Merge branch 'master' of github.com:pytorch/vision into restore_cache

Author: NicolasHug

CONTRIBUTING.md

@@ -33,7 +33,7 @@ clear and has sufficient instructions to be able to reproduce the issue.
 ### Install PyTorch Nightly 
 
 ```bash
-conda install pytorch -c pytorch-nightly
+conda install pytorch -c pytorch-nightly -c conda-forge
 # or with pip (see https://pytorch.org/get-started/locally/)
 # pip install numpy
 # pip install --pre torch -f https://download.pytorch.org/whl/nightly/cu102/torch_nightly.html

test/test_transforms.py

@@ -620,6 +620,20 @@ def test_to_tensor(self):
         output = trans(img)
         self.assertTrue(np.allclose(input_data.numpy(), output.numpy()))
 
+    def test_to_tensor_with_other_default_dtypes(self):
+        current_def_dtype = torch.get_default_dtype()
+
+        t = transforms.ToTensor()
+        np_arr = np.random.randint(0, 255, (32, 32, 3), dtype=np.uint8)
+        img = Image.fromarray(np_arr)
+
+        for dtype in [torch.float16, torch.float, torch.double]:
+            torch.set_default_dtype(dtype)
+            res = t(img)
+            self.assertTrue(res.dtype == dtype, msg=f"{res.dtype} vs {dtype}")
+
+        torch.set_default_dtype(current_def_dtype)
+
     def test_max_value(self):
         for dtype in int_dtypes():
             self.assertEqual(F_t._max_value(dtype), torch.iinfo(dtype).max)

torchvision/models/detection/backbone_utils.py

@@ -96,7 +96,8 @@ def resnet_fpn_backbone(
     # select layers that wont be frozen
     assert 0 <= trainable_layers <= 5
     layers_to_train = ['layer4', 'layer3', 'layer2', 'layer1', 'conv1'][:trainable_layers]
-    # freeze layers only if pretrained backbone is used
+    if trainable_layers == 5:
+        layers_to_train.append('bn1')
     for name, parameter in backbone.named_parameters():
         if all([not name.startswith(layer) for layer in layers_to_train]):
             parameter.requires_grad_(False)
@@ -152,7 +153,6 @@ def mobilenet_backbone(
     assert 0 <= trainable_layers <= num_stages
     freeze_before = num_stages if trainable_layers == 0 else stage_indices[num_stages - trainable_layers]
 
-    # freeze layers only if pretrained backbone is used
     for b in backbone[:freeze_before]:
         for parameter in b.parameters():
             parameter.requires_grad_(False)

torchvision/transforms/functional.py

@@ -104,6 +104,8 @@ def to_tensor(pic):
     if _is_numpy(pic) and not _is_numpy_image(pic):
         raise ValueError('pic should be 2/3 dimensional. Got {} dimensions.'.format(pic.ndim))
 
+    default_float_dtype = torch.get_default_dtype()
+
     if isinstance(pic, np.ndarray):
         # handle numpy array
         if pic.ndim == 2:
@@ -112,12 +114,12 @@ def to_tensor(pic):
         img = torch.from_numpy(pic.transpose((2, 0, 1))).contiguous()
         # backward compatibility
         if isinstance(img, torch.ByteTensor):
-            return img.float().div(255)
+            return img.to(dtype=default_float_dtype).div(255)
         else:
             return img
 
     if accimage is not None and isinstance(pic, accimage.Image):
-        nppic = np.zeros([pic.channels, pic.height, pic.width], dtype=np.float32)
+        nppic = np.zeros([pic.channels, pic.height, pic.width], dtype=default_float_dtype)
         pic.copyto(nppic)
         return torch.from_numpy(nppic)
 
@@ -137,7 +139,7 @@ def to_tensor(pic):
     # put it from HWC to CHW format
     img = img.permute((2, 0, 1)).contiguous()
     if isinstance(img, torch.ByteTensor):
-        return img.float().div(255)
+        return img.to(dtype=default_float_dtype).div(255)
     else:
         return img