src/aiunn/finetune.py

@@ -60,7 +60,33 @@ class aiuNNDataset(torch.utils.data.Dataset):
             'high_res': augmented_high['image']
         }
 
-def finetune_model(model: AIIA, datasets: list[str], batch_size=1, epochs=10, accumulation_steps=8, use_checkpoint=False):
+class Upscaler(nn.Module):
+    """
+    Wraps the base model to perform upsampling and a final convolution.
+    The base model produces a feature map of size 512x512.
+    We then upsample by a factor of 2 (to get 1024x1024)
+    and use a convolution to map the hidden features to 3 output channels.
+    """
+    def __init__(self, base_model: AIIABase):
+        super(Upscaler, self).__init__()
+        self.base_model = base_model
+        self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False)
+        self.final_conv = nn.Conv2d(
+            base_model.config.hidden_size,
+            base_model.config.num_channels,
+            kernel_size=3,
+            padding=1
+        )
+        
+    def forward(self, x):
+        # Get the feature maps from the base model (expected shape: [B, 512, 512, 512])
+        features = self.base_model(x)
+        # Upsample the features to match high resolution (1024x1024)
+        upsampled = self.upsample(features)
+        # Convert from hidden features to output channels
+        return self.final_conv(upsampled)
+
+def finetune_model(model: nn.Module, datasets: list[str], batch_size=1, epochs=10, accumulation_steps=8, use_checkpoint=False):
     # Load and concatenate datasets.
     loaded_datasets = [aiuNNDataset(d) for d in datasets]
     combined_dataset = torch.utils.data.ConcatDataset(loaded_datasets)
@@ -93,7 +119,7 @@ def finetune_model(model: AIIA, datasets: list[str], batch_size=1, epochs=10, ac
     
     model = model.to(device)
     criterion = nn.MSELoss()
-    optimizer = torch.optim.Adam(model.parameters(), lr=model.config.learning_rate)
+    optimizer = torch.optim.Adam(model.parameters(), lr=model.base_model.config.learning_rate)
     scaler = GradScaler()
     best_val_loss = float('inf')
     
@@ -108,10 +134,11 @@ def finetune_model(model: AIIA, datasets: list[str], batch_size=1, epochs=10, ac
             high_res = batch['high_res'].to(device)
             with autocast(device_type="cuda"):
                 if use_checkpoint:
-                    outputs = checkpoint(lambda x: model(x), low_res)
+                    # Use checkpointing if requested.
+                    features = checkpoint(lambda x: model(x), low_res)
                 else:
-                    outputs = model(low_res)
+                    features = model(low_res)
-                loss = criterion(outputs, high_res) / accumulation_steps
+                loss = criterion(features, high_res) / accumulation_steps
             scaler.scale(loss).backward()
             train_loss += loss.item() * accumulation_steps
             if i % accumulation_steps == 0:
@@ -142,7 +169,7 @@ def finetune_model(model: AIIA, datasets: list[str], batch_size=1, epochs=10, ac
         print(f"Epoch {epoch+1}, Validation Loss: {avg_val_loss:.4f}")
         if avg_val_loss < best_val_loss:
             best_val_loss = avg_val_loss
-            model.save("best_model")
+            model.base_model.save("best_model")
     return model
 
 def main():
@@ -150,19 +177,14 @@ def main():
     ACCUMULATION_STEPS = 8
     USE_CHECKPOINT = True
 
-    model = AIIABase.load("/root/vision/AIIA/AIIA-base-512")
+    # Load the base model using the config values (hidden_size=512, num_channels=3, etc.)
+    base_model = AIIABase.load("/root/vision/AIIA/AIIA-base-512")
     
-    # Upsample output if a 'chunked_' attribute exists, ensuring spatial dimensions match the high resolution images.
+    # Wrap the base model in our Upscaler so that the output is upsampled to 1024x1024
-    if hasattr(model, 'chunked_'):
+    model = Upscaler(base_model)
-        model.add_module('final_upsample', nn.Upsample(scale_factor=2, mode='bilinear'))
     
-    # Append a final convolutional layer using values from model.config.
+    print("Modified model architecture with upsampling wrapper:")
-    # This converts the hidden feature maps (512 channels) to the 3 required output channels.
+    print(base_model.config)
-    final_conv = nn.Conv2d(model.config.hidden_size, model.config.num_channels, kernel_size=3, padding=1)
-    model.add_module('final_layer', final_conv)
-    
-    print("Modified model architecture:")
-    print(model.config)
     
     finetune_model(
         model=model,