Spaces:
Sleeping
Sleeping
app.py and dependencies
Browse files- app.py +239 -0
- requirements.txt +6 -0
app.py
ADDED
@@ -0,0 +1,239 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from torch import nn
|
4 |
+
from torchvision import transforms
|
5 |
+
from PIL import Image
|
6 |
+
import numpy as np
|
7 |
+
import os
|
8 |
+
import random
|
9 |
+
import torch.nn.functional as F
|
10 |
+
from huggingface_hub import hf_hub_download
|
11 |
+
|
12 |
+
# DEFINICI脫N DE BLOQUES DE RED
|
13 |
+
class ResBlk(nn.Module):
|
14 |
+
def __init__(self, dim_in, dim_out, normalize=False, downsample=False):
|
15 |
+
super().__init__()
|
16 |
+
self.conv1 = nn.Conv2d(dim_in, dim_out, 3, 1, 1)
|
17 |
+
self.norm1 = nn.InstanceNorm2d(dim_out, affine=True) if normalize else None
|
18 |
+
self.relu1 = nn.ReLU(inplace=True)
|
19 |
+
self.conv2 = nn.Conv2d(dim_out, dim_out, 3, 1, 1)
|
20 |
+
self.norm2 = nn.InstanceNorm2d(dim_out, affine=True) if normalize else None
|
21 |
+
self.relu2 = nn.ReLU(inplace=True)
|
22 |
+
self.downsample = downsample
|
23 |
+
if self.downsample:
|
24 |
+
self.avg_pool = nn.AvgPool2d(2)
|
25 |
+
|
26 |
+
def forward(self, x):
|
27 |
+
residual = x
|
28 |
+
out = self.conv1(x)
|
29 |
+
if self.norm1:
|
30 |
+
out = self.norm1(out)
|
31 |
+
out = self.relu1(out)
|
32 |
+
out = self.conv2(out)
|
33 |
+
if self.norm2:
|
34 |
+
out = self.norm2(out)
|
35 |
+
out = self.relu2(out)
|
36 |
+
if self.downsample:
|
37 |
+
out = self.avg_pool(out)
|
38 |
+
residual = self.avg_pool(residual)
|
39 |
+
out = out + residual
|
40 |
+
return out
|
41 |
+
|
42 |
+
class AdainResBlk(nn.Module):
|
43 |
+
def __init__(self, dim_in, dim_out, style_dim, upsample=False):
|
44 |
+
super().__init__()
|
45 |
+
self.conv1 = nn.Conv2d(dim_in, dim_out, 3, 1, 1)
|
46 |
+
self.norm1 = AdaIN(dim_out, style_dim)
|
47 |
+
self.relu1 = nn.ReLU(inplace=True)
|
48 |
+
self.conv2 = nn.Conv2d(dim_out, dim_out, 3, 1, 1)
|
49 |
+
self.norm2 = AdaIN(dim_out, style_dim)
|
50 |
+
self.relu2 = nn.ReLU(inplace=True)
|
51 |
+
self.upsample = upsample
|
52 |
+
|
53 |
+
def forward(self, x, s):
|
54 |
+
residual = x
|
55 |
+
if self.upsample:
|
56 |
+
residual = F.interpolate(residual, scale_factor=2, mode='nearest')
|
57 |
+
out = self.conv1(x)
|
58 |
+
out = self.norm1(out, s)
|
59 |
+
out = self.relu1(out)
|
60 |
+
if self.upsample:
|
61 |
+
out = F.interpolate(out, scale_factor=2, mode='nearest')
|
62 |
+
out = self.conv2(out)
|
63 |
+
out = self.norm2(out, s)
|
64 |
+
out = self.relu2(out)
|
65 |
+
out = out + residual
|
66 |
+
return out
|
67 |
+
|
68 |
+
class AdaIN(nn.Module):
|
69 |
+
def __init__(self, num_features, style_dim):
|
70 |
+
super().__init__()
|
71 |
+
self.norm = nn.InstanceNorm2d(num_features, affine=False)
|
72 |
+
self.fc = nn.Linear(style_dim, num_features * 2)
|
73 |
+
|
74 |
+
def forward(self, x, s):
|
75 |
+
h = self.fc(s)
|
76 |
+
gamma, beta = torch.chunk(h, 2, dim=1)
|
77 |
+
gamma = gamma.unsqueeze(2).unsqueeze(3)
|
78 |
+
beta = beta.unsqueeze(2).unsqueeze(3)
|
79 |
+
return (1 + gamma) * self.norm(x) + beta
|
80 |
+
|
81 |
+
class MappingNetwork(nn.Module):
|
82 |
+
def __init__(self, latent_dim, style_dim, num_domains):
|
83 |
+
super().__init__()
|
84 |
+
layers = []
|
85 |
+
layers += [nn.Linear(latent_dim + num_domains, 512)]
|
86 |
+
layers += [nn.ReLU()]
|
87 |
+
for _ in range(3):
|
88 |
+
layers += [nn.Linear(512, 512)]
|
89 |
+
layers += [nn.ReLU()]
|
90 |
+
self.shared = nn.Sequential(*layers)
|
91 |
+
self.unshared = nn.ModuleList()
|
92 |
+
for _ in range(num_domains):
|
93 |
+
self.unshared += [nn.Linear(512, style_dim)]
|
94 |
+
|
95 |
+
def forward(self, z, y):
|
96 |
+
h = torch.cat([z, y], dim=1)
|
97 |
+
h = self.shared(h)
|
98 |
+
out = []
|
99 |
+
for layer in self.unshared:
|
100 |
+
out += [layer(h)]
|
101 |
+
out = torch.stack(out, dim=1) # (batch, num_domains, style_dim)
|
102 |
+
idx = torch.LongTensor(range(y.size(0))).unsqueeze(1).to(y.device)
|
103 |
+
s = torch.gather(out, 1, idx.unsqueeze(2).expand(-1, -1, out.size(2))).squeeze(1)
|
104 |
+
return s
|
105 |
+
|
106 |
+
class StyleEncoder(nn.Module):
|
107 |
+
def __init__(self, img_size=256, style_dim=64, num_domains=3, max_conv_dim=512):
|
108 |
+
super().__init__()
|
109 |
+
dim_in = 64
|
110 |
+
blocks = []
|
111 |
+
blocks += [nn.Conv2d(3, dim_in, 3, 1, 1)]
|
112 |
+
repeat_num = int(np.log2(img_size)) - 2
|
113 |
+
for _ in range(repeat_num):
|
114 |
+
dim_out = min(dim_in*2, max_conv_dim)
|
115 |
+
blocks += [ResBlk(dim_in, dim_out, downsample=True)]
|
116 |
+
dim_in = dim_out
|
117 |
+
self.shared = nn.Sequential(*blocks)
|
118 |
+
self.unshared = nn.ModuleList()
|
119 |
+
for _ in range(num_domains):
|
120 |
+
self.unshared += [nn.Linear(dim_in * (img_size // (2**repeat_num))**2, style_dim)]
|
121 |
+
|
122 |
+
def forward(self, x, y):
|
123 |
+
h = self.shared(x)
|
124 |
+
h = h.view(h.size(0), -1)
|
125 |
+
out = []
|
126 |
+
for layer in self.unshared:
|
127 |
+
out += [layer(h)]
|
128 |
+
out = torch.stack(out, dim=1) # (batch, num_domains, style_dim)
|
129 |
+
idx = torch.LongTensor(range(y.size(0))).unsqueeze(1).to(y.device)
|
130 |
+
s = torch.gather(out, 1, idx.unsqueeze(2).expand(-1, -1, out.size(2))).squeeze(1)
|
131 |
+
return s
|
132 |
+
|
133 |
+
# DEFINICI脫N DEL GENERADOR
|
134 |
+
class Generator(nn.Module):
|
135 |
+
def __init__(self, img_size=256, style_dim=64, max_conv_dim=512):
|
136 |
+
super().__init__()
|
137 |
+
dim_in = 64
|
138 |
+
blocks = []
|
139 |
+
blocks += [nn.Conv2d(3, dim_in, 3, 1, 1)]
|
140 |
+
repeat_num = int(np.log2(img_size)) - 4
|
141 |
+
for _ in range(repeat_num):
|
142 |
+
dim_out = min(dim_in*2, max_conv_dim)
|
143 |
+
blocks += [ResBlk(dim_in, dim_out, normalize=True, downsample=True)]
|
144 |
+
dim_in = dim_out
|
145 |
+
self.encode = nn.Sequential(*blocks)
|
146 |
+
|
147 |
+
self.decode = nn.ModuleList()
|
148 |
+
for _ in range(repeat_num):
|
149 |
+
dim_out = dim_in // 2
|
150 |
+
self.decode += [AdainResBlk(dim_in, dim_out, style_dim, upsample=True)]
|
151 |
+
dim_in = dim_out
|
152 |
+
self.to_rgb = nn.Sequential(
|
153 |
+
nn.InstanceNorm2d(dim_in, affine=True),
|
154 |
+
nn.ReLU(inplace=True),
|
155 |
+
nn.Conv2d(dim_in, 3, 1, 1, 0)
|
156 |
+
)
|
157 |
+
|
158 |
+
def forward(self, x, s):
|
159 |
+
x = self.encode(x)
|
160 |
+
for block in self.decode:
|
161 |
+
x = block(x, s)
|
162 |
+
out = self.to_rgb(x)
|
163 |
+
return out
|
164 |
+
|
165 |
+
# FUNCI脫N PARA CARGAR EL MODELO
|
166 |
+
def load_pretrained_model(ckpt_path, img_size=256, style_dim=64, num_domains=3, device='cpu'):
|
167 |
+
G = Generator(img_size, style_dim).to(device)
|
168 |
+
M = MappingNetwork(16, style_dim, num_domains).to(device) # Suponiendo latent_dim=16
|
169 |
+
S = StyleEncoder(img_size, style_dim, num_domains).to(device)
|
170 |
+
checkpoint = torch.load(ckpt_path, map_location=device)
|
171 |
+
G.load_state_dict(checkpoint['generator'])
|
172 |
+
M.load_state_dict(checkpoint['mapping_network'])
|
173 |
+
S.load_state_dict(checkpoint['style_encoder'])
|
174 |
+
G.eval()
|
175 |
+
S.eval()
|
176 |
+
return G, S
|
177 |
+
|
178 |
+
# FUNCI脫N PARA COMBINAR ESTILOS
|
179 |
+
def combine_styles(source_image, reference_image, generator, style_encoder, target_domain_idx, device='cpu'):
|
180 |
+
transform = transforms.Compose([
|
181 |
+
transforms.Resize((256, 256)), # Ajustar al tama帽o de entrada de tu modelo
|
182 |
+
transforms.ToTensor(),
|
183 |
+
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
|
184 |
+
])
|
185 |
+
|
186 |
+
source_img = transform(source_image).unsqueeze(0).to(device)
|
187 |
+
reference_img = transform(reference_image).unsqueeze(0).to(device)
|
188 |
+
target_domain = torch.tensor([target_domain_idx]).unsqueeze(0).to(device) # Crear un tensor para el dominio objetivo
|
189 |
+
|
190 |
+
with torch.no_grad():
|
191 |
+
style_ref = style_encoder(reference_img, target_domain) # Usar el mismo 铆ndice de dominio que la referencia
|
192 |
+
generated_image = generator(source_img, style_ref)
|
193 |
+
generated_image = (generated_image + 1) / 2.0 # Desnormalizar a [0, 1]
|
194 |
+
generated_image = generated_image.squeeze(0).cpu().permute(1, 2, 0).numpy()
|
195 |
+
generated_image = (generated_image * 255).astype(np.uint8)
|
196 |
+
return Image.fromarray(generated_image)
|
197 |
+
|
198 |
+
# CONFIGURACI脫N DE GRADIO
|
199 |
+
def create_interface(generator, style_encoder, domain_names, device='cpu'):
|
200 |
+
def predict(source_img, ref_img, target_domain):
|
201 |
+
target_domain_idx = domain_names.index(target_domain)
|
202 |
+
return combine_styles(source_img, ref_img, generator, style_encoder, target_domain_idx, device)
|
203 |
+
|
204 |
+
iface = gr.Interface(
|
205 |
+
fn=predict,
|
206 |
+
inputs=[
|
207 |
+
gr.Image(label="Imagen Fuente"),
|
208 |
+
gr.Image(label="Imagen de Referencia"),
|
209 |
+
gr.Dropdown(choices=domain_names, label="Dominio de Referencia (para el estilo)"),
|
210 |
+
],
|
211 |
+
outputs=gr.Image(label="Imagen Generada"),
|
212 |
+
title="AutoStyleGAN - Transferencia de Estilo de Carros",
|
213 |
+
description="Selecciona una imagen de carro fuente y una imagen de carro de referencia para transferir el estilo de la referencia a la fuente."
|
214 |
+
)
|
215 |
+
return iface
|
216 |
+
|
217 |
+
if __name__ == '__main__':
|
218 |
+
#CARGAR EL MODELO ENTRENADO
|
219 |
+
checkpoint_path = '10000_nets_ema.ckpt'
|
220 |
+
img_size = 128
|
221 |
+
style_dim = 64
|
222 |
+
num_domains = 3
|
223 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
224 |
+
|
225 |
+
try:
|
226 |
+
generator, style_encoder = load_pretrained_model(checkpoint_path, img_size, style_dim, num_domains, device)
|
227 |
+
print("Modelo cargado exitosamente.")
|
228 |
+
|
229 |
+
#DEFINIR LOS NOMBRES DE LOS DOMINIOS
|
230 |
+
domain_names = ["BMW", "Corvette", "Mazda"]
|
231 |
+
|
232 |
+
# CREAR E LANZAR LA INTERFAZ DE GRADIO
|
233 |
+
iface = create_interface(generator, style_encoder, domain_names, device)
|
234 |
+
iface.launch(share=True)
|
235 |
+
|
236 |
+
except FileNotFoundError:
|
237 |
+
print(f"Error: No se encontr贸 el archivo de checkpoint en '{checkpoint_path}'. Aseg煤rate de proporcionar la ruta correcta.")
|
238 |
+
except Exception as e:
|
239 |
+
print(f"Ocurri贸 un error al cargar el modelo: {e}")
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
torch
|
3 |
+
torchvision
|
4 |
+
Pillow
|
5 |
+
numpy
|
6 |
+
huggingface_hub
|