upload_folder(
|
| 1528 |
+
repo_id=repo_id,
|
| 1529 |
+
folder_path=args.output_dir,
|
| 1530 |
+
commit_message="End of training",
|
| 1531 |
+
ignore_patterns=["step_*", "epoch_*"],
|
| 1532 |
+
)
|
| 1533 |
+
|
| 1534 |
+
accelerator.end_training()
|
| 1535 |
+
|
| 1536 |
+
|
| 1537 |
+
if __name__ == "__main__":
|
| 1538 |
+
args = parse_args()
|
| 1539 |
+
main(args)
|
train_kmeans_segmentation.ipynb
ADDED
|
@@ -0,0 +1,578 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"cells": [
|
| 3 |
+
{
|
| 4 |
+
"cell_type": "markdown",
|
| 5 |
+
"id": "1c073d83-8e73-407a-a669-3a837a90aac6",
|
| 6 |
+
"metadata": {},
|
| 7 |
+
"source": [
|
| 8 |
+
"# Compute DINO features"
|
| 9 |
+
]
|
| 10 |
+
},
|
| 11 |
+
{
|
| 12 |
+
"cell_type": "code",
|
| 13 |
+
"execution_count": 2,
|
| 14 |
+
"id": "901240c3-5111-4733-97ee-69891e4e7184",
|
| 15 |
+
"metadata": {},
|
| 16 |
+
"outputs": [],
|
| 17 |
+
"source": [
|
| 18 |
+
"import argparse\n",
|
| 19 |
+
"import math\n",
|
| 20 |
+
"import os\n",
|
| 21 |
+
"\n",
|
| 22 |
+
"import torch\n",
|
| 23 |
+
"import torchpq\n",
|
| 24 |
+
"from omegaconf import OmegaConf\n",
|
| 25 |
+
"from torch.utils.data import DataLoader\n",
|
| 26 |
+
"from sklearn.decomposition import PCA\n",
|
| 27 |
+
"from torchvision.transforms import transforms\n",
|
| 28 |
+
"from tqdm import tqdm\n",
|
| 29 |
+
"from transformers.utils import constants\n",
|
| 30 |
+
"\n",
|
| 31 |
+
"from dreamcreature.dino import DINO\n",
|
| 32 |
+
"from dreamcreature.dataset import ImageDataset\n",
|
| 33 |
+
"\n",
|
| 34 |
+
"MEAN = constants.IMAGENET_DEFAULT_MEAN\n",
|
| 35 |
+
"STD = constants.IMAGENET_DEFAULT_STD"
|
| 36 |
+
]
|
| 37 |
+
},
|
| 38 |
+
{
|
| 39 |
+
"cell_type": "code",
|
| 40 |
+
"execution_count": 4,
|
| 41 |
+
"id": "2f83c248-c3d5-4fe2-a111-ecdeda648214",
|
| 42 |
+
"metadata": {},
|
| 43 |
+
"outputs": [],
|
| 44 |
+
"source": [
|
| 45 |
+
"dataset_name = 'cub200_2011'\n",
|
| 46 |
+
"# dataset_name = 'dogs'\n",
|
| 47 |
+
"\n",
|
| 48 |
+
"rootdir = f'data/{dataset_name}'\n",
|
| 49 |
+
"resize = 256\n",
|
| 50 |
+
"crop = 224\n",
|
| 51 |
+
"\n",
|
| 52 |
+
"dataset = ImageDataset(rootdir,\n",
|
| 53 |
+
" 'train.txt',\n",
|
| 54 |
+
" transform=transforms.Compose([\n",
|
| 55 |
+
" transforms.Resize(resize, interpolation=transforms.InterpolationMode.BICUBIC),\n",
|
| 56 |
+
" transforms.CenterCrop(crop),\n",
|
| 57 |
+
" transforms.ToTensor(),\n",
|
| 58 |
+
" transforms.Normalize(MEAN, STD)\n",
|
| 59 |
+
" ]))"
|
| 60 |
+
]
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"cell_type": "code",
|
| 64 |
+
"execution_count": null,
|
| 65 |
+
"id": "946a88e5-b368-4cc3-a625-fd952650036b",
|
| 66 |
+
"metadata": {},
|
| 67 |
+
"outputs": [],
|
| 68 |
+
"source": [
|
| 69 |
+
"dataloader = DataLoader(dataset, 32, shuffle=False, drop_last=False, num_workers=4)\n",
|
| 70 |
+
"model = DINO()\n",
|
| 71 |
+
"model.eval()\n",
|
| 72 |
+
"\n",
|
| 73 |
+
"device = torch.device('cuda')\n",
|
| 74 |
+
"model = model.to(device)"
|
| 75 |
+
]
|
| 76 |
+
},
|
| 77 |
+
{
|
| 78 |
+
"cell_type": "code",
|
| 79 |
+
"execution_count": null,
|
| 80 |
+
"id": "f526a02a-745d-43df-94af-1a50ed438fda",
|
| 81 |
+
"metadata": {},
|
| 82 |
+
"outputs": [],
|
| 83 |
+
"source": [
|
| 84 |
+
"os.makedirs(config.rootdir + '/dinov2', exist_ok=True)\n",
|
| 85 |
+
"\n",
|
| 86 |
+
"image_feats = []\n",
|
| 87 |
+
"with tqdm(dataloader, bar_format='{l_bar}{bar:10}{r_bar}') as tepoch:\n",
|
| 88 |
+
" for i, (image, label, index) in enumerate(tepoch):\n",
|
| 89 |
+
" image = image.to(device)\n",
|
| 90 |
+
"\n",
|
| 91 |
+
" with torch.no_grad():\n",
|
| 92 |
+
" output = model.get_feat_maps(image) # (B, C, H, W)\n",
|
| 93 |
+
"\n",
|
| 94 |
+
" B, C, H, W = output.size()\n",
|
| 95 |
+
" output = output.reshape(B, C, H * W)\n",
|
| 96 |
+
" image_feats.append(output.cpu())\n",
|
| 97 |
+
"\n",
|
| 98 |
+
"image_feats = torch.cat(image_feats, dim=0) # (N, C, H*W)\n",
|
| 99 |
+
"torch.save(image_feats, rootdir + '/dinov2_image_feats.pth')"
|
| 100 |
+
]
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"cell_type": "markdown",
|
| 104 |
+
"id": "50f9ed32-5231-47d0-864e-cfbcd8b6d732",
|
| 105 |
+
"metadata": {},
|
| 106 |
+
"source": [
|
| 107 |
+
"# Train Kmeans Segmentation"
|
| 108 |
+
]
|
| 109 |
+
},
|
| 110 |
+
{
|
| 111 |
+
"cell_type": "code",
|
| 112 |
+
"execution_count": null,
|
| 113 |
+
"id": "da8b9bde-2c1a-40d5-a32c-dda98334fe17",
|
| 114 |
+
"metadata": {},
|
| 115 |
+
"outputs": [],
|
| 116 |
+
"source": [
|
| 117 |
+
"import torch\n",
|
| 118 |
+
"import random\n",
|
| 119 |
+
"import numpy as np\n",
|
| 120 |
+
"\n",
|
| 121 |
+
"dataset_name = 'cub200_2011'\n",
|
| 122 |
+
"# dataset_name = 'dogs'\n",
|
| 123 |
+
"\n",
|
| 124 |
+
"sd = torch.load(f'data/{dataset_name}/dinov2_image_feats.pth', map_location='cpu')\n",
|
| 125 |
+
"sd.size()"
|
| 126 |
+
]
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"cell_type": "code",
|
| 130 |
+
"execution_count": null,
|
| 131 |
+
"id": "67b0f5ba-e272-4aea-b17d-05a80e2ce025",
|
| 132 |
+
"metadata": {},
|
| 133 |
+
"outputs": [],
|
| 134 |
+
"source": [
|
| 135 |
+
"from dataset import code_to_int, int_to_caption\n",
|
| 136 |
+
"from dataset import ImageDataset\n",
|
| 137 |
+
"from torchvision.transforms import transforms\n",
|
| 138 |
+
"\n",
|
| 139 |
+
"ds = ImageDataset(f'data/{dataset_name}', transform=transforms.Compose([transforms.Resize(256), transforms.CenterCrop(224)]))\n",
|
| 140 |
+
"train_lines = open(f'data/{dataset_name}/train.txt').readlines()"
|
| 141 |
+
]
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"cell_type": "code",
|
| 145 |
+
"execution_count": null,
|
| 146 |
+
"id": "46c57aa3-de67-4d1a-b9a7-584687e437f5",
|
| 147 |
+
"metadata": {},
|
| 148 |
+
"outputs": [],
|
| 149 |
+
"source": [
|
| 150 |
+
"def set_seed(seed):\n",
|
| 151 |
+
" random.seed(seed)\n",
|
| 152 |
+
" np.random.seed(seed)\n",
|
| 153 |
+
" torch.manual_seed(seed)\n",
|
| 154 |
+
" torch.cuda.manual_seed(seed)\n",
|
| 155 |
+
"\n",
|
| 156 |
+
"set_seed(42)\n",
|
| 157 |
+
" \n",
|
| 158 |
+
"n = 100 # use small training sample to avoid OOM\n",
|
| 159 |
+
"randidx = torch.randperm(len(sd))[:n]\n",
|
| 160 |
+
"randsd = sd[randidx].permute(0, 2, 1) # (N, HW, C)\n",
|
| 161 |
+
"randsd.size()"
|
| 162 |
+
]
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"cell_type": "code",
|
| 166 |
+
"execution_count": null,
|
| 167 |
+
"id": "fab9e741-ff9a-4b6a-8ee3-41bc1d43cb52",
|
| 168 |
+
"metadata": {},
|
| 169 |
+
"outputs": [],
|
| 170 |
+
"source": [
|
| 171 |
+
"import numpy as np\n",
|
| 172 |
+
"import torchpq\n",
|
| 173 |
+
"import torch.nn.functional as F\n",
|
| 174 |
+
"import matplotlib.pyplot as plt\n",
|
| 175 |
+
"import random\n",
|
| 176 |
+
"from sklearn.decomposition import PCA\n",
|
| 177 |
+
"\n",
|
| 178 |
+
"set_seed(42)\n",
|
| 179 |
+
"\n",
|
| 180 |
+
"fg_kmeans = torchpq.clustering.KMeans(n_clusters=2,\n",
|
| 181 |
+
" distance=\"cosine\",\n",
|
| 182 |
+
" verbose=1,\n",
|
| 183 |
+
" n_redo=5,\n",
|
| 184 |
+
" max_iter=1000)\n",
|
| 185 |
+
"fg_labels = fg_kmeans.fit(randsd.reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(n, -1)"
|
| 186 |
+
]
|
| 187 |
+
},
|
| 188 |
+
{
|
| 189 |
+
"cell_type": "code",
|
| 190 |
+
"execution_count": null,
|
| 191 |
+
"id": "9a12550b-1c66-48a0-9bd6-a39b928cf57d",
|
| 192 |
+
"metadata": {},
|
| 193 |
+
"outputs": [],
|
| 194 |
+
"source": [
|
| 195 |
+
"torch.unique(fg_labels, return_counts=True)"
|
| 196 |
+
]
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"cell_type": "code",
|
| 200 |
+
"execution_count": null,
|
| 201 |
+
"id": "b9a3b9f0-fc73-4e44-bca8-a2a2da16df38",
|
| 202 |
+
"metadata": {},
|
| 203 |
+
"outputs": [],
|
| 204 |
+
"source": [
|
| 205 |
+
"for i in range(100):\n",
|
| 206 |
+
" plt.subplot(10, 10, i+1)\n",
|
| 207 |
+
" plt.imshow(fg_labels[i].reshape(16, 16))\n",
|
| 208 |
+
" plt.axis('off')"
|
| 209 |
+
]
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"cell_type": "code",
|
| 213 |
+
"execution_count": null,
|
| 214 |
+
"id": "6f8e6121-74ee-4539-bf61-9d0ba2198ef5",
|
| 215 |
+
"metadata": {},
|
| 216 |
+
"outputs": [],
|
| 217 |
+
"source": [
|
| 218 |
+
"fg_idx = 0 # this have to do manual inspection, based on the visualization above\n",
|
| 219 |
+
"bg_idx = 1 - fg_idx\n",
|
| 220 |
+
"\n",
|
| 221 |
+
"randsd_bgnorm = []\n",
|
| 222 |
+
"randsd_nobg = []\n",
|
| 223 |
+
"randsd_bgmean = []\n",
|
| 224 |
+
"\n",
|
| 225 |
+
"for i in range(n):\n",
|
| 226 |
+
" bgnorm_mean = randsd[i][fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)\n",
|
| 227 |
+
" \n",
|
| 228 |
+
" if fg_idx == 0:\n",
|
| 229 |
+
" bg_mask = fg_labels[i]\n",
|
| 230 |
+
" else:\n",
|
| 231 |
+
" bg_mask = 1 - fg_labels[i]\n",
|
| 232 |
+
" \n",
|
| 233 |
+
" bg_mask = bg_mask.unsqueeze(1)\n",
|
| 234 |
+
" bgnorm = (randsd[i] * (1 - bg_mask)) + (bgnorm_mean * bg_mask)\n",
|
| 235 |
+
" \n",
|
| 236 |
+
" randsd_bgnorm.append(bgnorm)\n",
|
| 237 |
+
" randsd_nobg.append(randsd[i] * (1 - bg_mask) + (-1 * bg_mask))\n",
|
| 238 |
+
" randsd_bgmean.append(bgnorm_mean)\n",
|
| 239 |
+
" \n",
|
| 240 |
+
"randsd_bgnorm = torch.stack(randsd_bgnorm, dim=0)\n",
|
| 241 |
+
"randsd_nobg = torch.stack(randsd_nobg, dim=0)\n",
|
| 242 |
+
"randsd_bgmean = torch.cat(randsd_bgmean, dim=0)"
|
| 243 |
+
]
|
| 244 |
+
},
|
| 245 |
+
{
|
| 246 |
+
"cell_type": "code",
|
| 247 |
+
"execution_count": null,
|
| 248 |
+
"id": "f90785f8-d1e4-4dc9-9e6d-c242058639a5",
|
| 249 |
+
"metadata": {},
|
| 250 |
+
"outputs": [],
|
| 251 |
+
"source": [
|
| 252 |
+
"set_seed(42)\n",
|
| 253 |
+
"M = 8\n",
|
| 254 |
+
"\n",
|
| 255 |
+
"coarse_kmeans = torchpq.clustering.KMeans(n_clusters=M,\n",
|
| 256 |
+
" distance=\"cosine\",\n",
|
| 257 |
+
" verbose=1,\n",
|
| 258 |
+
" n_redo=5,\n",
|
| 259 |
+
" max_iter=1000)\n",
|
| 260 |
+
"coarse_labels = coarse_kmeans.fit(randsd_nobg.reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(n, -1)"
|
| 261 |
+
]
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"cell_type": "code",
|
| 265 |
+
"execution_count": null,
|
| 266 |
+
"id": "d4930216-de8a-420b-b7cd-45260299b7b9",
|
| 267 |
+
"metadata": {},
|
| 268 |
+
"outputs": [],
|
| 269 |
+
"source": [
|
| 270 |
+
"for i in range(100):\n",
|
| 271 |
+
" plt.subplot(10, 10, i+1)\n",
|
| 272 |
+
" plt.imshow(coarse_labels[i].reshape(16, 16))\n",
|
| 273 |
+
" plt.axis('off')"
|
| 274 |
+
]
|
| 275 |
+
},
|
| 276 |
+
{
|
| 277 |
+
"cell_type": "code",
|
| 278 |
+
"execution_count": null,
|
| 279 |
+
"id": "d5e72d5e-df24-485d-85aa-de17ba381d84",
|
| 280 |
+
"metadata": {},
|
| 281 |
+
"outputs": [],
|
| 282 |
+
"source": [
|
| 283 |
+
"import torch\n",
|
| 284 |
+
"import numpy as np\n",
|
| 285 |
+
"import matplotlib.pyplot as plt\n",
|
| 286 |
+
"\n",
|
| 287 |
+
"disp = coarse_labels[0].reshape(16, 16)\n",
|
| 288 |
+
"\n",
|
| 289 |
+
"plt.imshow(disp)\n",
|
| 290 |
+
"plt.axis('off')"
|
| 291 |
+
]
|
| 292 |
+
},
|
| 293 |
+
{
|
| 294 |
+
"cell_type": "code",
|
| 295 |
+
"execution_count": null,
|
| 296 |
+
"id": "b508e11d-082a-40a6-83db-4d306c0f9f00",
|
| 297 |
+
"metadata": {},
|
| 298 |
+
"outputs": [],
|
| 299 |
+
"source": [
|
| 300 |
+
"torch.unique(coarse_labels, return_counts=True)"
|
| 301 |
+
]
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"cell_type": "code",
|
| 305 |
+
"execution_count": null,
|
| 306 |
+
"id": "8f4bf418-2297-4dc5-8bdd-15af7cf44c7a",
|
| 307 |
+
"metadata": {},
|
| 308 |
+
"outputs": [],
|
| 309 |
+
"source": [
|
| 310 |
+
"sd_bgnorm = []\n",
|
| 311 |
+
"sd_nobg = []\n",
|
| 312 |
+
"sd_bgmean = []\n",
|
| 313 |
+
"\n",
|
| 314 |
+
"inp = sd.permute(0, 2, 1)\n",
|
| 315 |
+
"N = inp.size(0)\n",
|
| 316 |
+
"\n",
|
| 317 |
+
"sd_fg_labels = []\n",
|
| 318 |
+
"bs = 1000\n",
|
| 319 |
+
"for bidx in range(N // bs + 1):\n",
|
| 320 |
+
" if bidx * bs >= N:\n",
|
| 321 |
+
" break\n",
|
| 322 |
+
" \n",
|
| 323 |
+
" start_bidx = bidx*bs\n",
|
| 324 |
+
" end_bidx = min((bidx+1)*bs, N)\n",
|
| 325 |
+
" \n",
|
| 326 |
+
" sd_fg_labels.append(fg_kmeans.predict(inp[start_bidx:end_bidx].reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(end_bidx - start_bidx, -1))\n",
|
| 327 |
+
" \n",
|
| 328 |
+
"sd_fg_labels = torch.cat(sd_fg_labels, dim=0)\n",
|
| 329 |
+
"\n",
|
| 330 |
+
"for i in range(N):\n",
|
| 331 |
+
" bgnorm_mean = inp[i][sd_fg_labels[i] == bg_idx].mean(dim=0, keepdim=True)\n",
|
| 332 |
+
" \n",
|
| 333 |
+
" if fg_idx == 0:\n",
|
| 334 |
+
" bg_mask = sd_fg_labels[i]\n",
|
| 335 |
+
" else:\n",
|
| 336 |
+
" bg_mask = 1 - sd_fg_labels[i]\n",
|
| 337 |
+
" \n",
|
| 338 |
+
" bg_mask = bg_mask.unsqueeze(1)\n",
|
| 339 |
+
" bgnorm = (inp[i] * (1 - bg_mask)) + (bgnorm_mean * bg_mask)\n",
|
| 340 |
+
" \n",
|
| 341 |
+
" sd_bgnorm.append(bgnorm)\n",
|
| 342 |
+
" sd_nobg.append(inp[i] * (1 - bg_mask) + (-1 * bg_mask))\n",
|
| 343 |
+
" sd_bgmean.append(bgnorm_mean)\n",
|
| 344 |
+
" print(i, end='\\r')\n",
|
| 345 |
+
" \n",
|
| 346 |
+
"sd_bgnorm = torch.stack(sd_bgnorm, dim=0)\n",
|
| 347 |
+
"sd_nobg = torch.stack(sd_nobg, dim=0)\n",
|
| 348 |
+
"sd_bgmean = torch.cat(sd_bgmean, dim=0)"
|
| 349 |
+
]
|
| 350 |
+
},
|
| 351 |
+
{
|
| 352 |
+
"cell_type": "code",
|
| 353 |
+
"execution_count": null,
|
| 354 |
+
"id": "d8026046-2ac5-4d62-9d18-4cbf7e2ebdb0",
|
| 355 |
+
"metadata": {},
|
| 356 |
+
"outputs": [],
|
| 357 |
+
"source": [
|
| 358 |
+
"sd_coarse_labels = []\n",
|
| 359 |
+
"bs = 1000\n",
|
| 360 |
+
"for bidx in range(N // bs + 1):\n",
|
| 361 |
+
" if bidx * bs >= N:\n",
|
| 362 |
+
" break\n",
|
| 363 |
+
" \n",
|
| 364 |
+
" start_bidx = bidx*bs\n",
|
| 365 |
+
" end_bidx = min((bidx+1)*bs, N)\n",
|
| 366 |
+
" \n",
|
| 367 |
+
" sd_coarse_labels.append(coarse_kmeans.predict(sd_nobg[start_bidx:end_bidx].reshape(-1, 768).t().contiguous().cuda()).cpu().reshape(end_bidx - start_bidx, -1))\n",
|
| 368 |
+
" \n",
|
| 369 |
+
"sd_coarse_labels = torch.cat(sd_coarse_labels, dim=0)"
|
| 370 |
+
]
|
| 371 |
+
},
|
| 372 |
+
{
|
| 373 |
+
"cell_type": "code",
|
| 374 |
+
"execution_count": null,
|
| 375 |
+
"id": "80a46c2d-6cd4-4344-b56f-4ff1fa7235d3",
|
| 376 |
+
"metadata": {},
|
| 377 |
+
"outputs": [],
|
| 378 |
+
"source": [
|
| 379 |
+
"for i in range(100):\n",
|
| 380 |
+
" plt.subplot(10, 10, i+1)\n",
|
| 381 |
+
" coarse_mask = sd_coarse_labels[i].reshape(16, 16)\n",
|
| 382 |
+
" plt.imshow(coarse_mask)\n",
|
| 383 |
+
" plt.axis('off')"
|
| 384 |
+
]
|
| 385 |
+
},
|
| 386 |
+
{
|
| 387 |
+
"cell_type": "code",
|
| 388 |
+
"execution_count": null,
|
| 389 |
+
"id": "e41a997b-cccc-40c2-a983-c6092ffe69be",
|
| 390 |
+
"metadata": {},
|
| 391 |
+
"outputs": [],
|
| 392 |
+
"source": [
|
| 393 |
+
"torch.save(sd_coarse_labels.reshape(N, 16, 16).long().cpu(), f'data/{dataset_name}/coarse_mask_m8.pth')"
|
| 394 |
+
]
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"cell_type": "code",
|
| 398 |
+
"execution_count": null,
|
| 399 |
+
"id": "cc918d47-35ca-4177-b63b-a12f4f4a3d5a",
|
| 400 |
+
"metadata": {},
|
| 401 |
+
"outputs": [],
|
| 402 |
+
"source": [
|
| 403 |
+
"torch.unique(sd_coarse_labels, return_counts=True)"
|
| 404 |
+
]
|
| 405 |
+
},
|
| 406 |
+
{
|
| 407 |
+
"cell_type": "code",
|
| 408 |
+
"execution_count": null,
|
| 409 |
+
"id": "9d4beb2d-e385-4ecc-9f90-287d7bc13c0d",
|
| 410 |
+
"metadata": {},
|
| 411 |
+
"outputs": [],
|
| 412 |
+
"source": [
|
| 413 |
+
"from tqdm.auto import tqdm\n",
|
| 414 |
+
"\n",
|
| 415 |
+
"sd_fgmean = []\n",
|
| 416 |
+
"\n",
|
| 417 |
+
"inp = sd.permute(0, 2, 1)\n",
|
| 418 |
+
"N = inp.size(0)\n",
|
| 419 |
+
"M = 8\n",
|
| 420 |
+
"\n",
|
| 421 |
+
"for i in tqdm(range(N)):\n",
|
| 422 |
+
" mean_feats = []\n",
|
| 423 |
+
" for m in range(M):\n",
|
| 424 |
+
" coarse_mask = sd_coarse_labels[i] == m\n",
|
| 425 |
+
" if coarse_mask.sum().item() == 0:\n",
|
| 426 |
+
" m_mean_feats = torch.zeros(1, 768)\n",
|
| 427 |
+
" else:\n",
|
| 428 |
+
" m_mean_feats = inp[i][coarse_mask].mean(dim=0, keepdim=True)\n",
|
| 429 |
+
" \n",
|
| 430 |
+
" mean_feats.append(m_mean_feats)\n",
|
| 431 |
+
" \n",
|
| 432 |
+
" mean_feats = torch.cat(mean_feats, dim=0)\n",
|
| 433 |
+
" sd_fgmean.append(mean_feats)\n",
|
| 434 |
+
" print(i, end='\\r')\n",
|
| 435 |
+
" \n",
|
| 436 |
+
"sd_fgmean = torch.stack(sd_fgmean, dim=0)"
|
| 437 |
+
]
|
| 438 |
+
},
|
| 439 |
+
{
|
| 440 |
+
"cell_type": "code",
|
| 441 |
+
"execution_count": null,
|
| 442 |
+
"id": "5034a91a-ae6d-468b-a738-f9ec0d019d72",
|
| 443 |
+
"metadata": {},
|
| 444 |
+
"outputs": [],
|
| 445 |
+
"source": [
|
| 446 |
+
"N = inp.size(0)\n",
|
| 447 |
+
"M = 8\n",
|
| 448 |
+
"K = 256\n",
|
| 449 |
+
"bgm = {'cub200_2011': 7, 'dogs': 1}[dataset_name] # 7 for cub, 1 for dog, this means which index is background\n",
|
| 450 |
+
"\n",
|
| 451 |
+
"final_labels = torch.ones(N, M) * K\n",
|
| 452 |
+
"\n",
|
| 453 |
+
"set_seed(42)\n",
|
| 454 |
+
"\n",
|
| 455 |
+
"zero_mean_idxs = []\n",
|
| 456 |
+
"fine_feats = []\n",
|
| 457 |
+
"fine_kmeans_trained = []\n",
|
| 458 |
+
"\n",
|
| 459 |
+
"for m in range(M):\n",
|
| 460 |
+
" fine_kmeans = torchpq.clustering.KMeans(n_clusters=K,\n",
|
| 461 |
+
" distance=\"cosine\",\n",
|
| 462 |
+
" verbose=1,\n",
|
| 463 |
+
" n_redo=5,\n",
|
| 464 |
+
" max_iter=1000)\n",
|
| 465 |
+
" \n",
|
| 466 |
+
" if m == bgm:\n",
|
| 467 |
+
" fine_labels = fine_kmeans.fit(sd_bgmean.t().contiguous().cuda()).cpu()\n",
|
| 468 |
+
" final_labels[:, m] = fine_labels\n",
|
| 469 |
+
" else:\n",
|
| 470 |
+
" fine_inp = sd_fgmean[:, m].reshape(-1, 768)\n",
|
| 471 |
+
" fine_labels = fine_kmeans.fit(fine_inp.t().contiguous().cuda()).cpu()\n",
|
| 472 |
+
" \n",
|
| 473 |
+
" final_labels[:, m] = fine_labels\n",
|
| 474 |
+
" \n",
|
| 475 |
+
" fine_kmeans_trained.append(fine_kmeans)\n",
|
| 476 |
+
" \n",
|
| 477 |
+
" fine_feats.append(fine_kmeans.centroizds.cpu().t()[fine_labels])\n",
|
| 478 |
+
" \n",
|
| 479 |
+
" print('zero mean', torch.arange(K)[fine_kmeans.centroids.t().sum(dim=-1).cpu() == 0].tolist())\n",
|
| 480 |
+
" zero_mean_idxs.append(torch.arange(K)[fine_kmeans.centroids.t().sum(dim=-1).cpu() == 0].tolist())\n",
|
| 481 |
+
" \n",
|
| 482 |
+
"fine_feats = torch.cat(fine_feats, dim=1)\n",
|
| 483 |
+
"print(fine_feats.size())"
|
| 484 |
+
]
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"cell_type": "code",
|
| 488 |
+
"execution_count": null,
|
| 489 |
+
"id": "450bc5c1-cc41-44b3-ab6a-c0c65eb1210a",
|
| 490 |
+
"metadata": {},
|
| 491 |
+
"outputs": [],
|
| 492 |
+
"source": [
|
| 493 |
+
"torch.save({\n",
|
| 494 |
+
" 'foreground_background': fg_kmeans,\n",
|
| 495 |
+
" 'coarse_kmeans': coarse_kmeans,\n",
|
| 496 |
+
" 'fine_kmeans': fine_kmeans_trained,\n",
|
| 497 |
+
"}, f'data/{dataset_name}/pretrained_kmeans.pth')"
|
| 498 |
+
]
|
| 499 |
+
},
|
| 500 |
+
{
|
| 501 |
+
"cell_type": "code",
|
| 502 |
+
"execution_count": null,
|
| 503 |
+
"id": "9ad2f095-f99d-489e-b645-8933b6f66372",
|
| 504 |
+
"metadata": {},
|
| 505 |
+
"outputs": [],
|
| 506 |
+
"source": [
|
| 507 |
+
"from tqdm.auto import tqdm\n",
|
| 508 |
+
"\n",
|
| 509 |
+
"final_code_captions = []\n",
|
| 510 |
+
"counts = [[0 for _ in range(K)] for _ in range(M)]\n",
|
| 511 |
+
"\n",
|
| 512 |
+
"for i in tqdm(range(N)):\n",
|
| 513 |
+
" m_labels = final_labels[i] # M\n",
|
| 514 |
+
" \n",
|
| 515 |
+
" line = []\n",
|
| 516 |
+
" for m in range(M):\n",
|
| 517 |
+
" k = m_labels[m].long().item()\n",
|
| 518 |
+
" \n",
|
| 519 |
+
" if k not in zero_mean_idxs[m]:\n",
|
| 520 |
+
" line.append(f'{m}:{k}')\n",
|
| 521 |
+
" counts[m][k] += 1\n",
|
| 522 |
+
" \n",
|
| 523 |
+
" assert len(line) != 0, f'error at {i}'\n",
|
| 524 |
+
" final_code_captions.append(' '.join(line))"
|
| 525 |
+
]
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"cell_type": "code",
|
| 529 |
+
"execution_count": null,
|
| 530 |
+
"id": "1be7233d-d9fe-4e0a-bda9-93fc45f8e982",
|
| 531 |
+
"metadata": {},
|
| 532 |
+
"outputs": [],
|
| 533 |
+
"source": [
|
| 534 |
+
"import matplotlib.pyplot as plt\n",
|
| 535 |
+
"\n",
|
| 536 |
+
"for m in range(M):\n",
|
| 537 |
+
" if max(counts[m]) == 0:\n",
|
| 538 |
+
" continue\n",
|
| 539 |
+
" \n",
|
| 540 |
+
" plt.scatter(range(K), counts[m])\n",
|
| 541 |
+
" print(m, min(counts[m]), max(counts[m]), np.mean(counts[m]))"
|
| 542 |
+
]
|
| 543 |
+
},
|
| 544 |
+
{
|
| 545 |
+
"cell_type": "code",
|
| 546 |
+
"execution_count": null,
|
| 547 |
+
"id": "e344ab48-fba6-4c17-a88c-e2528c0ac5cf",
|
| 548 |
+
"metadata": {},
|
| 549 |
+
"outputs": [],
|
| 550 |
+
"source": [
|
| 551 |
+
"with open(f'data/{dataset_name}/train_caps_better_m{M}_k{K}.txt', 'w+') as f:\n",
|
| 552 |
+
" for line in final_code_captions:\n",
|
| 553 |
+
" f.write(line + '\\n')"
|
| 554 |
+
]
|
| 555 |
+
}
|
| 556 |
+
],
|
| 557 |
+
"metadata": {
|
| 558 |
+
"kernelspec": {
|
| 559 |
+
"display_name": "Python 3 (ipykernel)",
|
| 560 |
+
"language": "python",
|
| 561 |
+
"name": "python3"
|
| 562 |
+
},
|
| 563 |
+
"language_info": {
|
| 564 |
+
"codemirror_mode": {
|
| 565 |
+
"name": "ipython",
|
| 566 |
+
"version": 3
|
| 567 |
+
},
|
| 568 |
+
"file_extension": ".py",
|
| 569 |
+
"mimetype": "text/x-python",
|
| 570 |
+
"name": "python",
|
| 571 |
+
"nbconvert_exporter": "python",
|
| 572 |
+
"pygments_lexer": "ipython3",
|
| 573 |
+
"version": "3.10.8"
|
| 574 |
+
}
|
| 575 |
+
},
|
| 576 |
+
"nbformat": 4,
|
| 577 |
+
"nbformat_minor": 5
|
| 578 |
+
}
|
utils.py
ADDED
|
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from diffusers.models.attention_processor import LoRAAttnProcessor
|
| 3 |
+
|
| 4 |
+
|
| 5 |
+
def add_tokens(tokenizer, text_encoder, placeholder_token, num_vec_per_token=1, initializer_token=None):
|
| 6 |
+
"""
|
| 7 |
+
Add tokens to the tokenizer and set the initial value of token embeddings
|
| 8 |
+
"""
|
| 9 |
+
tokenizer.add_placeholder_tokens(placeholder_token, num_vec_per_token=num_vec_per_token)
|
| 10 |
+
text_encoder.resize_token_embeddings(len(tokenizer))
|
| 11 |
+
token_embeds = text_encoder.get_input_embeddings().weight.data
|
| 12 |
+
placeholder_token_ids = tokenizer.encode(placeholder_token, add_special_tokens=False)
|
| 13 |
+
if initializer_token:
|
| 14 |
+
token_ids = tokenizer.encode(initializer_token, add_special_tokens=False)
|
| 15 |
+
for i, placeholder_token_id in enumerate(placeholder_token_ids):
|
| 16 |
+
token_embeds[placeholder_token_id] = token_embeds[token_ids[i * len(token_ids) // num_vec_per_token]]
|
| 17 |
+
else:
|
| 18 |
+
for i, placeholder_token_id in enumerate(placeholder_token_ids):
|
| 19 |
+
token_embeds[placeholder_token_id] = torch.randn_like(token_embeds[placeholder_token_id])
|
| 20 |
+
return placeholder_token_ids
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def tokenize_prompt(tokenizer, prompt, replace_token=False):
|
| 24 |
+
text_inputs = tokenizer(
|
| 25 |
+
prompt,
|
| 26 |
+
replace_token=replace_token,
|
| 27 |
+
padding="max_length",
|
| 28 |
+
max_length=tokenizer.model_max_length,
|
| 29 |
+
truncation=True,
|
| 30 |
+
return_tensors="pt",
|
| 31 |
+
)
|
| 32 |
+
text_input_ids = text_inputs.input_ids
|
| 33 |
+
return text_input_ids
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
def get_processor(self, return_deprecated_lora: bool = False):
|
| 37 |
+
r"""
|
| 38 |
+
Get the attention processor in use.
|
| 39 |
+
|
| 40 |
+
Args:
|
| 41 |
+
return_deprecated_lora (`bool`, *optional*, defaults to `False`):
|
| 42 |
+
Set to `True` to return the deprecated LoRA attention processor.
|
| 43 |
+
|
| 44 |
+
Returns:
|
| 45 |
+
"AttentionProcessor": The attention processor in use.
|
| 46 |
+
"""
|
| 47 |
+
if not return_deprecated_lora:
|
| 48 |
+
return self.processor
|
| 49 |
+
|
| 50 |
+
# TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible
|
| 51 |
+
# serialization format for LoRA Attention Processors. It should be deleted once the integration
|
| 52 |
+
# with PEFT is completed.
|
| 53 |
+
is_lora_activated = {
|
| 54 |
+
name: module.lora_layer is not None
|
| 55 |
+
for name, module in self.named_modules()
|
| 56 |
+
if hasattr(module, "lora_layer")
|
| 57 |
+
}
|
| 58 |
+
|
| 59 |
+
# 1. if no layer has a LoRA activated we can return the processor as usual
|
| 60 |
+
if not any(is_lora_activated.values()):
|
| 61 |
+
return self.processor
|
| 62 |
+
|
| 63 |
+
# If doesn't apply LoRA do `add_k_proj` or `add_v_proj`
|
| 64 |
+
is_lora_activated.pop("add_k_proj", None)
|
| 65 |
+
is_lora_activated.pop("add_v_proj", None)
|
| 66 |
+
# 2. else it is not posssible that only some layers have LoRA activated
|
| 67 |
+
if not all(is_lora_activated.values()):
|
| 68 |
+
raise ValueError(
|
| 69 |
+
f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"
|
| 70 |
+
)
|
| 71 |
+
|
| 72 |
+
# 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor
|
| 73 |
+
# non_lora_processor_cls_name = self.processor.__class__.__name__
|
| 74 |
+
# lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name)
|
| 75 |
+
|
| 76 |
+
hidden_size = self.inner_dim
|
| 77 |
+
|
| 78 |
+
# now create a LoRA attention processor from the LoRA layers
|
| 79 |
+
kwargs = {
|
| 80 |
+
"cross_attention_dim": self.cross_attention_dim,
|
| 81 |
+
"rank": self.to_q.lora_layer.rank,
|
| 82 |
+
"network_alpha": self.to_q.lora_layer.network_alpha,
|
| 83 |
+
"q_rank": self.to_q.lora_layer.rank,
|
| 84 |
+
"q_hidden_size": self.to_q.lora_layer.out_features,
|
| 85 |
+
"k_rank": self.to_k.lora_layer.rank,
|
| 86 |
+
"k_hidden_size": self.to_k.lora_layer.out_features,
|
| 87 |
+
"v_rank": self.to_v.lora_layer.rank,
|
| 88 |
+
"v_hidden_size": self.to_v.lora_layer.out_features,
|
| 89 |
+
"out_rank": self.to_out[0].lora_layer.rank,
|
| 90 |
+
"out_hidden_size": self.to_out[0].lora_layer.out_features,
|
| 91 |
+
}
|
| 92 |
+
|
| 93 |
+
if hasattr(self.processor, "attention_op"):
|
| 94 |
+
kwargs["attention_op"] = self.processor.attention_op
|
| 95 |
+
|
| 96 |
+
lora_processor = LoRAAttnProcessor(hidden_size, **kwargs)
|
| 97 |
+
lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
|
| 98 |
+
lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
|
| 99 |
+
lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
|
| 100 |
+
lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
|
| 101 |
+
|
| 102 |
+
return lora_processor
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
def get_attn_processors(self):
|
| 106 |
+
r"""
|
| 107 |
+
Returns:
|
| 108 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
| 109 |
+
indexed by its weight name.
|
| 110 |
+
"""
|
| 111 |
+
# set recursively
|
| 112 |
+
processors = {}
|
| 113 |
+
|
| 114 |
+
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors):
|
| 115 |
+
if hasattr(module, "get_processor"):
|
| 116 |
+
processors[f"{name}.processor"] = get_processor(module, return_deprecated_lora=True)
|
| 117 |
+
|
| 118 |
+
for sub_name, child in module.named_children():
|
| 119 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
| 120 |
+
|
| 121 |
+
return processors
|
| 122 |
+
|
| 123 |
+
for name, module in self.named_children():
|
| 124 |
+
fn_recursive_add_processors(name, module, processors)
|
| 125 |
+
|
| 126 |
+
return processors
|
| 127 |
+
|