Upload train_medsiglip.py with huggingface_hub

train_medsiglip.py

@@ -67,13 +67,14 @@ class Config:
     num_seg_classes: int = 3  # background, symphysis, head
 
     # Training
-    batch_size: int = 8  # Smaller batch for larger model
+    batch_size: int = 4  # Reduced for memory when encoder unfrozen
     num_epochs: int = 30
     learning_rate: float = 5e-5  # Lower LR for fine-tuning
     weight_decay: float = 0.01
     warmup_epochs: int = 2
-    gradient_accumulation: int = 4
+    gradient_accumulation: int = 8  # Increased to maintain effective batch size
     freeze_encoder_epochs: int = 3  # Freeze encoder initially
+    use_gradient_checkpointing: bool = True  # Save memory
 
     # Output
     output_dir: Path = Path("./outputs")
@@ -362,11 +363,20 @@ class LaborViewMedSigLIP(nn.Module):
             param.requires_grad = False
         print("Encoder frozen")
 
-    def unfreeze_encoder(self):
-        """Unfreeze the vision encoder"""
+    def unfreeze_encoder(self, use_gradient_checkpointing=True):
+        """Unfreeze the vision encoder with optional gradient checkpointing"""
         for param in self.vision_encoder.parameters():
             param.requires_grad = True
-        print("Encoder unfrozen")
+
+        # Enable gradient checkpointing to save memory
+        if use_gradient_checkpointing:
+            if hasattr(self.vision_encoder, 'gradient_checkpointing_enable'):
+                self.vision_encoder.gradient_checkpointing_enable()
+                print("Encoder unfrozen with gradient checkpointing")
+            else:
+                print("Encoder unfrozen (gradient checkpointing not available)")
+        else:
+            print("Encoder unfrozen")
 
 
 def train_epoch(model, loader, optimizer, scheduler, scaler, device, config, epoch):
@@ -614,7 +624,9 @@ def main():
     for epoch in range(config.num_epochs):
         # Unfreeze encoder after initial epochs
         if epoch == config.freeze_encoder_epochs:
-            model.unfreeze_encoder()
+            # Clear memory before unfreezing
+            torch.cuda.empty_cache()
+            model.unfreeze_encoder(use_gradient_checkpointing=config.use_gradient_checkpointing)
             # Recreate optimizer with all parameters
             optimizer = AdamW(
                 model.parameters(),