Update app.py
Browse files
app.py
CHANGED
|
@@ -10,9 +10,8 @@ import torch.nn as nn
|
|
| 10 |
# 1. DEFINICIÓN DEL MODELO VAE
|
| 11 |
# -----------------------------
|
| 12 |
class VAE(nn.Module):
|
| 13 |
-
def __init__(self, input_dim, h_dim=400, z_dim=
|
| 14 |
super().__init__()
|
| 15 |
-
self.z_dim = z_dim
|
| 16 |
# Encoder
|
| 17 |
self.img_2hid = nn.Linear(input_dim, h_dim)
|
| 18 |
self.hid_2mu = nn.Linear(h_dim, z_dim)
|
|
@@ -26,8 +25,7 @@ class VAE(nn.Module):
|
|
| 26 |
|
| 27 |
def encode(self, x):
|
| 28 |
h = self.relu(self.img_2hid(x))
|
| 29 |
-
mu = self.hid_2mu(h)
|
| 30 |
-
sigma = self.hid_2sigma(h)
|
| 31 |
return mu, sigma
|
| 32 |
|
| 33 |
def decode(self, z):
|
|
@@ -37,47 +35,48 @@ class VAE(nn.Module):
|
|
| 37 |
def forward(self, x):
|
| 38 |
mu, sigma = self.encode(x)
|
| 39 |
epsilon = torch.randn_like(sigma)
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
return
|
| 43 |
|
| 44 |
# -----------------------------
|
| 45 |
-
# 2. CARGAR MODELO DESDE HUGGING FACE
|
| 46 |
# -----------------------------
|
| 47 |
-
REPO_ID = "Bmo411/VAE" #
|
| 48 |
MODEL_FILENAME = "vae_complete_model.pth"
|
| 49 |
|
| 50 |
-
# Descargar
|
| 51 |
model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILENAME)
|
| 52 |
|
| 53 |
-
|
|
|
|
|
|
|
| 54 |
|
| 55 |
-
#
|
| 56 |
torch.serialization.add_safe_globals({"VAE": VAE})
|
| 57 |
|
| 58 |
-
# Cargar
|
|
|
|
| 59 |
model = torch.load(model_path, map_location=device, weights_only=False)
|
| 60 |
model.to(device)
|
| 61 |
model.eval()
|
| 62 |
|
|
|
|
|
|
|
|
|
|
| 63 |
# -----------------------------
|
| 64 |
-
# 3. GENERAR IMAGEN
|
| 65 |
# -----------------------------
|
| 66 |
-
def generate_image(
|
| 67 |
with torch.no_grad():
|
| 68 |
-
# Muestra del espacio latente
|
| 69 |
z = torch.randn(1, z_dim).to(device)
|
| 70 |
-
out = model.decode(z)
|
| 71 |
-
|
| 72 |
-
# Convertir a forma imagen (1, 1, 100, 100)
|
| 73 |
out = out.view(1, 1, 100, 100)
|
| 74 |
|
| 75 |
-
# Guardar imagen temporal
|
| 76 |
output_path = "generated_sample.png"
|
| 77 |
save_image(out, output_path)
|
| 78 |
|
| 79 |
-
|
| 80 |
-
img = Image.open(output_path).convert("L")
|
| 81 |
return img
|
| 82 |
|
| 83 |
# -----------------------------
|
|
@@ -85,10 +84,10 @@ def generate_image(z_dim=40):
|
|
| 85 |
# -----------------------------
|
| 86 |
iface = gr.Interface(
|
| 87 |
fn=generate_image,
|
| 88 |
-
inputs=
|
| 89 |
outputs="image",
|
| 90 |
title="Generador de Imagen con VAE",
|
| 91 |
-
description="Genera una imagen aleatoria
|
| 92 |
)
|
| 93 |
|
| 94 |
iface.launch()
|
|
|
|
| 10 |
# 1. DEFINICIÓN DEL MODELO VAE
|
| 11 |
# -----------------------------
|
| 12 |
class VAE(nn.Module):
|
| 13 |
+
def __init__(self, input_dim, h_dim=400, z_dim=20): # NOTA: z_dim por defecto en 20
|
| 14 |
super().__init__()
|
|
|
|
| 15 |
# Encoder
|
| 16 |
self.img_2hid = nn.Linear(input_dim, h_dim)
|
| 17 |
self.hid_2mu = nn.Linear(h_dim, z_dim)
|
|
|
|
| 25 |
|
| 26 |
def encode(self, x):
|
| 27 |
h = self.relu(self.img_2hid(x))
|
| 28 |
+
mu, sigma = self.hid_2mu(h), self.hid_2sigma(h)
|
|
|
|
| 29 |
return mu, sigma
|
| 30 |
|
| 31 |
def decode(self, z):
|
|
|
|
| 35 |
def forward(self, x):
|
| 36 |
mu, sigma = self.encode(x)
|
| 37 |
epsilon = torch.randn_like(sigma)
|
| 38 |
+
z_reparametrized = mu + sigma * epsilon
|
| 39 |
+
x_reconstructed = self.decode(z_reparametrized)
|
| 40 |
+
return x_reconstructed, mu, sigma
|
| 41 |
|
| 42 |
# -----------------------------
|
| 43 |
+
# 2. CARGAR EL MODELO DESDE HUGGING FACE
|
| 44 |
# -----------------------------
|
| 45 |
+
REPO_ID = "Bmo411/VAE" # <-- reemplaza con tu repo si cambia
|
| 46 |
MODEL_FILENAME = "vae_complete_model.pth"
|
| 47 |
|
| 48 |
+
# Descargar modelo automáticamente
|
| 49 |
model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILENAME)
|
| 50 |
|
| 51 |
+
# Inicializar arquitectura del modelo
|
| 52 |
+
input_dim = 100 * 100
|
| 53 |
+
dummy_model = VAE(input_dim=input_dim, z_dim=20) # la arquitectura base es necesaria para cargar pesos
|
| 54 |
|
| 55 |
+
# Permitir deserialización segura
|
| 56 |
torch.serialization.add_safe_globals({"VAE": VAE})
|
| 57 |
|
| 58 |
+
# Cargar modelo completo (no solo pesos)
|
| 59 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 60 |
model = torch.load(model_path, map_location=device, weights_only=False)
|
| 61 |
model.to(device)
|
| 62 |
model.eval()
|
| 63 |
|
| 64 |
+
# Detectar z_dim automáticamente desde el decoder
|
| 65 |
+
z_dim = model.z_2hid.in_features
|
| 66 |
+
|
| 67 |
# -----------------------------
|
| 68 |
+
# 3. FUNCIÓN PARA GENERAR IMAGEN
|
| 69 |
# -----------------------------
|
| 70 |
+
def generate_image():
|
| 71 |
with torch.no_grad():
|
|
|
|
| 72 |
z = torch.randn(1, z_dim).to(device)
|
| 73 |
+
out = model.decode(z)
|
|
|
|
|
|
|
| 74 |
out = out.view(1, 1, 100, 100)
|
| 75 |
|
|
|
|
| 76 |
output_path = "generated_sample.png"
|
| 77 |
save_image(out, output_path)
|
| 78 |
|
| 79 |
+
img = Image.open(output_path).convert("L") # Convertir a escala de grises
|
|
|
|
| 80 |
return img
|
| 81 |
|
| 82 |
# -----------------------------
|
|
|
|
| 84 |
# -----------------------------
|
| 85 |
iface = gr.Interface(
|
| 86 |
fn=generate_image,
|
| 87 |
+
inputs=[],
|
| 88 |
outputs="image",
|
| 89 |
title="Generador de Imagen con VAE",
|
| 90 |
+
description=f"Genera una imagen aleatoria desde el VAE entrenado. Dimensión latente del modelo detectada: {z_dim}"
|
| 91 |
)
|
| 92 |
|
| 93 |
iface.launch()
|