Add detach to the tensors in handlers

monai/handlers/classification_saver.py

@@ -104,6 +104,8 @@ def __call__(self, engine: Engine) -> None:
             self._filenames.extend(filenames)
         outputs = self.output_transform(engine.state.output)
         if outputs is not None:
+            if isinstance(outputs, torch.Tensor):
+                outputs = outputs.detach()
             self._outputs.append(outputs)
 
     def _finalize(self, engine: Engine) -> None:

monai/handlers/iteration_metric.py

@@ -77,6 +77,10 @@ def update(self, output: Sequence[torch.Tensor]) -> None:
         y_pred, y = output
 
         def _compute(y_pred, y):
+            if isinstance(y_pred, torch.Tensor):
+                y_pred = y_pred.detach()
+            if isinstance(y, torch.Tensor):
+                y = y.detach()
             score = self.metric_fn(y_pred, y)
             return score[0] if isinstance(score, (tuple, list)) else score
 

monai/handlers/transform_inverter.py

@@ -130,8 +130,11 @@ def __call__(self, engine: Engine) -> None:
                     align_corners=None,
                 )
 
+            output = engine.state.output[output_key]
+            if isinstance(output, torch.Tensor):
+                output = output.detach()
             segs_dict = {
-                batch_key: engine.state.output[output_key],
+                batch_key: output,
                 transform_key: transform_info,
             }
             meta_dict_key = f"{batch_key}_{self.meta_key_postfix}"

tests/test_lmdbdataset.py

@@ -187,7 +187,7 @@ def setUp(self):
     def tearDown(self):
         shutil.rmtree(self.tempdir)
 
-    @DistCall(nnodes=1, nproc_per_node=2)
+    @DistCall(nnodes=1, nproc_per_node=1)
     def test_mp_cache(self):
         items = [[list(range(i))] for i in range(5)]