[Fix] Fix image dtype when enable_normalize=False. (#301)

* [Fix] Fix image dtype when enable_normalize=False.

* update ut

* move to collate

* update ut

mmengine/model/base_model/data_preprocessor.py

@@ -53,7 +53,7 @@ class BaseDataPreprocessor(nn.Module):
             Tuple[List[torch.Tensor], Optional[list]]: Unstacked list of input
             tensor and list of labels at target device.
         """
-        inputs = [_data['inputs'].to(self._device) for _data in data]
+        inputs = [_data['inputs'].to(self._device).float() for _data in data]
         batch_data_samples: List[BaseDataElement] = []
         # Model can get predictions without any data samples.
         for _data in data:

tests/test_model/test_base_model/test_data_preprocessor.py

@@ -28,6 +28,7 @@ class TestBaseDataPreprocessor(TestCase):
         ]
 
         batch_inputs, batch_labels = base_data_preprocessor(data)
+        self.assertTrue(torch.is_floating_point(batch_inputs))
         self.assertEqual(batch_inputs.shape, (2, 1, 3, 5))
 
         assert_allclose(input1, batch_inputs[0])
@@ -38,14 +39,17 @@ class TestBaseDataPreprocessor(TestCase):
         if torch.cuda.is_available():
             base_data_preprocessor = base_data_preprocessor.cuda()
             batch_inputs, batch_labels = base_data_preprocessor(data)
+            self.assertTrue(torch.is_floating_point(batch_inputs))
             self.assertEqual(batch_inputs.device.type, 'cuda')
 
             base_data_preprocessor = base_data_preprocessor.cpu()
             batch_inputs, batch_labels = base_data_preprocessor(data)
+            self.assertTrue(torch.is_floating_point(batch_inputs))
             self.assertEqual(batch_inputs.device.type, 'cpu')
 
             base_data_preprocessor = base_data_preprocessor.to('cuda:0')
             batch_inputs, batch_labels = base_data_preprocessor(data)
+            self.assertTrue(torch.is_floating_point(batch_inputs))
             self.assertEqual(batch_inputs.device.type, 'cuda')
 
 
@@ -122,6 +126,7 @@ class TestImageDataPreprocessor(TestBaseDataPreprocessor):
 
         target_inputs = [target_inputs1, target_inputs2]
         inputs, data_samples = data_preprocessor(data, True)
+        self.assertTrue(torch.is_floating_point(inputs))
 
         target_data_samples = [data_sample1, data_sample2]
         for input_, data_sample, target_input, target_data_sample in zip(
@@ -142,6 +147,7 @@ class TestImageDataPreprocessor(TestBaseDataPreprocessor):
 
         target_inputs = [target_inputs1, target_inputs2]
         inputs, data_samples = data_preprocessor(data, True)
+        self.assertTrue(torch.is_floating_point(inputs))
 
         target_data_samples = [data_sample1, data_sample2]
         for input_, data_sample, target_input, target_data_sample in zip(