WIP: save local changes before rebase

Signed-off-by: 林旻佑 <[email protected]>

Author: 林旻佑

monai/inferers/utils.py

@@ -36,14 +36,14 @@
 tqdm, _ = optional_import("tqdm", name="tqdm")
 _nearest_mode = "nearest-exact"
 
-__all__ = ["sliding_window_inference"]
+__all__ = ["ensure_channel_first","sliding_window_inference"]
 
 def ensure_channel_first(
     x: torch.Tensor,
     spatial_ndim: Optional[int] = None,
     channel_hint: Optional[int] = None,
     threshold: int = 32,
-) -> Tuple[torch.Tensor, int]:
+) -> tuple[torch.Tensor, int]:
     """
     Normalize a tensor to channel-first layout (N, C, spatial...).
 

monai/metrics/meandice.py

@@ -309,18 +309,22 @@ def __call__(self, y_pred: torch.Tensor, y: torch.Tensor) -> torch.Tensor | tupl
                 the number of channels is inferred from ``y_pred.shape[1]`` when ``num_classes is None``.
             y: ground truth with shape (batch_size, num_classes or 1, spatial_dims...).
         """
+       
+
         # --- Normalize layout to channel-first (N, C, spatial...) ---
-        n_ch = self.num_classes or y_pred.shape[1]
+        # Prefer a strong signal when available.
+        if self.num_classes is not None:
+            y_pred, _ = ensure_channel_first(y_pred, channel_hint=self.num_classes)
+        else:
+            y_pred, _ = ensure_channel_first(y_pred)
 
-        # Always normalize y_pred with hint
-        y_pred, _ = ensure_channel_first(y_pred, channel_hint=n_ch)
+        # Infer channels after normalization (or use provided).
+        n_ch = self.num_classes or y_pred.shape[1]
 
-        # Normalize y if it looks like channel-last (last dim = 1 or n_ch)
+        # Normalize y if it plausibly is channel-last.
         if y.ndim == y_pred.ndim and y.shape[-1] in (1, n_ch):
             y, _ = ensure_channel_first(y, channel_hint=n_ch)
 
-
-
         _apply_argmax, _threshold = self.apply_argmax, self.threshold
         if self.num_classes is None:
             n_pred_ch = y_pred.shape[1]  # y_pred is in one-hot format or multi-channel scores