Spaces:
Running
on
Zero
Running
on
Zero
update examles
Browse files
examples/default/input_params/output_20250426091716_0_input_params.json
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"prompt": "anime, cute female vocals, kawaii pop, j-pop, childish, piano, guitar, synthesizer, fast, happy, cheerful, lighthearted",
|
3 |
+
"lyrics": "[Chorus]\nใญใใ้กใ่ตคใใ๏ผ\nใฉใใใใฎ๏ผ ็ฑใใใใฎ๏ผ\nใใใจใๆใฃใฆใใฎ๏ผ\nใญใใ่จใฃใฆใ๏ผ\n\nใฉใใใฆใใใช็ฎใง่ฆใใฎ๏ผ\n็งใๆชใใใจใใ๏ผ\nไฝใ้้ใใใฎ๏ผ\nใ้กใใใใใฆโฆ ๆใใใโฆ\nใ ใใใใใใฆใโฆ\n\n[Bridge]\n็ฎใ้ใใฆใใใใฃใจ่ใๅใใฆใ\nไฝใ่ฆใชใใฃใใใชใใใใใ\nๆใใชใใงโฆ ่จฑใใฆใโฆ\n\n[Chorus]\nใญใใ้กใ่ตคใใ๏ผ\nใฉใใใใฎ๏ผ ็ฑใใใใฎ๏ผ\nใใใจใๆใฃใฆใใฎ๏ผ\nใญใใ่จใฃใฆใ๏ผ\n\nใฉใใใฆใใใช็ฎใง่ฆใใฎ๏ผ\n็งใๆชใใใจใใ๏ผ\nไฝใ้้ใใใฎ๏ผ\nใ้กใใใใใฆโฆ ๆใใใโฆ\nใ ใใใใใใฆใโฆ\n\n[Bridge 2]\nๅพ
ใฃใฆใใใ็งใๆชใใชใใ\nใใใใชใใใฃใฆ่จใใใใ\nใขใคในใฏใชใผใ ใใใใใใ\nใใๆใใชใใง๏ผ\n\nOooohโฆ ่จใฃใฆใ๏ผ",
|
4 |
+
"audio_duration": 160,
|
5 |
+
"infer_step": 60,
|
6 |
+
"guidance_scale": 15,
|
7 |
+
"scheduler_type": "euler",
|
8 |
+
"cfg_type": "apg",
|
9 |
+
"omega_scale": 10,
|
10 |
+
"guidance_interval": 0.5,
|
11 |
+
"guidance_interval_decay": 0,
|
12 |
+
"min_guidance_scale": 3,
|
13 |
+
"use_erg_tag": true,
|
14 |
+
"use_erg_lyric": true,
|
15 |
+
"use_erg_diffusion": true,
|
16 |
+
"oss_steps": [],
|
17 |
+
"timecosts": {
|
18 |
+
"preprocess": 0.0282442569732666,
|
19 |
+
"diffusion": 12.104875326156616,
|
20 |
+
"latent2audio": 1.587641954421997
|
21 |
+
},
|
22 |
+
"actual_seeds": [
|
23 |
+
4028738662
|
24 |
+
]
|
25 |
+
}
|
examples/zh_rap_lora/input_params/output_20250512120348_0_input_params.json
CHANGED
@@ -22,7 +22,7 @@
|
|
22 |
"latent2audio": 0.5694489479064941
|
23 |
},
|
24 |
"actual_seeds": [
|
25 |
-
|
26 |
],
|
27 |
"retake_seeds": [
|
28 |
1603201617
|
|
|
22 |
"latent2audio": 0.5694489479064941
|
23 |
},
|
24 |
"actual_seeds": [
|
25 |
+
226581098
|
26 |
],
|
27 |
"retake_seeds": [
|
28 |
1603201617
|
examples/zh_rap_lora/input_params/output_20250512160830_0_input_params.json
DELETED
@@ -1,45 +0,0 @@
|
|
1 |
-
{
|
2 |
-
"lora_name_or_path": "/root/sag_train/data/ace_step_v1_chinese_rap_lora_80k",
|
3 |
-
"task": "text2music",
|
4 |
-
"prompt": "articulate, spoken word, young adult, rap music, male, clear, energetic, warm, relaxed, breathy, night club",
|
5 |
-
"lyrics": "[verse]\n่ฟ ่ฟ ่ฐ ๅ ๅจ ๆดพ ๅฏน ๅ ๅค\nๆ ็ ่ ่ข\nๅ ่ขซ ้ฉด ่ธข ่ฟ\nไธ ๅฏน ๅฒ\n่ ๅคด ๆ ็ป ไธ ไผ ่ฏด\nไฝ ๆฅ ๆ ๆ ๆ ๅฐฑ ่ทช\nๅผ ๅฑ ็ด ๆฅ ๅดฉ ๆบ\n\n[chorus]\nๅฐฑ ๅช ไนฑ ๅช ๅฟต ๅช ้ ๅช\nๅด ๅช ็ข ๅช ๆ ๅช ็ ๅช\n่ ๅช ่ข ๅช ๅ ๅช ๆต ๅช ็ณ ๅช\n่ท ๅช ็ ๅช ่ ๅช ๅฅ ๅช\nๆ ๅช ๆญ ๅช ่ฏ ๅช ๅ
จ ๅช ๅฟ ๅช\nไธ ๅช ๅผ ๅช ๅด ๅช ๅฐฑ ๅช ๅบ ๅช\nๅช ๅช ๅฉ ๅช ไธ ๅช ๅฐด ๅช ๅฐฌ ๅช ๅ ๅช ๅฟ\n่๏ผ\n\n[verse]\n้ ้ ้ ้ ไบ\nไธ ๅฃ ๆฐ ๅ
จ ๅฟต ้\n้ ้ ้ ้ ไบ\n่ ๅคด ๆ ็ป ็ฉ ้
\n็ฉ ็ฉ ็ฉ ็ฉ ้
\n็ฉ ้
็ฉ ้
\nๆ ๅญ ๅ
จ ้จ ไนฑ ๅฅ\n่ง ไผ ็ฌ ๅฐ ๅ ่ก\n\n[verse]\nไฝ ็ ๆญ ่ฏ ๆ ็ ๅฉ ๆขฆ\nๅฑ ๅฎ ็ด ๆฅ ็คพ ๆญป\n่ฐ ่ท ๅฐ ๅค ๅคช ็ฉบ\n่ง ไผ ่กจ ๆ
่ฃ ๅผ\nไฝ ็ฌ ๆ ่\nๆ ็ฌ ไฝ ไธ ๆ\n่ฟ ๅซ ่บ ๆฏ ่กจ ๆผ\nไธ ๆ ไฝ ๆฅ๏ผ\n\n[verse]\n่ฟ ่ฟ ่ฐ ๅ ๅจ ๆดพ ๅฏน ไธข ไบบ\nๆ ็ ไธ ็\nๅทฒ ็ป ๅฝป ๅบ ๅดฉ ๆบ\nๆฒก ๆ ๅฎ ็พ\nๅช ๆ ็ฟป ่ฝฆ ็ฐ ๅบ\nไปฅ ๅ ่ง ไผ ็ ๅฒ ่ฎฝ\n\n[chorus]\nๅฐฑ ๅช ไนฑ ๅช ๅฟต ๅช ้ ๅช\nๅด ๅช ็ข ๅช ๆ ๅช ็ ๅช\n่ ๅช ่ข ๅช ๅ ๅช ๆต ๅช ็ณ ๅช\n่ท ๅช ็ ๅช ่ ๅช ๅฅ ๅช\nๆ ๅช ๆญ ๅช ่ฏ ๅช ๅ
จ ๅช ๅฟ ๅช\nไธ ๅช ๅผ ๅช ๅด ๅช ๅฐฑ ๅช ๅบ ๅช\nๅช ๅช ๅฉ ๅช ไธ ๅช ๅฐด ๅช ๅฐฌ ๅช ๅ ๅช ๅฟ\n่๏ผ\n\n[verse]\n้ ้ ้ ้ ไบ\nไธ ๅฃ ๆฐ ๅ
จ ๅฟต ้\n้ ้ ้ ้ ไบ\n่ ๅคด ๆ ็ป ็ฉ ้
\n็ฉ ็ฉ ็ฉ ็ฉ ้
\n็ฉ ้
็ฉ ้
\nๆ ๅญ ๅ
จ ้จ ไนฑ ๅฅ\n่ง ไผ ็ฌ ๅฐ ๅ ่ก\n\n[verse]\nไฝ ็ ๆญ ่ฏ ๆ ็ ๅฉ ๆขฆ\nๅฑ ๅฎ ็ด ๆฅ ็คพ ๆญป\n่ฐ ่ท ๅฐ ๅค ๅคช ็ฉบ\n่ง ไผ ่กจ ๆ
่ฃ ๅผ\nไฝ ็ฌ ๆ ่\nๆ ็ฌ ไฝ ไธ ๆ\n่ฟ ๅซ ่บ ๆฏ ่กจ ๆผ\nไธ ๆ ไฝ ๆฅ๏ผ",
|
6 |
-
"audio_duration": 169.12,
|
7 |
-
"infer_step": 60,
|
8 |
-
"guidance_scale": 15,
|
9 |
-
"scheduler_type": "euler",
|
10 |
-
"cfg_type": "apg",
|
11 |
-
"omega_scale": 10,
|
12 |
-
"guidance_interval": 0.5,
|
13 |
-
"guidance_interval_decay": 0,
|
14 |
-
"min_guidance_scale": 3,
|
15 |
-
"use_erg_tag": true,
|
16 |
-
"use_erg_lyric": false,
|
17 |
-
"use_erg_diffusion": true,
|
18 |
-
"oss_steps": [],
|
19 |
-
"timecosts": {
|
20 |
-
"preprocess": 0.041605472564697266,
|
21 |
-
"diffusion": 14.009192705154419,
|
22 |
-
"latent2audio": 1.55946946144104
|
23 |
-
},
|
24 |
-
"actual_seeds": [
|
25 |
-
547563805
|
26 |
-
],
|
27 |
-
"retake_seeds": [
|
28 |
-
2702917060
|
29 |
-
],
|
30 |
-
"retake_variance": 0.5,
|
31 |
-
"guidance_scale_text": 0,
|
32 |
-
"guidance_scale_lyric": 0,
|
33 |
-
"repaint_start": 0,
|
34 |
-
"repaint_end": 0,
|
35 |
-
"edit_n_min": 0.0,
|
36 |
-
"edit_n_max": 1.0,
|
37 |
-
"edit_n_avg": 1,
|
38 |
-
"src_audio_path": null,
|
39 |
-
"edit_target_prompt": null,
|
40 |
-
"edit_target_lyrics": null,
|
41 |
-
"audio2audio_enable": false,
|
42 |
-
"ref_audio_strength": 0.5,
|
43 |
-
"ref_audio_input": null,
|
44 |
-
"audio_path": "./outputs/output_20250512160830_0.wav"
|
45 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
pipeline_ace_step.py
CHANGED
@@ -12,9 +12,15 @@ import math
|
|
12 |
from huggingface_hub import hf_hub_download, snapshot_download
|
13 |
|
14 |
# from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
15 |
-
from schedulers.scheduling_flow_match_euler_discrete import
|
16 |
-
|
17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
from diffusers.utils.torch_utils import randn_tensor
|
19 |
from transformers import UMT5EncoderModel, AutoTokenizer
|
20 |
|
@@ -22,23 +28,42 @@ from language_segmentation import LangSegment
|
|
22 |
from music_dcae.music_dcae_pipeline import MusicDCAE
|
23 |
from models.ace_step_transformer import ACEStepTransformer2DModel
|
24 |
from models.lyrics_utils.lyric_tokenizer import VoiceBpeTokenizer
|
25 |
-
from apg_guidance import
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
import torchaudio
|
27 |
import torio
|
28 |
|
29 |
|
30 |
torch.backends.cudnn.benchmark = False
|
31 |
-
torch.set_float32_matmul_precision(
|
32 |
torch.backends.cudnn.deterministic = True
|
33 |
torch.backends.cuda.matmul.allow_tf32 = True
|
34 |
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
35 |
|
36 |
|
37 |
SUPPORT_LANGUAGES = {
|
38 |
-
"en": 259,
|
39 |
-
"
|
40 |
-
"
|
41 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
}
|
43 |
|
44 |
structure_pattern = re.compile(r"\[.*?\]")
|
@@ -56,7 +81,16 @@ REPO_ID = "ACE-Step/ACE-Step-v1-3.5B"
|
|
56 |
# class ACEStepPipeline(DiffusionPipeline):
|
57 |
class ACEStepPipeline:
|
58 |
|
59 |
-
def __init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
if not checkpoint_dir:
|
61 |
if persistent_storage_path is None:
|
62 |
checkpoint_dir = os.path.join(os.path.dirname(__file__), "checkpoints")
|
@@ -64,7 +98,11 @@ class ACEStepPipeline:
|
|
64 |
checkpoint_dir = os.path.join(persistent_storage_path, "checkpoints")
|
65 |
ensure_directory_exists(checkpoint_dir)
|
66 |
self.checkpoint_dir = checkpoint_dir
|
67 |
-
device =
|
|
|
|
|
|
|
|
|
68 |
if device.type == "cpu" and torch.backends.mps.is_available():
|
69 |
device = torch.device("mps")
|
70 |
self.dtype = torch.bfloat16 if dtype == "bfloat16" else torch.float32
|
@@ -74,17 +112,25 @@ class ACEStepPipeline:
|
|
74 |
self.loaded = False
|
75 |
self.torch_compile = torch_compile
|
76 |
self.lora_path = "none"
|
77 |
-
|
78 |
def load_lora(self, lora_name_or_path):
|
79 |
if lora_name_or_path != self.lora_path and lora_name_or_path != "none":
|
80 |
if not os.path.exists(lora_name_or_path):
|
81 |
-
lora_download_path = snapshot_download(
|
|
|
|
|
82 |
else:
|
83 |
lora_download_path = lora_name_or_path
|
84 |
if self.lora_path != "none":
|
85 |
self.ace_step_transformer.unload_lora()
|
86 |
-
self.ace_step_transformer.load_lora_adapter(
|
87 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
self.lora_path = lora_name_or_path
|
89 |
elif self.lora_path != "none" and lora_name_or_path == "none":
|
90 |
logger.info("No lora weights to load.")
|
@@ -99,55 +145,124 @@ class ACEStepPipeline:
|
|
99 |
text_encoder_model_path = os.path.join(checkpoint_dir, "umt5-base")
|
100 |
|
101 |
files_exist = (
|
102 |
-
os.path.exists(os.path.join(dcae_model_path, "config.json"))
|
103 |
-
os.path.exists(
|
104 |
-
|
105 |
-
|
106 |
-
os.path.exists(os.path.join(
|
107 |
-
os.path.exists(
|
108 |
-
|
109 |
-
|
110 |
-
os.path.exists(os.path.join(
|
111 |
-
os.path.exists(
|
112 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
)
|
114 |
|
115 |
if not files_exist:
|
116 |
-
logger.info(
|
|
|
|
|
117 |
|
118 |
# download music dcae model
|
119 |
os.makedirs(dcae_model_path, exist_ok=True)
|
120 |
-
hf_hub_download(
|
121 |
-
|
122 |
-
|
123 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
|
125 |
# download vocoder model
|
126 |
os.makedirs(vocoder_model_path, exist_ok=True)
|
127 |
-
hf_hub_download(
|
128 |
-
|
129 |
-
|
130 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
|
132 |
# download ace_step transformer model
|
133 |
os.makedirs(ace_step_model_path, exist_ok=True)
|
134 |
-
hf_hub_download(
|
135 |
-
|
136 |
-
|
137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
|
139 |
# download text encoder model
|
140 |
os.makedirs(text_encoder_model_path, exist_ok=True)
|
141 |
-
hf_hub_download(
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
|
152 |
logger.info("Models downloaded")
|
153 |
|
@@ -156,29 +271,131 @@ class ACEStepPipeline:
|
|
156 |
ace_step_checkpoint_path = ace_step_model_path
|
157 |
text_encoder_checkpoint_path = text_encoder_model_path
|
158 |
|
159 |
-
self.music_dcae = MusicDCAE(
|
|
|
|
|
|
|
160 |
self.music_dcae.to(device).eval().to(self.dtype)
|
161 |
|
162 |
-
self.ace_step_transformer = ACEStepTransformer2DModel.from_pretrained(
|
|
|
|
|
163 |
self.ace_step_transformer.to(device).eval().to(self.dtype)
|
164 |
|
165 |
lang_segment = LangSegment()
|
166 |
|
167 |
-
lang_segment.setfilters(
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
self.lang_segment = lang_segment
|
176 |
self.lyric_tokenizer = VoiceBpeTokenizer()
|
177 |
-
text_encoder_model = UMT5EncoderModel.from_pretrained(
|
|
|
|
|
178 |
text_encoder_model = text_encoder_model.to(device).to(self.dtype)
|
179 |
text_encoder_model.requires_grad_(False)
|
180 |
self.text_encoder_model = text_encoder_model
|
181 |
-
self.text_tokenizer = AutoTokenizer.from_pretrained(
|
|
|
|
|
182 |
self.loaded = True
|
183 |
|
184 |
# compile
|
@@ -188,7 +405,13 @@ class ACEStepPipeline:
|
|
188 |
self.text_encoder_model = torch.compile(self.text_encoder_model)
|
189 |
|
190 |
def get_text_embeddings(self, texts, device, text_max_length=256):
|
191 |
-
inputs = self.text_tokenizer(
|
|
|
|
|
|
|
|
|
|
|
|
|
192 |
inputs = {key: value.to(device) for key, value in inputs.items()}
|
193 |
if self.text_encoder_model.device != device:
|
194 |
self.text_encoder_model.to(device)
|
@@ -197,62 +420,87 @@ class ACEStepPipeline:
|
|
197 |
last_hidden_states = outputs.last_hidden_state
|
198 |
attention_mask = inputs["attention_mask"]
|
199 |
return last_hidden_states, attention_mask
|
200 |
-
|
201 |
-
def get_text_embeddings_null(
|
202 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
inputs = {key: value.to(device) for key, value in inputs.items()}
|
204 |
if self.text_encoder_model.device != device:
|
205 |
self.text_encoder_model.to(device)
|
206 |
-
|
207 |
def forward_with_temperature(inputs, tau=0.01, l_min=8, l_max=10):
|
208 |
handlers = []
|
209 |
-
|
210 |
def hook(module, input, output):
|
211 |
output[:] *= tau
|
212 |
return output
|
213 |
-
|
214 |
for i in range(l_min, l_max):
|
215 |
-
handler =
|
|
|
|
|
|
|
|
|
216 |
handlers.append(handler)
|
217 |
-
|
218 |
with torch.no_grad():
|
219 |
outputs = self.text_encoder_model(**inputs)
|
220 |
last_hidden_states = outputs.last_hidden_state
|
221 |
-
|
222 |
for hook in handlers:
|
223 |
hook.remove()
|
224 |
-
|
225 |
return last_hidden_states
|
226 |
-
|
227 |
last_hidden_states = forward_with_temperature(inputs, tau, l_min, l_max)
|
228 |
return last_hidden_states
|
229 |
|
230 |
def set_seeds(self, batch_size, manual_seeds=None):
|
231 |
-
|
232 |
if manual_seeds is not None:
|
233 |
if isinstance(manual_seeds, str):
|
234 |
if "," in manual_seeds:
|
235 |
-
|
236 |
elif manual_seeds.isdigit():
|
237 |
-
|
238 |
-
|
239 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
240 |
actual_seeds = []
|
241 |
for i in range(batch_size):
|
242 |
-
|
243 |
-
if
|
244 |
-
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
|
|
|
|
|
|
|
|
|
|
251 |
return random_generators, actual_seeds
|
252 |
|
253 |
def get_lang(self, text):
|
254 |
language = "en"
|
255 |
-
try:
|
256 |
_ = self.lang_segment.getTexts(text)
|
257 |
langCounts = self.lang_segment.getCounts()
|
258 |
language = langCounts[0][0]
|
@@ -286,7 +534,9 @@ class ACEStepPipeline:
|
|
286 |
else:
|
287 |
token_idx = self.lyric_tokenizer.encode(line, lang)
|
288 |
if debug:
|
289 |
-
toks = self.lyric_tokenizer.batch_decode(
|
|
|
|
|
290 |
logger.info(f"debbug {line} --> {lang} --> {toks}")
|
291 |
lyric_token_idx = lyric_token_idx + token_idx + [2]
|
292 |
except Exception as e:
|
@@ -315,11 +565,13 @@ class ACEStepPipeline:
|
|
315 |
attention_mask=None,
|
316 |
momentum_buffer=None,
|
317 |
momentum_buffer_tar=None,
|
318 |
-
return_src_pred=True
|
319 |
):
|
320 |
noise_pred_src = None
|
321 |
if return_src_pred:
|
322 |
-
src_latent_model_input =
|
|
|
|
|
323 |
timestep = t.expand(src_latent_model_input.shape[0])
|
324 |
# source
|
325 |
noise_pred_src = self.ace_step_transformer(
|
@@ -334,7 +586,9 @@ class ACEStepPipeline:
|
|
334 |
).sample
|
335 |
|
336 |
if do_classifier_free_guidance:
|
337 |
-
noise_pred_with_cond_src, noise_pred_uncond_src = noise_pred_src.chunk(
|
|
|
|
|
338 |
if cfg_type == "apg":
|
339 |
noise_pred_src = apg_forward(
|
340 |
pred_cond=noise_pred_with_cond_src,
|
@@ -349,7 +603,9 @@ class ACEStepPipeline:
|
|
349 |
cfg_strength=guidance_scale,
|
350 |
)
|
351 |
|
352 |
-
tar_latent_model_input =
|
|
|
|
|
353 |
timestep = t.expand(tar_latent_model_input.shape[0])
|
354 |
# target
|
355 |
noise_pred_tar = self.ace_step_transformer(
|
@@ -419,26 +675,52 @@ class ACEStepPipeline:
|
|
419 |
T_steps = infer_steps
|
420 |
frame_length = src_latents.shape[-1]
|
421 |
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
422 |
-
|
423 |
-
timesteps, T_steps = retrieve_timesteps(
|
|
|
|
|
424 |
|
425 |
if do_classifier_free_guidance:
|
426 |
attention_mask = torch.cat([attention_mask] * 2, dim=0)
|
427 |
-
|
428 |
-
encoder_text_hidden_states = torch.cat(
|
|
|
|
|
|
|
|
|
|
|
|
|
429 |
text_attention_mask = torch.cat([text_attention_mask] * 2, dim=0)
|
430 |
|
431 |
-
target_encoder_text_hidden_states = torch.cat(
|
432 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
433 |
|
434 |
-
speaker_embds = torch.cat(
|
435 |
-
|
|
|
|
|
|
|
|
|
436 |
|
437 |
-
lyric_token_ids = torch.cat(
|
|
|
|
|
438 |
lyric_mask = torch.cat([lyric_mask, torch.zeros_like(lyric_mask)], 0)
|
439 |
|
440 |
-
target_lyric_token_ids = torch.cat(
|
441 |
-
|
|
|
|
|
|
|
|
|
442 |
|
443 |
momentum_buffer = MomentumBuffer()
|
444 |
momentum_buffer_tar = MomentumBuffer()
|
@@ -455,10 +737,10 @@ class ACEStepPipeline:
|
|
455 |
if i < n_min:
|
456 |
continue
|
457 |
|
458 |
-
t_i = t/1000
|
459 |
|
460 |
-
if i+1 < len(timesteps):
|
461 |
-
t_im1 = (timesteps[i+1])/1000
|
462 |
else:
|
463 |
t_im1 = torch.zeros_like(t_i).to(t_i.device)
|
464 |
|
@@ -466,7 +748,12 @@ class ACEStepPipeline:
|
|
466 |
# Calculate the average of the V predictions
|
467 |
V_delta_avg = torch.zeros_like(x_src)
|
468 |
for k in range(n_avg):
|
469 |
-
fwd_noise = randn_tensor(
|
|
|
|
|
|
|
|
|
|
|
470 |
|
471 |
zt_src = (1 - t_i) * x_src + (t_i) * fwd_noise
|
472 |
|
@@ -490,22 +777,29 @@ class ACEStepPipeline:
|
|
490 |
guidance_scale=guidance_scale,
|
491 |
target_guidance_scale=target_guidance_scale,
|
492 |
attention_mask=attention_mask,
|
493 |
-
momentum_buffer=momentum_buffer
|
494 |
)
|
495 |
-
V_delta_avg += (1 / n_avg) * (
|
|
|
|
|
496 |
|
497 |
# propagate direct ODE
|
498 |
zt_edit = zt_edit.to(torch.float32)
|
499 |
zt_edit = zt_edit + (t_im1 - t_i) * V_delta_avg
|
500 |
zt_edit = zt_edit.to(V_delta_avg.dtype)
|
501 |
-
else:
|
502 |
if i == n_max:
|
503 |
-
fwd_noise = randn_tensor(
|
|
|
|
|
|
|
|
|
|
|
504 |
scheduler._init_step_index(t)
|
505 |
sigma = scheduler.sigmas[scheduler.step_index]
|
506 |
xt_src = sigma * fwd_noise + (1.0 - sigma) * x_src
|
507 |
xt_tar = zt_edit + xt_src - x_src
|
508 |
-
|
509 |
_, Vt_tar = self.calc_v(
|
510 |
zt_src=None,
|
511 |
zt_tar=xt_tar,
|
@@ -527,13 +821,13 @@ class ACEStepPipeline:
|
|
527 |
momentum_buffer_tar=momentum_buffer_tar,
|
528 |
return_src_pred=False,
|
529 |
)
|
530 |
-
|
531 |
dtype = Vt_tar.dtype
|
532 |
xt_tar = xt_tar.to(torch.float32)
|
533 |
prev_sample = xt_tar + (t_im1 - t_i) * Vt_tar
|
534 |
-
prev_sample = prev_sample.to(dtype)
|
535 |
xt_tar = prev_sample
|
536 |
-
|
537 |
target_latents = zt_edit if xt_tar is None else xt_tar
|
538 |
return target_latents
|
539 |
|
@@ -551,7 +845,12 @@ class ACEStepPipeline:
|
|
551 |
timesteps = scheduler.timesteps.unsqueeze(1).to(gt_latents.dtype)
|
552 |
indices = indices.to(timesteps.device).to(gt_latents.dtype).unsqueeze(1)
|
553 |
nearest_idx = torch.argmin(torch.cdist(indices, timesteps), dim=1)
|
554 |
-
sigma =
|
|
|
|
|
|
|
|
|
|
|
555 |
while len(sigma.shape) < gt_latents.ndim:
|
556 |
sigma = sigma.unsqueeze(-1)
|
557 |
noisy_image = sigma * noise + (1.0 - sigma) * gt_latents
|
@@ -595,15 +894,30 @@ class ACEStepPipeline:
|
|
595 |
ref_latents=None,
|
596 |
):
|
597 |
|
598 |
-
logger.info(
|
|
|
|
|
|
|
|
|
599 |
do_classifier_free_guidance = True
|
600 |
if guidance_scale == 0.0 or guidance_scale == 1.0:
|
601 |
do_classifier_free_guidance = False
|
602 |
-
|
603 |
do_double_condition_guidance = False
|
604 |
-
if
|
|
|
|
|
|
|
|
|
|
|
605 |
do_double_condition_guidance = True
|
606 |
-
logger.info(
|
|
|
|
|
|
|
|
|
|
|
|
|
607 |
|
608 |
device = encoder_text_hidden_states.device
|
609 |
dtype = encoder_text_hidden_states.dtype
|
@@ -619,7 +933,7 @@ class ACEStepPipeline:
|
|
619 |
num_train_timesteps=1000,
|
620 |
shift=3.0,
|
621 |
)
|
622 |
-
|
623 |
frame_length = int(duration * 44100 / 512 / 8)
|
624 |
if src_latents is not None:
|
625 |
frame_length = src_latents.shape[-1]
|
@@ -630,31 +944,60 @@ class ACEStepPipeline:
|
|
630 |
if len(oss_steps) > 0:
|
631 |
infer_steps = max(oss_steps)
|
632 |
scheduler.set_timesteps
|
633 |
-
timesteps, num_inference_steps = retrieve_timesteps(
|
|
|
|
|
|
|
|
|
|
|
634 |
new_timesteps = torch.zeros(len(oss_steps), dtype=dtype, device=device)
|
635 |
for idx in range(len(oss_steps)):
|
636 |
-
new_timesteps[idx] = timesteps[oss_steps[idx]-1]
|
637 |
num_inference_steps = len(oss_steps)
|
638 |
sigmas = (new_timesteps / 1000).float().cpu().numpy()
|
639 |
-
timesteps, num_inference_steps = retrieve_timesteps(
|
640 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
641 |
else:
|
642 |
-
timesteps, num_inference_steps = retrieve_timesteps(
|
643 |
-
|
644 |
-
|
645 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
646 |
is_repaint = False
|
647 |
-
is_extend
|
648 |
if add_retake_noise:
|
649 |
n_min = int(infer_steps * (1 - retake_variance))
|
650 |
-
retake_variance =
|
651 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
652 |
repaint_start_frame = int(repaint_start * 44100 / 512 / 8)
|
653 |
repaint_end_frame = int(repaint_end * 44100 / 512 / 8)
|
654 |
x0 = src_latents
|
655 |
# retake
|
656 |
-
is_repaint =
|
657 |
-
|
658 |
is_extend = (repaint_start_frame < 0) or (repaint_end_frame > frame_length)
|
659 |
if is_extend:
|
660 |
is_repaint = True
|
@@ -662,13 +1005,23 @@ class ACEStepPipeline:
|
|
662 |
# TODO: train a mask aware repainting controlnet
|
663 |
# to make sure mean = 0, std = 1
|
664 |
if not is_repaint:
|
665 |
-
target_latents =
|
|
|
|
|
|
|
666 |
elif not is_extend:
|
667 |
-
# if repaint_end_frame
|
668 |
-
repaint_mask = torch.zeros(
|
|
|
|
|
669 |
repaint_mask[:, :, :, repaint_start_frame:repaint_end_frame] = 1.0
|
670 |
-
repaint_noise =
|
671 |
-
|
|
|
|
|
|
|
|
|
|
|
672 |
zt_edit = x0.clone()
|
673 |
z0 = repaint_noise
|
674 |
elif is_extend:
|
@@ -684,73 +1037,107 @@ class ACEStepPipeline:
|
|
684 |
if repaint_start_frame < 0:
|
685 |
left_pad_frame_length = abs(repaint_start_frame)
|
686 |
frame_length = left_pad_frame_length + gt_latents.shape[-1]
|
687 |
-
extend_gt_latents = torch.nn.functional.pad(
|
|
|
|
|
688 |
if frame_length > max_infer_fame_length:
|
689 |
right_trim_length = frame_length - max_infer_fame_length
|
690 |
-
extend_gt_latents = extend_gt_latents[
|
691 |
-
|
|
|
|
|
|
|
|
|
692 |
frame_length = max_infer_fame_length
|
693 |
repaint_start_frame = 0
|
694 |
gt_latents = extend_gt_latents
|
695 |
-
|
696 |
if repaint_end_frame > src_latents_length:
|
697 |
right_pad_frame_length = repaint_end_frame - gt_latents.shape[-1]
|
698 |
frame_length = gt_latents.shape[-1] + right_pad_frame_length
|
699 |
-
extend_gt_latents = torch.nn.functional.pad(
|
|
|
|
|
700 |
if frame_length > max_infer_fame_length:
|
701 |
left_trim_length = frame_length - max_infer_fame_length
|
702 |
-
extend_gt_latents = extend_gt_latents[
|
703 |
-
|
|
|
|
|
|
|
|
|
704 |
frame_length = max_infer_fame_length
|
705 |
repaint_end_frame = frame_length
|
706 |
gt_latents = extend_gt_latents
|
707 |
|
708 |
-
repaint_mask = torch.zeros(
|
|
|
|
|
709 |
if left_pad_frame_length > 0:
|
710 |
-
repaint_mask[
|
711 |
if right_pad_frame_length > 0:
|
712 |
-
repaint_mask[
|
713 |
x0 = gt_latents
|
714 |
padd_list = []
|
715 |
if left_pad_frame_length > 0:
|
716 |
padd_list.append(retake_latents[:, :, :, :left_pad_frame_length])
|
717 |
-
padd_list.append(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
718 |
if right_pad_frame_length > 0:
|
719 |
padd_list.append(retake_latents[:, :, :, -right_pad_frame_length:])
|
720 |
target_latents = torch.cat(padd_list, dim=-1)
|
721 |
-
assert
|
|
|
|
|
722 |
zt_edit = x0.clone()
|
723 |
z0 = target_latents
|
724 |
|
725 |
init_timestep = 1000
|
726 |
if audio2audio_enable and ref_latents is not None:
|
727 |
-
target_latents, init_timestep = self.add_latents_noise(
|
|
|
|
|
|
|
|
|
|
|
728 |
|
729 |
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
730 |
-
|
731 |
# guidance interval
|
732 |
start_idx = int(num_inference_steps * ((1 - guidance_interval) / 2))
|
733 |
end_idx = int(num_inference_steps * (guidance_interval / 2 + 0.5))
|
734 |
-
logger.info(
|
|
|
|
|
735 |
|
736 |
momentum_buffer = MomentumBuffer()
|
737 |
|
738 |
def forward_encoder_with_temperature(self, inputs, tau=0.01, l_min=4, l_max=6):
|
739 |
handlers = []
|
740 |
-
|
741 |
def hook(module, input, output):
|
742 |
output[:] *= tau
|
743 |
return output
|
744 |
-
|
745 |
for i in range(l_min, l_max):
|
746 |
-
handler = self.ace_step_transformer.lyric_encoder.encoders[
|
|
|
|
|
747 |
handlers.append(handler)
|
748 |
-
|
749 |
-
encoder_hidden_states, encoder_hidden_mask =
|
750 |
-
|
|
|
|
|
751 |
for hook in handlers:
|
752 |
hook.remove()
|
753 |
-
|
754 |
return encoder_hidden_states
|
755 |
|
756 |
# P(speaker, text, lyric)
|
@@ -767,12 +1154,16 @@ class ACEStepPipeline:
|
|
767 |
encoder_hidden_states_null = forward_encoder_with_temperature(
|
768 |
self,
|
769 |
inputs={
|
770 |
-
"encoder_text_hidden_states":
|
|
|
|
|
|
|
|
|
771 |
"text_attention_mask": text_attention_mask,
|
772 |
"speaker_embeds": torch.zeros_like(speaker_embds),
|
773 |
"lyric_token_idx": lyric_token_ids,
|
774 |
"lyric_mask": lyric_mask,
|
775 |
-
}
|
776 |
)
|
777 |
else:
|
778 |
# P(null_speaker, null_text, null_lyric)
|
@@ -783,7 +1174,7 @@ class ACEStepPipeline:
|
|
783 |
torch.zeros_like(lyric_token_ids),
|
784 |
lyric_mask,
|
785 |
)
|
786 |
-
|
787 |
encoder_hidden_states_no_lyric = None
|
788 |
if do_double_condition_guidance:
|
789 |
# P(null_speaker, text, lyric_weaker)
|
@@ -796,7 +1187,7 @@ class ACEStepPipeline:
|
|
796 |
"speaker_embeds": torch.zeros_like(speaker_embds),
|
797 |
"lyric_token_idx": lyric_token_ids,
|
798 |
"lyric_mask": lyric_mask,
|
799 |
-
}
|
800 |
)
|
801 |
# P(null_speaker, text, no_lyric)
|
802 |
else:
|
@@ -808,26 +1199,34 @@ class ACEStepPipeline:
|
|
808 |
lyric_mask,
|
809 |
)
|
810 |
|
811 |
-
def forward_diffusion_with_temperature(
|
|
|
|
|
812 |
handlers = []
|
813 |
-
|
814 |
def hook(module, input, output):
|
815 |
output[:] *= tau
|
816 |
return output
|
817 |
-
|
818 |
for i in range(l_min, l_max):
|
819 |
-
handler = self.ace_step_transformer.transformer_blocks[
|
|
|
|
|
820 |
handlers.append(handler)
|
821 |
-
handler = self.ace_step_transformer.transformer_blocks[
|
|
|
|
|
822 |
handlers.append(handler)
|
823 |
|
824 |
-
sample = self.ace_step_transformer.decode(
|
825 |
-
|
|
|
|
|
826 |
for hook in handlers:
|
827 |
hook.remove()
|
828 |
-
|
829 |
return sample
|
830 |
-
|
831 |
for i, t in tqdm(enumerate(timesteps), total=num_inference_steps):
|
832 |
|
833 |
if t > init_timestep:
|
@@ -850,8 +1249,15 @@ class ACEStepPipeline:
|
|
850 |
# compute current guidance scale
|
851 |
if guidance_interval_decay > 0:
|
852 |
# Linearly interpolate to calculate the current guidance scale
|
853 |
-
progress = (i - start_idx) / (
|
854 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
855 |
else:
|
856 |
current_guidance_scale = guidance_scale
|
857 |
|
@@ -869,7 +1275,10 @@ class ACEStepPipeline:
|
|
869 |
).sample
|
870 |
|
871 |
noise_pred_with_only_text_cond = None
|
872 |
-
if
|
|
|
|
|
|
|
873 |
noise_pred_with_only_text_cond = self.ace_step_transformer.decode(
|
874 |
hidden_states=latent_model_input,
|
875 |
attention_mask=attention_mask,
|
@@ -901,7 +1310,10 @@ class ACEStepPipeline:
|
|
901 |
timestep=timestep,
|
902 |
).sample
|
903 |
|
904 |
-
if
|
|
|
|
|
|
|
905 |
noise_pred = cfg_double_condition_forward(
|
906 |
cond_output=noise_pred_with_cond,
|
907 |
uncond_output=noise_pred_uncond,
|
@@ -930,7 +1342,7 @@ class ACEStepPipeline:
|
|
930 |
guidance_scale=current_guidance_scale,
|
931 |
i=i,
|
932 |
zero_steps=zero_steps,
|
933 |
-
use_zero_init=use_zero_init
|
934 |
)
|
935 |
else:
|
936 |
latent_model_input = latents
|
@@ -945,9 +1357,9 @@ class ACEStepPipeline:
|
|
945 |
).sample
|
946 |
|
947 |
if is_repaint and i >= n_min:
|
948 |
-
t_i = t/1000
|
949 |
-
if i+1 < len(timesteps):
|
950 |
-
t_im1 = (timesteps[i+1])/1000
|
951 |
else:
|
952 |
t_im1 = torch.zeros_like(t_i).to(t_i.device)
|
953 |
dtype = noise_pred.dtype
|
@@ -956,18 +1368,37 @@ class ACEStepPipeline:
|
|
956 |
prev_sample = prev_sample.to(dtype)
|
957 |
target_latents = prev_sample
|
958 |
zt_src = (1 - t_im1) * x0 + (t_im1) * z0
|
959 |
-
target_latents = torch.where(
|
|
|
|
|
960 |
else:
|
961 |
-
target_latents = scheduler.step(
|
|
|
|
|
|
|
|
|
|
|
|
|
962 |
|
963 |
if is_extend:
|
964 |
if to_right_pad_gt_latents is not None:
|
965 |
-
target_latents = torch.cat(
|
|
|
|
|
966 |
if to_left_pad_gt_latents is not None:
|
967 |
-
target_latents = torch.cat(
|
|
|
|
|
968 |
return target_latents
|
969 |
|
970 |
-
def latents2audio(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
971 |
output_audio_paths = []
|
972 |
bs = latents.shape[0]
|
973 |
audio_lengths = [target_wav_duration_second * sample_rate] * bs
|
@@ -976,11 +1407,15 @@ class ACEStepPipeline:
|
|
976 |
_, pred_wavs = self.music_dcae.decode(pred_latents, sr=sample_rate)
|
977 |
pred_wavs = [pred_wav.cpu().float() for pred_wav in pred_wavs]
|
978 |
for i in tqdm(range(bs)):
|
979 |
-
output_audio_path = self.save_wav_file(
|
|
|
|
|
980 |
output_audio_paths.append(output_audio_path)
|
981 |
return output_audio_paths
|
982 |
|
983 |
-
def save_wav_file(
|
|
|
|
|
984 |
if save_path is None:
|
985 |
logger.warning("save_path is None, using default path ./outputs/")
|
986 |
base_path = f"./outputs"
|
@@ -989,9 +1424,17 @@ class ACEStepPipeline:
|
|
989 |
base_path = save_path
|
990 |
ensure_directory_exists(base_path)
|
991 |
|
992 |
-
output_path_flac =
|
|
|
|
|
993 |
target_wav = target_wav.float()
|
994 |
-
torchaudio.save(
|
|
|
|
|
|
|
|
|
|
|
|
|
995 |
return output_path_flac
|
996 |
|
997 |
def infer_latents(self, input_audio_path):
|
@@ -1017,7 +1460,7 @@ class ACEStepPipeline:
|
|
1017 |
omega_scale: int = 10.0,
|
1018 |
manual_seeds: list = None,
|
1019 |
guidance_interval: float = 0.5,
|
1020 |
-
guidance_interval_decay: float = 0
|
1021 |
min_guidance_scale: float = 3.0,
|
1022 |
use_erg_tag: bool = True,
|
1023 |
use_erg_lyric: bool = True,
|
@@ -1060,22 +1503,30 @@ class ACEStepPipeline:
|
|
1060 |
start_time = time.time()
|
1061 |
|
1062 |
random_generators, actual_seeds = self.set_seeds(batch_size, manual_seeds)
|
1063 |
-
retake_random_generators, actual_retake_seeds = self.set_seeds(
|
|
|
|
|
1064 |
|
1065 |
if isinstance(oss_steps, str) and len(oss_steps) > 0:
|
1066 |
oss_steps = list(map(int, oss_steps.split(",")))
|
1067 |
else:
|
1068 |
oss_steps = []
|
1069 |
-
|
1070 |
texts = [prompt]
|
1071 |
-
encoder_text_hidden_states, text_attention_mask = self.get_text_embeddings(
|
|
|
|
|
1072 |
encoder_text_hidden_states = encoder_text_hidden_states.repeat(batch_size, 1, 1)
|
1073 |
text_attention_mask = text_attention_mask.repeat(batch_size, 1)
|
1074 |
|
1075 |
encoder_text_hidden_states_null = None
|
1076 |
if use_erg_tag:
|
1077 |
-
encoder_text_hidden_states_null = self.get_text_embeddings_null(
|
1078 |
-
|
|
|
|
|
|
|
|
|
1079 |
|
1080 |
# not support for released checkpoint
|
1081 |
speaker_embeds = torch.zeros(batch_size, 512).to(self.device).to(self.dtype)
|
@@ -1086,8 +1537,18 @@ class ACEStepPipeline:
|
|
1086 |
if len(lyrics) > 0:
|
1087 |
lyric_token_idx = self.tokenize_lyrics(lyrics, debug=debug)
|
1088 |
lyric_mask = [1] * len(lyric_token_idx)
|
1089 |
-
lyric_token_idx =
|
1090 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1091 |
|
1092 |
if audio_duration <= 0:
|
1093 |
audio_duration = random.uniform(30.0, 240.0)
|
@@ -1102,16 +1563,24 @@ class ACEStepPipeline:
|
|
1102 |
if task == "retake":
|
1103 |
repaint_start = 0
|
1104 |
repaint_end = audio_duration
|
1105 |
-
|
1106 |
src_latents = None
|
1107 |
if src_audio_path is not None:
|
1108 |
-
assert src_audio_path is not None and task in (
|
1109 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1110 |
src_latents = self.infer_latents(src_audio_path)
|
1111 |
|
1112 |
ref_latents = None
|
1113 |
if ref_audio_input is not None and audio2audio_enable:
|
1114 |
-
assert
|
|
|
|
|
1115 |
assert os.path.exists(
|
1116 |
ref_audio_input
|
1117 |
), f"ref_audio_input {ref_audio_input} does not exist"
|
@@ -1119,17 +1588,39 @@ class ACEStepPipeline:
|
|
1119 |
|
1120 |
if task == "edit":
|
1121 |
texts = [edit_target_prompt]
|
1122 |
-
target_encoder_text_hidden_states, target_text_attention_mask =
|
1123 |
-
|
1124 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1125 |
|
1126 |
-
target_lyric_token_idx =
|
1127 |
-
|
|
|
|
|
|
|
|
|
1128 |
if len(edit_target_lyrics) > 0:
|
1129 |
-
target_lyric_token_idx = self.tokenize_lyrics(
|
|
|
|
|
1130 |
target_lyric_mask = [1] * len(target_lyric_token_idx)
|
1131 |
-
target_lyric_token_idx =
|
1132 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1133 |
|
1134 |
target_speaker_embeds = speaker_embeds.clone()
|
1135 |
|
@@ -1145,7 +1636,7 @@ class ACEStepPipeline:
|
|
1145 |
target_lyric_token_ids=target_lyric_token_idx,
|
1146 |
target_lyric_mask=target_lyric_mask,
|
1147 |
src_latents=src_latents,
|
1148 |
-
random_generators=retake_random_generators,
|
1149 |
infer_steps=infer_step,
|
1150 |
guidance_scale=guidance_scale,
|
1151 |
n_min=edit_n_min,
|
@@ -1233,7 +1724,7 @@ class ACEStepPipeline:
|
|
1233 |
"repaint_end": repaint_end,
|
1234 |
"edit_n_min": edit_n_min,
|
1235 |
"edit_n_max": edit_n_max,
|
1236 |
-
"edit_n_avg": edit_n_avg,
|
1237 |
"src_audio_path": src_audio_path,
|
1238 |
"edit_target_prompt": edit_target_prompt,
|
1239 |
"edit_target_lyrics": edit_target_lyrics,
|
@@ -1243,7 +1734,9 @@ class ACEStepPipeline:
|
|
1243 |
}
|
1244 |
# save input_params_json
|
1245 |
for output_audio_path in output_paths:
|
1246 |
-
input_params_json_save_path = output_audio_path.replace(
|
|
|
|
|
1247 |
input_params_json["audio_path"] = output_audio_path
|
1248 |
with open(input_params_json_save_path, "w", encoding="utf-8") as f:
|
1249 |
json.dump(input_params_json, f, indent=4, ensure_ascii=False)
|
|
|
12 |
from huggingface_hub import hf_hub_download, snapshot_download
|
13 |
|
14 |
# from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
15 |
+
from schedulers.scheduling_flow_match_euler_discrete import (
|
16 |
+
FlowMatchEulerDiscreteScheduler,
|
17 |
+
)
|
18 |
+
from schedulers.scheduling_flow_match_heun_discrete import (
|
19 |
+
FlowMatchHeunDiscreteScheduler,
|
20 |
+
)
|
21 |
+
from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3 import (
|
22 |
+
retrieve_timesteps,
|
23 |
+
)
|
24 |
from diffusers.utils.torch_utils import randn_tensor
|
25 |
from transformers import UMT5EncoderModel, AutoTokenizer
|
26 |
|
|
|
28 |
from music_dcae.music_dcae_pipeline import MusicDCAE
|
29 |
from models.ace_step_transformer import ACEStepTransformer2DModel
|
30 |
from models.lyrics_utils.lyric_tokenizer import VoiceBpeTokenizer
|
31 |
+
from apg_guidance import (
|
32 |
+
apg_forward,
|
33 |
+
MomentumBuffer,
|
34 |
+
cfg_forward,
|
35 |
+
cfg_zero_star,
|
36 |
+
cfg_double_condition_forward,
|
37 |
+
)
|
38 |
import torchaudio
|
39 |
import torio
|
40 |
|
41 |
|
42 |
torch.backends.cudnn.benchmark = False
|
43 |
+
torch.set_float32_matmul_precision("high")
|
44 |
torch.backends.cudnn.deterministic = True
|
45 |
torch.backends.cuda.matmul.allow_tf32 = True
|
46 |
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
47 |
|
48 |
|
49 |
SUPPORT_LANGUAGES = {
|
50 |
+
"en": 259,
|
51 |
+
"de": 260,
|
52 |
+
"fr": 262,
|
53 |
+
"es": 284,
|
54 |
+
"it": 285,
|
55 |
+
"pt": 286,
|
56 |
+
"pl": 294,
|
57 |
+
"tr": 295,
|
58 |
+
"ru": 267,
|
59 |
+
"cs": 293,
|
60 |
+
"nl": 297,
|
61 |
+
"ar": 5022,
|
62 |
+
"zh": 5023,
|
63 |
+
"ja": 5412,
|
64 |
+
"hu": 5753,
|
65 |
+
"ko": 6152,
|
66 |
+
"hi": 6680,
|
67 |
}
|
68 |
|
69 |
structure_pattern = re.compile(r"\[.*?\]")
|
|
|
81 |
# class ACEStepPipeline(DiffusionPipeline):
|
82 |
class ACEStepPipeline:
|
83 |
|
84 |
+
def __init__(
|
85 |
+
self,
|
86 |
+
checkpoint_dir=None,
|
87 |
+
device_id=0,
|
88 |
+
dtype="bfloat16",
|
89 |
+
text_encoder_checkpoint_path=None,
|
90 |
+
persistent_storage_path=None,
|
91 |
+
torch_compile=False,
|
92 |
+
**kwargs,
|
93 |
+
):
|
94 |
if not checkpoint_dir:
|
95 |
if persistent_storage_path is None:
|
96 |
checkpoint_dir = os.path.join(os.path.dirname(__file__), "checkpoints")
|
|
|
98 |
checkpoint_dir = os.path.join(persistent_storage_path, "checkpoints")
|
99 |
ensure_directory_exists(checkpoint_dir)
|
100 |
self.checkpoint_dir = checkpoint_dir
|
101 |
+
device = (
|
102 |
+
torch.device(f"cuda:{device_id}")
|
103 |
+
if torch.cuda.is_available()
|
104 |
+
else torch.device("cpu")
|
105 |
+
)
|
106 |
if device.type == "cpu" and torch.backends.mps.is_available():
|
107 |
device = torch.device("mps")
|
108 |
self.dtype = torch.bfloat16 if dtype == "bfloat16" else torch.float32
|
|
|
112 |
self.loaded = False
|
113 |
self.torch_compile = torch_compile
|
114 |
self.lora_path = "none"
|
115 |
+
|
116 |
def load_lora(self, lora_name_or_path):
|
117 |
if lora_name_or_path != self.lora_path and lora_name_or_path != "none":
|
118 |
if not os.path.exists(lora_name_or_path):
|
119 |
+
lora_download_path = snapshot_download(
|
120 |
+
lora_name_or_path, cache_dir=self.checkpoint_dir
|
121 |
+
)
|
122 |
else:
|
123 |
lora_download_path = lora_name_or_path
|
124 |
if self.lora_path != "none":
|
125 |
self.ace_step_transformer.unload_lora()
|
126 |
+
self.ace_step_transformer.load_lora_adapter(
|
127 |
+
os.path.join(lora_download_path, "pytorch_lora_weights.safetensors"),
|
128 |
+
adapter_name="zh_rap_lora",
|
129 |
+
with_alpha=True,
|
130 |
+
)
|
131 |
+
logger.info(
|
132 |
+
f"Loading lora weights from: {lora_name_or_path} download path is: {lora_download_path}"
|
133 |
+
)
|
134 |
self.lora_path = lora_name_or_path
|
135 |
elif self.lora_path != "none" and lora_name_or_path == "none":
|
136 |
logger.info("No lora weights to load.")
|
|
|
145 |
text_encoder_model_path = os.path.join(checkpoint_dir, "umt5-base")
|
146 |
|
147 |
files_exist = (
|
148 |
+
os.path.exists(os.path.join(dcae_model_path, "config.json"))
|
149 |
+
and os.path.exists(
|
150 |
+
os.path.join(dcae_model_path, "diffusion_pytorch_model.safetensors")
|
151 |
+
)
|
152 |
+
and os.path.exists(os.path.join(vocoder_model_path, "config.json"))
|
153 |
+
and os.path.exists(
|
154 |
+
os.path.join(vocoder_model_path, "diffusion_pytorch_model.safetensors")
|
155 |
+
)
|
156 |
+
and os.path.exists(os.path.join(ace_step_model_path, "config.json"))
|
157 |
+
and os.path.exists(
|
158 |
+
os.path.join(ace_step_model_path, "diffusion_pytorch_model.safetensors")
|
159 |
+
)
|
160 |
+
and os.path.exists(os.path.join(text_encoder_model_path, "config.json"))
|
161 |
+
and os.path.exists(
|
162 |
+
os.path.join(text_encoder_model_path, "model.safetensors")
|
163 |
+
)
|
164 |
+
and os.path.exists(
|
165 |
+
os.path.join(text_encoder_model_path, "special_tokens_map.json")
|
166 |
+
)
|
167 |
+
and os.path.exists(
|
168 |
+
os.path.join(text_encoder_model_path, "tokenizer_config.json")
|
169 |
+
)
|
170 |
+
and os.path.exists(os.path.join(text_encoder_model_path, "tokenizer.json"))
|
171 |
)
|
172 |
|
173 |
if not files_exist:
|
174 |
+
logger.info(
|
175 |
+
f"Checkpoint directory {checkpoint_dir} is not complete, downloading from Hugging Face Hub"
|
176 |
+
)
|
177 |
|
178 |
# download music dcae model
|
179 |
os.makedirs(dcae_model_path, exist_ok=True)
|
180 |
+
hf_hub_download(
|
181 |
+
repo_id=REPO_ID,
|
182 |
+
subfolder="music_dcae_f8c8",
|
183 |
+
filename="config.json",
|
184 |
+
local_dir=checkpoint_dir,
|
185 |
+
local_dir_use_symlinks=False,
|
186 |
+
)
|
187 |
+
hf_hub_download(
|
188 |
+
repo_id=REPO_ID,
|
189 |
+
subfolder="music_dcae_f8c8",
|
190 |
+
filename="diffusion_pytorch_model.safetensors",
|
191 |
+
local_dir=checkpoint_dir,
|
192 |
+
local_dir_use_symlinks=False,
|
193 |
+
)
|
194 |
|
195 |
# download vocoder model
|
196 |
os.makedirs(vocoder_model_path, exist_ok=True)
|
197 |
+
hf_hub_download(
|
198 |
+
repo_id=REPO_ID,
|
199 |
+
subfolder="music_vocoder",
|
200 |
+
filename="config.json",
|
201 |
+
local_dir=checkpoint_dir,
|
202 |
+
local_dir_use_symlinks=False,
|
203 |
+
)
|
204 |
+
hf_hub_download(
|
205 |
+
repo_id=REPO_ID,
|
206 |
+
subfolder="music_vocoder",
|
207 |
+
filename="diffusion_pytorch_model.safetensors",
|
208 |
+
local_dir=checkpoint_dir,
|
209 |
+
local_dir_use_symlinks=False,
|
210 |
+
)
|
211 |
|
212 |
# download ace_step transformer model
|
213 |
os.makedirs(ace_step_model_path, exist_ok=True)
|
214 |
+
hf_hub_download(
|
215 |
+
repo_id=REPO_ID,
|
216 |
+
subfolder="ace_step_transformer",
|
217 |
+
filename="config.json",
|
218 |
+
local_dir=checkpoint_dir,
|
219 |
+
local_dir_use_symlinks=False,
|
220 |
+
)
|
221 |
+
hf_hub_download(
|
222 |
+
repo_id=REPO_ID,
|
223 |
+
subfolder="ace_step_transformer",
|
224 |
+
filename="diffusion_pytorch_model.safetensors",
|
225 |
+
local_dir=checkpoint_dir,
|
226 |
+
local_dir_use_symlinks=False,
|
227 |
+
)
|
228 |
|
229 |
# download text encoder model
|
230 |
os.makedirs(text_encoder_model_path, exist_ok=True)
|
231 |
+
hf_hub_download(
|
232 |
+
repo_id=REPO_ID,
|
233 |
+
subfolder="umt5-base",
|
234 |
+
filename="config.json",
|
235 |
+
local_dir=checkpoint_dir,
|
236 |
+
local_dir_use_symlinks=False,
|
237 |
+
)
|
238 |
+
hf_hub_download(
|
239 |
+
repo_id=REPO_ID,
|
240 |
+
subfolder="umt5-base",
|
241 |
+
filename="model.safetensors",
|
242 |
+
local_dir=checkpoint_dir,
|
243 |
+
local_dir_use_symlinks=False,
|
244 |
+
)
|
245 |
+
hf_hub_download(
|
246 |
+
repo_id=REPO_ID,
|
247 |
+
subfolder="umt5-base",
|
248 |
+
filename="special_tokens_map.json",
|
249 |
+
local_dir=checkpoint_dir,
|
250 |
+
local_dir_use_symlinks=False,
|
251 |
+
)
|
252 |
+
hf_hub_download(
|
253 |
+
repo_id=REPO_ID,
|
254 |
+
subfolder="umt5-base",
|
255 |
+
filename="tokenizer_config.json",
|
256 |
+
local_dir=checkpoint_dir,
|
257 |
+
local_dir_use_symlinks=False,
|
258 |
+
)
|
259 |
+
hf_hub_download(
|
260 |
+
repo_id=REPO_ID,
|
261 |
+
subfolder="umt5-base",
|
262 |
+
filename="tokenizer.json",
|
263 |
+
local_dir=checkpoint_dir,
|
264 |
+
local_dir_use_symlinks=False,
|
265 |
+
)
|
266 |
|
267 |
logger.info("Models downloaded")
|
268 |
|
|
|
271 |
ace_step_checkpoint_path = ace_step_model_path
|
272 |
text_encoder_checkpoint_path = text_encoder_model_path
|
273 |
|
274 |
+
self.music_dcae = MusicDCAE(
|
275 |
+
dcae_checkpoint_path=dcae_checkpoint_path,
|
276 |
+
vocoder_checkpoint_path=vocoder_checkpoint_path,
|
277 |
+
)
|
278 |
self.music_dcae.to(device).eval().to(self.dtype)
|
279 |
|
280 |
+
self.ace_step_transformer = ACEStepTransformer2DModel.from_pretrained(
|
281 |
+
ace_step_checkpoint_path, torch_dtype=self.dtype
|
282 |
+
)
|
283 |
self.ace_step_transformer.to(device).eval().to(self.dtype)
|
284 |
|
285 |
lang_segment = LangSegment()
|
286 |
|
287 |
+
lang_segment.setfilters(
|
288 |
+
[
|
289 |
+
"af",
|
290 |
+
"am",
|
291 |
+
"an",
|
292 |
+
"ar",
|
293 |
+
"as",
|
294 |
+
"az",
|
295 |
+
"be",
|
296 |
+
"bg",
|
297 |
+
"bn",
|
298 |
+
"br",
|
299 |
+
"bs",
|
300 |
+
"ca",
|
301 |
+
"cs",
|
302 |
+
"cy",
|
303 |
+
"da",
|
304 |
+
"de",
|
305 |
+
"dz",
|
306 |
+
"el",
|
307 |
+
"en",
|
308 |
+
"eo",
|
309 |
+
"es",
|
310 |
+
"et",
|
311 |
+
"eu",
|
312 |
+
"fa",
|
313 |
+
"fi",
|
314 |
+
"fo",
|
315 |
+
"fr",
|
316 |
+
"ga",
|
317 |
+
"gl",
|
318 |
+
"gu",
|
319 |
+
"he",
|
320 |
+
"hi",
|
321 |
+
"hr",
|
322 |
+
"ht",
|
323 |
+
"hu",
|
324 |
+
"hy",
|
325 |
+
"id",
|
326 |
+
"is",
|
327 |
+
"it",
|
328 |
+
"ja",
|
329 |
+
"jv",
|
330 |
+
"ka",
|
331 |
+
"kk",
|
332 |
+
"km",
|
333 |
+
"kn",
|
334 |
+
"ko",
|
335 |
+
"ku",
|
336 |
+
"ky",
|
337 |
+
"la",
|
338 |
+
"lb",
|
339 |
+
"lo",
|
340 |
+
"lt",
|
341 |
+
"lv",
|
342 |
+
"mg",
|
343 |
+
"mk",
|
344 |
+
"ml",
|
345 |
+
"mn",
|
346 |
+
"mr",
|
347 |
+
"ms",
|
348 |
+
"mt",
|
349 |
+
"nb",
|
350 |
+
"ne",
|
351 |
+
"nl",
|
352 |
+
"nn",
|
353 |
+
"no",
|
354 |
+
"oc",
|
355 |
+
"or",
|
356 |
+
"pa",
|
357 |
+
"pl",
|
358 |
+
"ps",
|
359 |
+
"pt",
|
360 |
+
"qu",
|
361 |
+
"ro",
|
362 |
+
"ru",
|
363 |
+
"rw",
|
364 |
+
"se",
|
365 |
+
"si",
|
366 |
+
"sk",
|
367 |
+
"sl",
|
368 |
+
"sq",
|
369 |
+
"sr",
|
370 |
+
"sv",
|
371 |
+
"sw",
|
372 |
+
"ta",
|
373 |
+
"te",
|
374 |
+
"th",
|
375 |
+
"tl",
|
376 |
+
"tr",
|
377 |
+
"ug",
|
378 |
+
"uk",
|
379 |
+
"ur",
|
380 |
+
"vi",
|
381 |
+
"vo",
|
382 |
+
"wa",
|
383 |
+
"xh",
|
384 |
+
"zh",
|
385 |
+
"zu",
|
386 |
+
]
|
387 |
+
)
|
388 |
self.lang_segment = lang_segment
|
389 |
self.lyric_tokenizer = VoiceBpeTokenizer()
|
390 |
+
text_encoder_model = UMT5EncoderModel.from_pretrained(
|
391 |
+
text_encoder_checkpoint_path, torch_dtype=self.dtype
|
392 |
+
).eval()
|
393 |
text_encoder_model = text_encoder_model.to(device).to(self.dtype)
|
394 |
text_encoder_model.requires_grad_(False)
|
395 |
self.text_encoder_model = text_encoder_model
|
396 |
+
self.text_tokenizer = AutoTokenizer.from_pretrained(
|
397 |
+
text_encoder_checkpoint_path
|
398 |
+
)
|
399 |
self.loaded = True
|
400 |
|
401 |
# compile
|
|
|
405 |
self.text_encoder_model = torch.compile(self.text_encoder_model)
|
406 |
|
407 |
def get_text_embeddings(self, texts, device, text_max_length=256):
|
408 |
+
inputs = self.text_tokenizer(
|
409 |
+
texts,
|
410 |
+
return_tensors="pt",
|
411 |
+
padding=True,
|
412 |
+
truncation=True,
|
413 |
+
max_length=text_max_length,
|
414 |
+
)
|
415 |
inputs = {key: value.to(device) for key, value in inputs.items()}
|
416 |
if self.text_encoder_model.device != device:
|
417 |
self.text_encoder_model.to(device)
|
|
|
420 |
last_hidden_states = outputs.last_hidden_state
|
421 |
attention_mask = inputs["attention_mask"]
|
422 |
return last_hidden_states, attention_mask
|
423 |
+
|
424 |
+
def get_text_embeddings_null(
|
425 |
+
self, texts, device, text_max_length=256, tau=0.01, l_min=8, l_max=10
|
426 |
+
):
|
427 |
+
inputs = self.text_tokenizer(
|
428 |
+
texts,
|
429 |
+
return_tensors="pt",
|
430 |
+
padding=True,
|
431 |
+
truncation=True,
|
432 |
+
max_length=text_max_length,
|
433 |
+
)
|
434 |
inputs = {key: value.to(device) for key, value in inputs.items()}
|
435 |
if self.text_encoder_model.device != device:
|
436 |
self.text_encoder_model.to(device)
|
437 |
+
|
438 |
def forward_with_temperature(inputs, tau=0.01, l_min=8, l_max=10):
|
439 |
handlers = []
|
440 |
+
|
441 |
def hook(module, input, output):
|
442 |
output[:] *= tau
|
443 |
return output
|
444 |
+
|
445 |
for i in range(l_min, l_max):
|
446 |
+
handler = (
|
447 |
+
self.text_encoder_model.encoder.block[i]
|
448 |
+
.layer[0]
|
449 |
+
.SelfAttention.q.register_forward_hook(hook)
|
450 |
+
)
|
451 |
handlers.append(handler)
|
452 |
+
|
453 |
with torch.no_grad():
|
454 |
outputs = self.text_encoder_model(**inputs)
|
455 |
last_hidden_states = outputs.last_hidden_state
|
456 |
+
|
457 |
for hook in handlers:
|
458 |
hook.remove()
|
459 |
+
|
460 |
return last_hidden_states
|
461 |
+
|
462 |
last_hidden_states = forward_with_temperature(inputs, tau, l_min, l_max)
|
463 |
return last_hidden_states
|
464 |
|
465 |
def set_seeds(self, batch_size, manual_seeds=None):
|
466 |
+
processed_input_seeds = None
|
467 |
if manual_seeds is not None:
|
468 |
if isinstance(manual_seeds, str):
|
469 |
if "," in manual_seeds:
|
470 |
+
processed_input_seeds = list(map(int, manual_seeds.split(",")))
|
471 |
elif manual_seeds.isdigit():
|
472 |
+
processed_input_seeds = int(manual_seeds)
|
473 |
+
elif isinstance(manual_seeds, list) and all(
|
474 |
+
isinstance(s, int) for s in manual_seeds
|
475 |
+
):
|
476 |
+
if len(manual_seeds) > 0:
|
477 |
+
processed_input_seeds = list(manual_seeds)
|
478 |
+
elif isinstance(manual_seeds, int):
|
479 |
+
processed_input_seeds = manual_seeds
|
480 |
+
random_generators = [
|
481 |
+
torch.Generator(device=self.device) for _ in range(batch_size)
|
482 |
+
]
|
483 |
actual_seeds = []
|
484 |
for i in range(batch_size):
|
485 |
+
current_seed_for_generator = None
|
486 |
+
if processed_input_seeds is None:
|
487 |
+
current_seed_for_generator = torch.randint(0, 2**32, (1,)).item()
|
488 |
+
elif isinstance(processed_input_seeds, int):
|
489 |
+
current_seed_for_generator = processed_input_seeds
|
490 |
+
elif isinstance(processed_input_seeds, list):
|
491 |
+
if i < len(processed_input_seeds):
|
492 |
+
current_seed_for_generator = processed_input_seeds[i]
|
493 |
+
else:
|
494 |
+
current_seed_for_generator = processed_input_seeds[-1]
|
495 |
+
if current_seed_for_generator is None:
|
496 |
+
current_seed_for_generator = torch.randint(0, 2**32, (1,)).item()
|
497 |
+
random_generators[i].manual_seed(current_seed_for_generator)
|
498 |
+
actual_seeds.append(current_seed_for_generator)
|
499 |
return random_generators, actual_seeds
|
500 |
|
501 |
def get_lang(self, text):
|
502 |
language = "en"
|
503 |
+
try:
|
504 |
_ = self.lang_segment.getTexts(text)
|
505 |
langCounts = self.lang_segment.getCounts()
|
506 |
language = langCounts[0][0]
|
|
|
534 |
else:
|
535 |
token_idx = self.lyric_tokenizer.encode(line, lang)
|
536 |
if debug:
|
537 |
+
toks = self.lyric_tokenizer.batch_decode(
|
538 |
+
[[tok_id] for tok_id in token_idx]
|
539 |
+
)
|
540 |
logger.info(f"debbug {line} --> {lang} --> {toks}")
|
541 |
lyric_token_idx = lyric_token_idx + token_idx + [2]
|
542 |
except Exception as e:
|
|
|
565 |
attention_mask=None,
|
566 |
momentum_buffer=None,
|
567 |
momentum_buffer_tar=None,
|
568 |
+
return_src_pred=True,
|
569 |
):
|
570 |
noise_pred_src = None
|
571 |
if return_src_pred:
|
572 |
+
src_latent_model_input = (
|
573 |
+
torch.cat([zt_src, zt_src]) if do_classifier_free_guidance else zt_src
|
574 |
+
)
|
575 |
timestep = t.expand(src_latent_model_input.shape[0])
|
576 |
# source
|
577 |
noise_pred_src = self.ace_step_transformer(
|
|
|
586 |
).sample
|
587 |
|
588 |
if do_classifier_free_guidance:
|
589 |
+
noise_pred_with_cond_src, noise_pred_uncond_src = noise_pred_src.chunk(
|
590 |
+
2
|
591 |
+
)
|
592 |
if cfg_type == "apg":
|
593 |
noise_pred_src = apg_forward(
|
594 |
pred_cond=noise_pred_with_cond_src,
|
|
|
603 |
cfg_strength=guidance_scale,
|
604 |
)
|
605 |
|
606 |
+
tar_latent_model_input = (
|
607 |
+
torch.cat([zt_tar, zt_tar]) if do_classifier_free_guidance else zt_tar
|
608 |
+
)
|
609 |
timestep = t.expand(tar_latent_model_input.shape[0])
|
610 |
# target
|
611 |
noise_pred_tar = self.ace_step_transformer(
|
|
|
675 |
T_steps = infer_steps
|
676 |
frame_length = src_latents.shape[-1]
|
677 |
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
678 |
+
|
679 |
+
timesteps, T_steps = retrieve_timesteps(
|
680 |
+
scheduler, T_steps, device, timesteps=None
|
681 |
+
)
|
682 |
|
683 |
if do_classifier_free_guidance:
|
684 |
attention_mask = torch.cat([attention_mask] * 2, dim=0)
|
685 |
+
|
686 |
+
encoder_text_hidden_states = torch.cat(
|
687 |
+
[
|
688 |
+
encoder_text_hidden_states,
|
689 |
+
torch.zeros_like(encoder_text_hidden_states),
|
690 |
+
],
|
691 |
+
0,
|
692 |
+
)
|
693 |
text_attention_mask = torch.cat([text_attention_mask] * 2, dim=0)
|
694 |
|
695 |
+
target_encoder_text_hidden_states = torch.cat(
|
696 |
+
[
|
697 |
+
target_encoder_text_hidden_states,
|
698 |
+
torch.zeros_like(target_encoder_text_hidden_states),
|
699 |
+
],
|
700 |
+
0,
|
701 |
+
)
|
702 |
+
target_text_attention_mask = torch.cat(
|
703 |
+
[target_text_attention_mask] * 2, dim=0
|
704 |
+
)
|
705 |
|
706 |
+
speaker_embds = torch.cat(
|
707 |
+
[speaker_embds, torch.zeros_like(speaker_embds)], 0
|
708 |
+
)
|
709 |
+
target_speaker_embeds = torch.cat(
|
710 |
+
[target_speaker_embeds, torch.zeros_like(target_speaker_embeds)], 0
|
711 |
+
)
|
712 |
|
713 |
+
lyric_token_ids = torch.cat(
|
714 |
+
[lyric_token_ids, torch.zeros_like(lyric_token_ids)], 0
|
715 |
+
)
|
716 |
lyric_mask = torch.cat([lyric_mask, torch.zeros_like(lyric_mask)], 0)
|
717 |
|
718 |
+
target_lyric_token_ids = torch.cat(
|
719 |
+
[target_lyric_token_ids, torch.zeros_like(target_lyric_token_ids)], 0
|
720 |
+
)
|
721 |
+
target_lyric_mask = torch.cat(
|
722 |
+
[target_lyric_mask, torch.zeros_like(target_lyric_mask)], 0
|
723 |
+
)
|
724 |
|
725 |
momentum_buffer = MomentumBuffer()
|
726 |
momentum_buffer_tar = MomentumBuffer()
|
|
|
737 |
if i < n_min:
|
738 |
continue
|
739 |
|
740 |
+
t_i = t / 1000
|
741 |
|
742 |
+
if i + 1 < len(timesteps):
|
743 |
+
t_im1 = (timesteps[i + 1]) / 1000
|
744 |
else:
|
745 |
t_im1 = torch.zeros_like(t_i).to(t_i.device)
|
746 |
|
|
|
748 |
# Calculate the average of the V predictions
|
749 |
V_delta_avg = torch.zeros_like(x_src)
|
750 |
for k in range(n_avg):
|
751 |
+
fwd_noise = randn_tensor(
|
752 |
+
shape=x_src.shape,
|
753 |
+
generator=random_generators,
|
754 |
+
device=device,
|
755 |
+
dtype=dtype,
|
756 |
+
)
|
757 |
|
758 |
zt_src = (1 - t_i) * x_src + (t_i) * fwd_noise
|
759 |
|
|
|
777 |
guidance_scale=guidance_scale,
|
778 |
target_guidance_scale=target_guidance_scale,
|
779 |
attention_mask=attention_mask,
|
780 |
+
momentum_buffer=momentum_buffer,
|
781 |
)
|
782 |
+
V_delta_avg += (1 / n_avg) * (
|
783 |
+
Vt_tar - Vt_src
|
784 |
+
) # - (hfg-1)*( x_src))
|
785 |
|
786 |
# propagate direct ODE
|
787 |
zt_edit = zt_edit.to(torch.float32)
|
788 |
zt_edit = zt_edit + (t_im1 - t_i) * V_delta_avg
|
789 |
zt_edit = zt_edit.to(V_delta_avg.dtype)
|
790 |
+
else: # i >= T_steps-n_min # regular sampling for last n_min steps
|
791 |
if i == n_max:
|
792 |
+
fwd_noise = randn_tensor(
|
793 |
+
shape=x_src.shape,
|
794 |
+
generator=random_generators,
|
795 |
+
device=device,
|
796 |
+
dtype=dtype,
|
797 |
+
)
|
798 |
scheduler._init_step_index(t)
|
799 |
sigma = scheduler.sigmas[scheduler.step_index]
|
800 |
xt_src = sigma * fwd_noise + (1.0 - sigma) * x_src
|
801 |
xt_tar = zt_edit + xt_src - x_src
|
802 |
+
|
803 |
_, Vt_tar = self.calc_v(
|
804 |
zt_src=None,
|
805 |
zt_tar=xt_tar,
|
|
|
821 |
momentum_buffer_tar=momentum_buffer_tar,
|
822 |
return_src_pred=False,
|
823 |
)
|
824 |
+
|
825 |
dtype = Vt_tar.dtype
|
826 |
xt_tar = xt_tar.to(torch.float32)
|
827 |
prev_sample = xt_tar + (t_im1 - t_i) * Vt_tar
|
828 |
+
prev_sample = prev_sample.to(dtype)
|
829 |
xt_tar = prev_sample
|
830 |
+
|
831 |
target_latents = zt_edit if xt_tar is None else xt_tar
|
832 |
return target_latents
|
833 |
|
|
|
845 |
timesteps = scheduler.timesteps.unsqueeze(1).to(gt_latents.dtype)
|
846 |
indices = indices.to(timesteps.device).to(gt_latents.dtype).unsqueeze(1)
|
847 |
nearest_idx = torch.argmin(torch.cdist(indices, timesteps), dim=1)
|
848 |
+
sigma = (
|
849 |
+
scheduler.sigmas[nearest_idx]
|
850 |
+
.flatten()
|
851 |
+
.to(gt_latents.device)
|
852 |
+
.to(gt_latents.dtype)
|
853 |
+
)
|
854 |
while len(sigma.shape) < gt_latents.ndim:
|
855 |
sigma = sigma.unsqueeze(-1)
|
856 |
noisy_image = sigma * noise + (1.0 - sigma) * gt_latents
|
|
|
894 |
ref_latents=None,
|
895 |
):
|
896 |
|
897 |
+
logger.info(
|
898 |
+
"cfg_type: {}, guidance_scale: {}, omega_scale: {}".format(
|
899 |
+
cfg_type, guidance_scale, omega_scale
|
900 |
+
)
|
901 |
+
)
|
902 |
do_classifier_free_guidance = True
|
903 |
if guidance_scale == 0.0 or guidance_scale == 1.0:
|
904 |
do_classifier_free_guidance = False
|
905 |
+
|
906 |
do_double_condition_guidance = False
|
907 |
+
if (
|
908 |
+
guidance_scale_text is not None
|
909 |
+
and guidance_scale_text > 1.0
|
910 |
+
and guidance_scale_lyric is not None
|
911 |
+
and guidance_scale_lyric > 1.0
|
912 |
+
):
|
913 |
do_double_condition_guidance = True
|
914 |
+
logger.info(
|
915 |
+
"do_double_condition_guidance: {}, guidance_scale_text: {}, guidance_scale_lyric: {}".format(
|
916 |
+
do_double_condition_guidance,
|
917 |
+
guidance_scale_text,
|
918 |
+
guidance_scale_lyric,
|
919 |
+
)
|
920 |
+
)
|
921 |
|
922 |
device = encoder_text_hidden_states.device
|
923 |
dtype = encoder_text_hidden_states.dtype
|
|
|
933 |
num_train_timesteps=1000,
|
934 |
shift=3.0,
|
935 |
)
|
936 |
+
|
937 |
frame_length = int(duration * 44100 / 512 / 8)
|
938 |
if src_latents is not None:
|
939 |
frame_length = src_latents.shape[-1]
|
|
|
944 |
if len(oss_steps) > 0:
|
945 |
infer_steps = max(oss_steps)
|
946 |
scheduler.set_timesteps
|
947 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
948 |
+
scheduler,
|
949 |
+
num_inference_steps=infer_steps,
|
950 |
+
device=device,
|
951 |
+
timesteps=None,
|
952 |
+
)
|
953 |
new_timesteps = torch.zeros(len(oss_steps), dtype=dtype, device=device)
|
954 |
for idx in range(len(oss_steps)):
|
955 |
+
new_timesteps[idx] = timesteps[oss_steps[idx] - 1]
|
956 |
num_inference_steps = len(oss_steps)
|
957 |
sigmas = (new_timesteps / 1000).float().cpu().numpy()
|
958 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
959 |
+
scheduler,
|
960 |
+
num_inference_steps=num_inference_steps,
|
961 |
+
device=device,
|
962 |
+
sigmas=sigmas,
|
963 |
+
)
|
964 |
+
logger.info(
|
965 |
+
f"oss_steps: {oss_steps}, num_inference_steps: {num_inference_steps} after remapping to timesteps {timesteps}"
|
966 |
+
)
|
967 |
else:
|
968 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
969 |
+
scheduler,
|
970 |
+
num_inference_steps=infer_steps,
|
971 |
+
device=device,
|
972 |
+
timesteps=None,
|
973 |
+
)
|
974 |
+
|
975 |
+
target_latents = randn_tensor(
|
976 |
+
shape=(bsz, 8, 16, frame_length),
|
977 |
+
generator=random_generators,
|
978 |
+
device=device,
|
979 |
+
dtype=dtype,
|
980 |
+
)
|
981 |
+
|
982 |
is_repaint = False
|
983 |
+
is_extend = False
|
984 |
if add_retake_noise:
|
985 |
n_min = int(infer_steps * (1 - retake_variance))
|
986 |
+
retake_variance = (
|
987 |
+
torch.tensor(retake_variance * math.pi / 2).to(device).to(dtype)
|
988 |
+
)
|
989 |
+
retake_latents = randn_tensor(
|
990 |
+
shape=(bsz, 8, 16, frame_length),
|
991 |
+
generator=retake_random_generators,
|
992 |
+
device=device,
|
993 |
+
dtype=dtype,
|
994 |
+
)
|
995 |
repaint_start_frame = int(repaint_start * 44100 / 512 / 8)
|
996 |
repaint_end_frame = int(repaint_end * 44100 / 512 / 8)
|
997 |
x0 = src_latents
|
998 |
# retake
|
999 |
+
is_repaint = repaint_end_frame - repaint_start_frame != frame_length
|
1000 |
+
|
1001 |
is_extend = (repaint_start_frame < 0) or (repaint_end_frame > frame_length)
|
1002 |
if is_extend:
|
1003 |
is_repaint = True
|
|
|
1005 |
# TODO: train a mask aware repainting controlnet
|
1006 |
# to make sure mean = 0, std = 1
|
1007 |
if not is_repaint:
|
1008 |
+
target_latents = (
|
1009 |
+
torch.cos(retake_variance) * target_latents
|
1010 |
+
+ torch.sin(retake_variance) * retake_latents
|
1011 |
+
)
|
1012 |
elif not is_extend:
|
1013 |
+
# if repaint_end_frame
|
1014 |
+
repaint_mask = torch.zeros(
|
1015 |
+
(bsz, 8, 16, frame_length), device=device, dtype=dtype
|
1016 |
+
)
|
1017 |
repaint_mask[:, :, :, repaint_start_frame:repaint_end_frame] = 1.0
|
1018 |
+
repaint_noise = (
|
1019 |
+
torch.cos(retake_variance) * target_latents
|
1020 |
+
+ torch.sin(retake_variance) * retake_latents
|
1021 |
+
)
|
1022 |
+
repaint_noise = torch.where(
|
1023 |
+
repaint_mask == 1.0, repaint_noise, target_latents
|
1024 |
+
)
|
1025 |
zt_edit = x0.clone()
|
1026 |
z0 = repaint_noise
|
1027 |
elif is_extend:
|
|
|
1037 |
if repaint_start_frame < 0:
|
1038 |
left_pad_frame_length = abs(repaint_start_frame)
|
1039 |
frame_length = left_pad_frame_length + gt_latents.shape[-1]
|
1040 |
+
extend_gt_latents = torch.nn.functional.pad(
|
1041 |
+
gt_latents, (left_pad_frame_length, 0), "constant", 0
|
1042 |
+
)
|
1043 |
if frame_length > max_infer_fame_length:
|
1044 |
right_trim_length = frame_length - max_infer_fame_length
|
1045 |
+
extend_gt_latents = extend_gt_latents[
|
1046 |
+
:, :, :, :max_infer_fame_length
|
1047 |
+
]
|
1048 |
+
to_right_pad_gt_latents = extend_gt_latents[
|
1049 |
+
:, :, :, -right_trim_length:
|
1050 |
+
]
|
1051 |
frame_length = max_infer_fame_length
|
1052 |
repaint_start_frame = 0
|
1053 |
gt_latents = extend_gt_latents
|
1054 |
+
|
1055 |
if repaint_end_frame > src_latents_length:
|
1056 |
right_pad_frame_length = repaint_end_frame - gt_latents.shape[-1]
|
1057 |
frame_length = gt_latents.shape[-1] + right_pad_frame_length
|
1058 |
+
extend_gt_latents = torch.nn.functional.pad(
|
1059 |
+
gt_latents, (0, right_pad_frame_length), "constant", 0
|
1060 |
+
)
|
1061 |
if frame_length > max_infer_fame_length:
|
1062 |
left_trim_length = frame_length - max_infer_fame_length
|
1063 |
+
extend_gt_latents = extend_gt_latents[
|
1064 |
+
:, :, :, -max_infer_fame_length:
|
1065 |
+
]
|
1066 |
+
to_left_pad_gt_latents = extend_gt_latents[
|
1067 |
+
:, :, :, :left_trim_length
|
1068 |
+
]
|
1069 |
frame_length = max_infer_fame_length
|
1070 |
repaint_end_frame = frame_length
|
1071 |
gt_latents = extend_gt_latents
|
1072 |
|
1073 |
+
repaint_mask = torch.zeros(
|
1074 |
+
(bsz, 8, 16, frame_length), device=device, dtype=dtype
|
1075 |
+
)
|
1076 |
if left_pad_frame_length > 0:
|
1077 |
+
repaint_mask[:, :, :, :left_pad_frame_length] = 1.0
|
1078 |
if right_pad_frame_length > 0:
|
1079 |
+
repaint_mask[:, :, :, -right_pad_frame_length:] = 1.0
|
1080 |
x0 = gt_latents
|
1081 |
padd_list = []
|
1082 |
if left_pad_frame_length > 0:
|
1083 |
padd_list.append(retake_latents[:, :, :, :left_pad_frame_length])
|
1084 |
+
padd_list.append(
|
1085 |
+
target_latents[
|
1086 |
+
:,
|
1087 |
+
:,
|
1088 |
+
:,
|
1089 |
+
left_trim_length : target_latents.shape[-1] - right_trim_length,
|
1090 |
+
]
|
1091 |
+
)
|
1092 |
if right_pad_frame_length > 0:
|
1093 |
padd_list.append(retake_latents[:, :, :, -right_pad_frame_length:])
|
1094 |
target_latents = torch.cat(padd_list, dim=-1)
|
1095 |
+
assert (
|
1096 |
+
target_latents.shape[-1] == x0.shape[-1]
|
1097 |
+
), f"{target_latents.shape=} {x0.shape=}"
|
1098 |
zt_edit = x0.clone()
|
1099 |
z0 = target_latents
|
1100 |
|
1101 |
init_timestep = 1000
|
1102 |
if audio2audio_enable and ref_latents is not None:
|
1103 |
+
target_latents, init_timestep = self.add_latents_noise(
|
1104 |
+
gt_latents=ref_latents,
|
1105 |
+
variance=(1 - ref_audio_strength),
|
1106 |
+
noise=target_latents,
|
1107 |
+
scheduler=scheduler,
|
1108 |
+
)
|
1109 |
|
1110 |
attention_mask = torch.ones(bsz, frame_length, device=device, dtype=dtype)
|
1111 |
+
|
1112 |
# guidance interval
|
1113 |
start_idx = int(num_inference_steps * ((1 - guidance_interval) / 2))
|
1114 |
end_idx = int(num_inference_steps * (guidance_interval / 2 + 0.5))
|
1115 |
+
logger.info(
|
1116 |
+
f"start_idx: {start_idx}, end_idx: {end_idx}, num_inference_steps: {num_inference_steps}"
|
1117 |
+
)
|
1118 |
|
1119 |
momentum_buffer = MomentumBuffer()
|
1120 |
|
1121 |
def forward_encoder_with_temperature(self, inputs, tau=0.01, l_min=4, l_max=6):
|
1122 |
handlers = []
|
1123 |
+
|
1124 |
def hook(module, input, output):
|
1125 |
output[:] *= tau
|
1126 |
return output
|
1127 |
+
|
1128 |
for i in range(l_min, l_max):
|
1129 |
+
handler = self.ace_step_transformer.lyric_encoder.encoders[
|
1130 |
+
i
|
1131 |
+
].self_attn.linear_q.register_forward_hook(hook)
|
1132 |
handlers.append(handler)
|
1133 |
+
|
1134 |
+
encoder_hidden_states, encoder_hidden_mask = (
|
1135 |
+
self.ace_step_transformer.encode(**inputs)
|
1136 |
+
)
|
1137 |
+
|
1138 |
for hook in handlers:
|
1139 |
hook.remove()
|
1140 |
+
|
1141 |
return encoder_hidden_states
|
1142 |
|
1143 |
# P(speaker, text, lyric)
|
|
|
1154 |
encoder_hidden_states_null = forward_encoder_with_temperature(
|
1155 |
self,
|
1156 |
inputs={
|
1157 |
+
"encoder_text_hidden_states": (
|
1158 |
+
encoder_text_hidden_states_null
|
1159 |
+
if encoder_text_hidden_states_null is not None
|
1160 |
+
else torch.zeros_like(encoder_text_hidden_states)
|
1161 |
+
),
|
1162 |
"text_attention_mask": text_attention_mask,
|
1163 |
"speaker_embeds": torch.zeros_like(speaker_embds),
|
1164 |
"lyric_token_idx": lyric_token_ids,
|
1165 |
"lyric_mask": lyric_mask,
|
1166 |
+
},
|
1167 |
)
|
1168 |
else:
|
1169 |
# P(null_speaker, null_text, null_lyric)
|
|
|
1174 |
torch.zeros_like(lyric_token_ids),
|
1175 |
lyric_mask,
|
1176 |
)
|
1177 |
+
|
1178 |
encoder_hidden_states_no_lyric = None
|
1179 |
if do_double_condition_guidance:
|
1180 |
# P(null_speaker, text, lyric_weaker)
|
|
|
1187 |
"speaker_embeds": torch.zeros_like(speaker_embds),
|
1188 |
"lyric_token_idx": lyric_token_ids,
|
1189 |
"lyric_mask": lyric_mask,
|
1190 |
+
},
|
1191 |
)
|
1192 |
# P(null_speaker, text, no_lyric)
|
1193 |
else:
|
|
|
1199 |
lyric_mask,
|
1200 |
)
|
1201 |
|
1202 |
+
def forward_diffusion_with_temperature(
|
1203 |
+
self, hidden_states, timestep, inputs, tau=0.01, l_min=15, l_max=20
|
1204 |
+
):
|
1205 |
handlers = []
|
1206 |
+
|
1207 |
def hook(module, input, output):
|
1208 |
output[:] *= tau
|
1209 |
return output
|
1210 |
+
|
1211 |
for i in range(l_min, l_max):
|
1212 |
+
handler = self.ace_step_transformer.transformer_blocks[
|
1213 |
+
i
|
1214 |
+
].attn.to_q.register_forward_hook(hook)
|
1215 |
handlers.append(handler)
|
1216 |
+
handler = self.ace_step_transformer.transformer_blocks[
|
1217 |
+
i
|
1218 |
+
].cross_attn.to_q.register_forward_hook(hook)
|
1219 |
handlers.append(handler)
|
1220 |
|
1221 |
+
sample = self.ace_step_transformer.decode(
|
1222 |
+
hidden_states=hidden_states, timestep=timestep, **inputs
|
1223 |
+
).sample
|
1224 |
+
|
1225 |
for hook in handlers:
|
1226 |
hook.remove()
|
1227 |
+
|
1228 |
return sample
|
1229 |
+
|
1230 |
for i, t in tqdm(enumerate(timesteps), total=num_inference_steps):
|
1231 |
|
1232 |
if t > init_timestep:
|
|
|
1249 |
# compute current guidance scale
|
1250 |
if guidance_interval_decay > 0:
|
1251 |
# Linearly interpolate to calculate the current guidance scale
|
1252 |
+
progress = (i - start_idx) / (
|
1253 |
+
end_idx - start_idx - 1
|
1254 |
+
) # ๅฝไธๅๅฐ[0,1]
|
1255 |
+
current_guidance_scale = (
|
1256 |
+
guidance_scale
|
1257 |
+
- (guidance_scale - min_guidance_scale)
|
1258 |
+
* progress
|
1259 |
+
* guidance_interval_decay
|
1260 |
+
)
|
1261 |
else:
|
1262 |
current_guidance_scale = guidance_scale
|
1263 |
|
|
|
1275 |
).sample
|
1276 |
|
1277 |
noise_pred_with_only_text_cond = None
|
1278 |
+
if (
|
1279 |
+
do_double_condition_guidance
|
1280 |
+
and encoder_hidden_states_no_lyric is not None
|
1281 |
+
):
|
1282 |
noise_pred_with_only_text_cond = self.ace_step_transformer.decode(
|
1283 |
hidden_states=latent_model_input,
|
1284 |
attention_mask=attention_mask,
|
|
|
1310 |
timestep=timestep,
|
1311 |
).sample
|
1312 |
|
1313 |
+
if (
|
1314 |
+
do_double_condition_guidance
|
1315 |
+
and noise_pred_with_only_text_cond is not None
|
1316 |
+
):
|
1317 |
noise_pred = cfg_double_condition_forward(
|
1318 |
cond_output=noise_pred_with_cond,
|
1319 |
uncond_output=noise_pred_uncond,
|
|
|
1342 |
guidance_scale=current_guidance_scale,
|
1343 |
i=i,
|
1344 |
zero_steps=zero_steps,
|
1345 |
+
use_zero_init=use_zero_init,
|
1346 |
)
|
1347 |
else:
|
1348 |
latent_model_input = latents
|
|
|
1357 |
).sample
|
1358 |
|
1359 |
if is_repaint and i >= n_min:
|
1360 |
+
t_i = t / 1000
|
1361 |
+
if i + 1 < len(timesteps):
|
1362 |
+
t_im1 = (timesteps[i + 1]) / 1000
|
1363 |
else:
|
1364 |
t_im1 = torch.zeros_like(t_i).to(t_i.device)
|
1365 |
dtype = noise_pred.dtype
|
|
|
1368 |
prev_sample = prev_sample.to(dtype)
|
1369 |
target_latents = prev_sample
|
1370 |
zt_src = (1 - t_im1) * x0 + (t_im1) * z0
|
1371 |
+
target_latents = torch.where(
|
1372 |
+
repaint_mask == 1.0, target_latents, zt_src
|
1373 |
+
)
|
1374 |
else:
|
1375 |
+
target_latents = scheduler.step(
|
1376 |
+
model_output=noise_pred,
|
1377 |
+
timestep=t,
|
1378 |
+
sample=target_latents,
|
1379 |
+
return_dict=False,
|
1380 |
+
omega=omega_scale,
|
1381 |
+
)[0]
|
1382 |
|
1383 |
if is_extend:
|
1384 |
if to_right_pad_gt_latents is not None:
|
1385 |
+
target_latents = torch.cat(
|
1386 |
+
[target_latents, to_right_pad_gt_latents], dim=-1
|
1387 |
+
)
|
1388 |
if to_left_pad_gt_latents is not None:
|
1389 |
+
target_latents = torch.cat(
|
1390 |
+
[to_right_pad_gt_latents, target_latents], dim=0
|
1391 |
+
)
|
1392 |
return target_latents
|
1393 |
|
1394 |
+
def latents2audio(
|
1395 |
+
self,
|
1396 |
+
latents,
|
1397 |
+
target_wav_duration_second=30,
|
1398 |
+
sample_rate=48000,
|
1399 |
+
save_path=None,
|
1400 |
+
format="mp3",
|
1401 |
+
):
|
1402 |
output_audio_paths = []
|
1403 |
bs = latents.shape[0]
|
1404 |
audio_lengths = [target_wav_duration_second * sample_rate] * bs
|
|
|
1407 |
_, pred_wavs = self.music_dcae.decode(pred_latents, sr=sample_rate)
|
1408 |
pred_wavs = [pred_wav.cpu().float() for pred_wav in pred_wavs]
|
1409 |
for i in tqdm(range(bs)):
|
1410 |
+
output_audio_path = self.save_wav_file(
|
1411 |
+
pred_wavs[i], i, sample_rate=sample_rate
|
1412 |
+
)
|
1413 |
output_audio_paths.append(output_audio_path)
|
1414 |
return output_audio_paths
|
1415 |
|
1416 |
+
def save_wav_file(
|
1417 |
+
self, target_wav, idx, save_path=None, sample_rate=48000, format="mp3"
|
1418 |
+
):
|
1419 |
if save_path is None:
|
1420 |
logger.warning("save_path is None, using default path ./outputs/")
|
1421 |
base_path = f"./outputs"
|
|
|
1424 |
base_path = save_path
|
1425 |
ensure_directory_exists(base_path)
|
1426 |
|
1427 |
+
output_path_flac = (
|
1428 |
+
f"{base_path}/output_{time.strftime('%Y%m%d%H%M%S')}_{idx}.{format}"
|
1429 |
+
)
|
1430 |
target_wav = target_wav.float()
|
1431 |
+
torchaudio.save(
|
1432 |
+
output_path_flac,
|
1433 |
+
target_wav,
|
1434 |
+
sample_rate=sample_rate,
|
1435 |
+
format=format,
|
1436 |
+
compression=torio.io.CodecConfig(bit_rate=320000),
|
1437 |
+
)
|
1438 |
return output_path_flac
|
1439 |
|
1440 |
def infer_latents(self, input_audio_path):
|
|
|
1460 |
omega_scale: int = 10.0,
|
1461 |
manual_seeds: list = None,
|
1462 |
guidance_interval: float = 0.5,
|
1463 |
+
guidance_interval_decay: float = 0.0,
|
1464 |
min_guidance_scale: float = 3.0,
|
1465 |
use_erg_tag: bool = True,
|
1466 |
use_erg_lyric: bool = True,
|
|
|
1503 |
start_time = time.time()
|
1504 |
|
1505 |
random_generators, actual_seeds = self.set_seeds(batch_size, manual_seeds)
|
1506 |
+
retake_random_generators, actual_retake_seeds = self.set_seeds(
|
1507 |
+
batch_size, retake_seeds
|
1508 |
+
)
|
1509 |
|
1510 |
if isinstance(oss_steps, str) and len(oss_steps) > 0:
|
1511 |
oss_steps = list(map(int, oss_steps.split(",")))
|
1512 |
else:
|
1513 |
oss_steps = []
|
1514 |
+
|
1515 |
texts = [prompt]
|
1516 |
+
encoder_text_hidden_states, text_attention_mask = self.get_text_embeddings(
|
1517 |
+
texts, self.device
|
1518 |
+
)
|
1519 |
encoder_text_hidden_states = encoder_text_hidden_states.repeat(batch_size, 1, 1)
|
1520 |
text_attention_mask = text_attention_mask.repeat(batch_size, 1)
|
1521 |
|
1522 |
encoder_text_hidden_states_null = None
|
1523 |
if use_erg_tag:
|
1524 |
+
encoder_text_hidden_states_null = self.get_text_embeddings_null(
|
1525 |
+
texts, self.device
|
1526 |
+
)
|
1527 |
+
encoder_text_hidden_states_null = encoder_text_hidden_states_null.repeat(
|
1528 |
+
batch_size, 1, 1
|
1529 |
+
)
|
1530 |
|
1531 |
# not support for released checkpoint
|
1532 |
speaker_embeds = torch.zeros(batch_size, 512).to(self.device).to(self.dtype)
|
|
|
1537 |
if len(lyrics) > 0:
|
1538 |
lyric_token_idx = self.tokenize_lyrics(lyrics, debug=debug)
|
1539 |
lyric_mask = [1] * len(lyric_token_idx)
|
1540 |
+
lyric_token_idx = (
|
1541 |
+
torch.tensor(lyric_token_idx)
|
1542 |
+
.unsqueeze(0)
|
1543 |
+
.to(self.device)
|
1544 |
+
.repeat(batch_size, 1)
|
1545 |
+
)
|
1546 |
+
lyric_mask = (
|
1547 |
+
torch.tensor(lyric_mask)
|
1548 |
+
.unsqueeze(0)
|
1549 |
+
.to(self.device)
|
1550 |
+
.repeat(batch_size, 1)
|
1551 |
+
)
|
1552 |
|
1553 |
if audio_duration <= 0:
|
1554 |
audio_duration = random.uniform(30.0, 240.0)
|
|
|
1563 |
if task == "retake":
|
1564 |
repaint_start = 0
|
1565 |
repaint_end = audio_duration
|
1566 |
+
|
1567 |
src_latents = None
|
1568 |
if src_audio_path is not None:
|
1569 |
+
assert src_audio_path is not None and task in (
|
1570 |
+
"repaint",
|
1571 |
+
"edit",
|
1572 |
+
"extend",
|
1573 |
+
), "src_audio_path is required for retake/repaint/extend task"
|
1574 |
+
assert os.path.exists(
|
1575 |
+
src_audio_path
|
1576 |
+
), f"src_audio_path {src_audio_path} does not exist"
|
1577 |
src_latents = self.infer_latents(src_audio_path)
|
1578 |
|
1579 |
ref_latents = None
|
1580 |
if ref_audio_input is not None and audio2audio_enable:
|
1581 |
+
assert (
|
1582 |
+
ref_audio_input is not None
|
1583 |
+
), "ref_audio_input is required for audio2audio task"
|
1584 |
assert os.path.exists(
|
1585 |
ref_audio_input
|
1586 |
), f"ref_audio_input {ref_audio_input} does not exist"
|
|
|
1588 |
|
1589 |
if task == "edit":
|
1590 |
texts = [edit_target_prompt]
|
1591 |
+
target_encoder_text_hidden_states, target_text_attention_mask = (
|
1592 |
+
self.get_text_embeddings(texts, self.device)
|
1593 |
+
)
|
1594 |
+
target_encoder_text_hidden_states = (
|
1595 |
+
target_encoder_text_hidden_states.repeat(batch_size, 1, 1)
|
1596 |
+
)
|
1597 |
+
target_text_attention_mask = target_text_attention_mask.repeat(
|
1598 |
+
batch_size, 1
|
1599 |
+
)
|
1600 |
|
1601 |
+
target_lyric_token_idx = (
|
1602 |
+
torch.tensor([0]).repeat(batch_size, 1).to(self.device).long()
|
1603 |
+
)
|
1604 |
+
target_lyric_mask = (
|
1605 |
+
torch.tensor([0]).repeat(batch_size, 1).to(self.device).long()
|
1606 |
+
)
|
1607 |
if len(edit_target_lyrics) > 0:
|
1608 |
+
target_lyric_token_idx = self.tokenize_lyrics(
|
1609 |
+
edit_target_lyrics, debug=True
|
1610 |
+
)
|
1611 |
target_lyric_mask = [1] * len(target_lyric_token_idx)
|
1612 |
+
target_lyric_token_idx = (
|
1613 |
+
torch.tensor(target_lyric_token_idx)
|
1614 |
+
.unsqueeze(0)
|
1615 |
+
.to(self.device)
|
1616 |
+
.repeat(batch_size, 1)
|
1617 |
+
)
|
1618 |
+
target_lyric_mask = (
|
1619 |
+
torch.tensor(target_lyric_mask)
|
1620 |
+
.unsqueeze(0)
|
1621 |
+
.to(self.device)
|
1622 |
+
.repeat(batch_size, 1)
|
1623 |
+
)
|
1624 |
|
1625 |
target_speaker_embeds = speaker_embeds.clone()
|
1626 |
|
|
|
1636 |
target_lyric_token_ids=target_lyric_token_idx,
|
1637 |
target_lyric_mask=target_lyric_mask,
|
1638 |
src_latents=src_latents,
|
1639 |
+
random_generators=retake_random_generators, # more diversity
|
1640 |
infer_steps=infer_step,
|
1641 |
guidance_scale=guidance_scale,
|
1642 |
n_min=edit_n_min,
|
|
|
1724 |
"repaint_end": repaint_end,
|
1725 |
"edit_n_min": edit_n_min,
|
1726 |
"edit_n_max": edit_n_max,
|
1727 |
+
"edit_n_avg": edit_n_avg,
|
1728 |
"src_audio_path": src_audio_path,
|
1729 |
"edit_target_prompt": edit_target_prompt,
|
1730 |
"edit_target_lyrics": edit_target_lyrics,
|
|
|
1734 |
}
|
1735 |
# save input_params_json
|
1736 |
for output_audio_path in output_paths:
|
1737 |
+
input_params_json_save_path = output_audio_path.replace(
|
1738 |
+
f".{format}", "_input_params.json"
|
1739 |
+
)
|
1740 |
input_params_json["audio_path"] = output_audio_path
|
1741 |
with open(input_params_json_save_path, "w", encoding="utf-8") as f:
|
1742 |
json.dump(input_params_json, f, indent=4, ensure_ascii=False)
|