Upload example.py with huggingface_hub
Browse files- example.py +80 -0
example.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import torch
|
3 |
+
from diffusers import DDPMScheduler
|
4 |
+
import json
|
5 |
+
from PIL import Image
|
6 |
+
import numpy as np
|
7 |
+
|
8 |
+
class LetterConditionedUnet(torch.nn.Module):
|
9 |
+
def __init__(self, num_classes=26, class_emb_size=8):
|
10 |
+
super().__init__()
|
11 |
+
from diffusers import UNet2DModel
|
12 |
+
|
13 |
+
self.class_emb = torch.nn.Embedding(num_classes, class_emb_size)
|
14 |
+
|
15 |
+
self.model = UNet2DModel(
|
16 |
+
sample_size=512,
|
17 |
+
in_channels=1 + class_emb_size,
|
18 |
+
out_channels=1,
|
19 |
+
layers_per_block=2,
|
20 |
+
block_out_channels=(64, 128, 256, 512, 512),
|
21 |
+
down_block_types=(
|
22 |
+
"DownBlock2D",
|
23 |
+
"DownBlock2D",
|
24 |
+
"AttnDownBlock2D",
|
25 |
+
"AttnDownBlock2D",
|
26 |
+
"AttnDownBlock2D",
|
27 |
+
),
|
28 |
+
up_block_types=(
|
29 |
+
"AttnUpBlock2D",
|
30 |
+
"AttnUpBlock2D",
|
31 |
+
"AttnUpBlock2D",
|
32 |
+
"UpBlock2D",
|
33 |
+
"UpBlock2D",
|
34 |
+
),
|
35 |
+
)
|
36 |
+
|
37 |
+
def forward(self, x, t, class_labels):
|
38 |
+
bs, ch, w, h = x.shape
|
39 |
+
class_cond = self.class_emb(class_labels)
|
40 |
+
class_cond = class_cond.view(bs, class_cond.shape[1], 1, 1).expand(bs, class_cond.shape[1], w, h)
|
41 |
+
net_input = torch.cat((x, class_cond), 1)
|
42 |
+
return self.model(net_input, t).sample
|
43 |
+
|
44 |
+
def generate_letter(letter, model_path="./"):
|
45 |
+
"""Genera una imagen de la letra especificada"""
|
46 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
47 |
+
|
48 |
+
# Cargar modelo
|
49 |
+
model = LetterConditionedUnet()
|
50 |
+
model.load_state_dict(torch.load(f"{model_path}/pytorch_model.bin", map_location=device))
|
51 |
+
model = model.to(device)
|
52 |
+
model.eval()
|
53 |
+
|
54 |
+
# Cargar scheduler
|
55 |
+
with open(f"{model_path}/scheduler_config.json", 'r') as f:
|
56 |
+
scheduler_config = json.load(f)
|
57 |
+
scheduler = DDPMScheduler(**scheduler_config)
|
58 |
+
|
59 |
+
# Preparar entrada
|
60 |
+
letter_label = ord(letter.upper()) - 65 # Convertir letra a n煤mero
|
61 |
+
x = torch.randn(1, 1, 512, 512, device=device)
|
62 |
+
labels = torch.tensor([letter_label], device=device)
|
63 |
+
|
64 |
+
# Generar
|
65 |
+
with torch.no_grad():
|
66 |
+
for t in scheduler.timesteps:
|
67 |
+
residual = model(x, t, labels)
|
68 |
+
x = scheduler.step(residual, t, x).prev_sample
|
69 |
+
|
70 |
+
# Convertir a imagen
|
71 |
+
image = x[0, 0].cpu().numpy()
|
72 |
+
image = (image + 1) / 2 # Desnormalizar de [-1,1] a [0,1]
|
73 |
+
image = (image * 255).astype(np.uint8)
|
74 |
+
|
75 |
+
return Image.fromarray(image, mode='L')
|
76 |
+
|
77 |
+
# Ejemplo de uso
|
78 |
+
if __name__ == "__main__":
|
79 |
+
letter_image = generate_letter('A')
|
80 |
+
letter_image.save('generated_letter_A.png')
|