Fix #5 app.py

app.py

@@ -13,70 +13,73 @@ from huggingface_hub import hf_hub_download
 class ResBlk(nn.Module):
     def __init__(self, dim_in, dim_out, normalize=False, downsample=False):
         super().__init__()
-        self.conv1 = nn.Conv2d(dim_in, dim_out, 3, 1, 1)
-        self.norm1 = nn.InstanceNorm2d(dim_out, affine=True) if normalize else None
-        self.relu1 = nn.ReLU(inplace=True)
-        self.conv2 = nn.Conv2d(dim_out, dim_out, 3, 1, 1)
-        self.norm2 = nn.InstanceNorm2d(dim_out, affine=True) if normalize else None
-        self.relu2 = nn.ReLU(inplace=True)
+        self.normalize = normalize
         self.downsample = downsample
-        if self.downsample:
-            self.avg_pool = nn.AvgPool2d(2)
+        self.main = nn.Sequential(
+            nn.Conv2d(dim_in, dim_out, 3, 1, 1),
+            nn.InstanceNorm2d(dim_out, affine=True) if normalize else nn.Identity(),
+            nn.ReLU(inplace=True),
+            nn.Conv2d(dim_out, dim_out, 3, 1, 1),
+            nn.InstanceNorm2d(dim_out, affine=True) if normalize else nn.Identity()
+        )
+        self.downsample_layer = nn.AvgPool2d(2) if downsample else nn.Identity()
+        self.skip = nn.Conv2d(dim_in, dim_out, 1, 1, 0, bias=False)
 
     def forward(self, x):
-        residual = x
-        out = self.conv1(x)
-        if self.norm1:
-            out = self.norm1(out)
-        out = self.relu1(out)
-        out = self.conv2(out)
-        if self.norm2:
-            out = self.norm2(out)
-        out = self.relu2(out)
-        if self.downsample:
-            out = self.avg_pool(out)
-            residual = self.avg_pool(residual)
-        out = out + residual
-        return out
+        out = self.main(x)
+        out = self.downsample_layer(out)
+        skip = self.skip(x)
+        skip = self.downsample_layer(skip)
+        return (out + skip) / math.sqrt(2)
 
 class AdainResBlk(nn.Module):
-    def __init__(self, dim_in, dim_out, style_dim, upsample=False):
+    def __init__(self, dim_in, dim_out, style_dim=64, w_hpf=1, upsample=False):
         super().__init__()
+        self.upsample = upsample
+        self.w_hpf = w_hpf
+
+        self.norm1 = AdaIN(dim_in, style_dim)
+        self.norm2 = AdaIN(dim_out, style_dim)
+        self.actv = nn.LeakyReLU(0.2)
         self.conv1 = nn.Conv2d(dim_in, dim_out, 3, 1, 1)
-        self.norm1 = AdaIN(dim_out, style_dim)
-        self.relu1 = nn.ReLU(inplace=True)
         self.conv2 = nn.Conv2d(dim_out, dim_out, 3, 1, 1)
-        self.norm2 = AdaIN(dim_out, style_dim)
-        self.relu2 = nn.ReLU(inplace=True)
-        self.upsample = upsample
+
+        if dim_in != dim_out:
+            self.skip = nn.Conv2d(dim_in, dim_out, 1, 1, 0)
+        else:
+            self.skip = nn.Identity()
 
     def forward(self, x, s):
-        residual = x
-        if self.upsample:
-            residual = F.interpolate(residual, scale_factor=2, mode='nearest')
-        out = self.conv1(x)
-        out = self.norm1(out, s)
-        out = self.relu1(out)
+        x_orig = x
+
         if self.upsample:
-            out = F.interpolate(out, scale_factor=2, mode='nearest')
-        out = self.conv2(out)
-        out = self.norm2(out, s)
-        out = self.relu2(out)
-        out = out + residual
+            x = F.interpolate(x, scale_factor=2, mode='nearest')
+            x_orig = F.interpolate(x_orig, scale_factor=2, mode='nearest')
+
+        h = self.norm1(x, s)
+        h = self.actv(h)
+        h = self.conv1(h)
+
+        h = self.norm2(h, s)
+        h = self.actv(h)
+        h = self.conv2(h)
+
+        skip = self.skip(x_orig)
+
+        out = (h + skip) / math.sqrt(2)
         return out
 
 class AdaIN(nn.Module):
     def __init__(self, num_features, style_dim):
-        super().__init__()
-        self.norm = nn.InstanceNorm2d(num_features, affine=False)
+        super(AdaIN, self).__init__()
         self.fc = nn.Linear(style_dim, num_features * 2)
 
     def forward(self, x, s):
         h = self.fc(s)
-        gamma, beta = torch.chunk(h, 2, dim=1)
+        gamma, beta = torch.chunk(h, chunks=2, dim=1)
         gamma = gamma.unsqueeze(2).unsqueeze(3)
         beta = beta.unsqueeze(2).unsqueeze(3)
-        return (1 + gamma) * self.norm(x) + beta
+        return (1 + gamma) * x + beta
 
 class MappingNetwork(nn.Module):
     def __init__(self, latent_dim, style_dim, num_domains):
@@ -145,7 +148,7 @@ class Generator(nn.Module):
         self.encode = nn.Sequential(*blocks)
 
         self.decode = nn.ModuleList()
-        for _ in range(repeat_num):
+        for i in range(repeat_num):
             dim_out = dim_in // 2
             self.decode += [AdainResBlk(dim_in, dim_out, style_dim, upsample=True)]
             dim_in = dim_out
@@ -214,22 +217,23 @@ def create_interface(generator, style_encoder, domain_names, device='cpu'):
     )
     return iface
 
+
 if __name__ == '__main__':
     #CARGAR EL MODELO ENTRENADO
-    checkpoint_path = 'iter/10500_nets_ema.ckpt'
+    checkpoint_path = 'iter/12000_nets_ema.ckpt'
     img_size = 128
-    style_dim = 64 
-    num_domains = 3 
+    style_dim = 64
+    num_domains = 3
     device = 'cuda' if torch.cuda.is_available() else 'cpu'
 
     try:
         generator, style_encoder = load_pretrained_model(checkpoint_path, img_size, style_dim, num_domains, device)
         print("Modelo cargado exitosamente.")
 
-        #DEFINIR LOS NOMBRES DE LOS DOMINIOS
+        # DEFINIR LOS NOMBRES DE LOS DOMINIOS
         domain_names = ["BMW", "Corvette", "Mazda"]
 
-        #  CREAR E LANZAR LA INTERFAZ DE GRADIO 
+        #  CREAR E LANZAR LA INTERFAZ DE GRADIO
         iface = create_interface(generator, style_encoder, domain_names, device)
         iface.launch(share=True)