fix just train lucchi++ monai_unet --fast-dev-run

Author: Donglai Wei

.gitignore

@@ -156,3 +156,4 @@ lib/
 .claude/
 .codex/
 .CLAUDE.md
+tmp/

connectomics/config/hydra_config.py

@@ -915,6 +915,7 @@ class TestTimeAugmentationConfig:
     - rotation90_axes: Uses spatial-only indices (e.g., [1, 2] for H-W plane where 0=D, 1=H, 2=W)
     """
 
+    enabled: bool = True  # Enable flip/rotation TTA (preprocessing still applies)
     flip_axes: Any = None  # TTA flip strategy: "all" (8 flips), null (no aug),
     # or list like [[2], [3]] (full tensor indices)
     rotation90_axes: Any = None  # TTA rotation90 strategy: "all" (3 planes × 4 rotations),

connectomics/inference/tta.py

@@ -198,14 +198,16 @@ def predict(self, images: torch.Tensor, mask: Optional[torch.Tensor] = None) ->
         if getattr(self.cfg.data, "do_2d", False) and images.size(2) == 1:
             images = images.squeeze(2)
 
-        # Get TTA configuration
+        # Get TTA configuration (respect enabled flag for augmentations)
         if hasattr(self.cfg, "inference") and hasattr(self.cfg.inference, "test_time_augmentation"):
-            tta_flip_axes_config = getattr(
-                self.cfg.inference.test_time_augmentation, "flip_axes", None
-            )
-            tta_rotation90_axes_config = getattr(
-                self.cfg.inference.test_time_augmentation, "rotation90_axes", None
-            )
+            tta_cfg = self.cfg.inference.test_time_augmentation
+            tta_enabled = getattr(tta_cfg, "enabled", True)
+            if tta_enabled:
+                tta_flip_axes_config = getattr(tta_cfg, "flip_axes", None)
+                tta_rotation90_axes_config = getattr(tta_cfg, "rotation90_axes", None)
+            else:
+                tta_flip_axes_config = None
+                tta_rotation90_axes_config = None
         else:
             tta_flip_axes_config = None
             tta_rotation90_axes_config = None

tutorials/lucchi++.yaml

@@ -172,12 +172,13 @@ optimization:
 
   # Scheduler - ReduceLROnPlateau for adaptive learning
   scheduler:
-    name: ReduceLROnPlateau           # Reduce LR when validation loss plateaus
+    name: ReduceLROnPlateau           # Reduce LR when training loss plateaus
     mode: min                         # Monitor minimum loss
     factor: 0.5                       # Reduce LR by 50%
     patience: 50                      # Wait 50 epochs before reducing
     threshold: 1.0e-4                 # Minimum change to qualify as improvement
     min_lr: 1.0e-6                    # Don't go below 1e-6
+    monitor: train_loss_total_epoch   # No validation; monitor training loss
 
 monitor:
   # Loss monitoring and validation frequency