Change import FFT according to interface of torch #246
Conversation
Codecov Report
@@ Coverage Diff @@
## master #246 +/- ##
==========================================
+ Coverage 95.66% 95.68% +0.02%
==========================================
Files 29 29
Lines 1798 1807 +9
==========================================
+ Hits 1720 1729 +9
Misses 78 78
Flags with carried forward coverage won't be shown. Click here to find out more.
|
|
In order to resolve blocked merging, the protection of master branch should be changed (required tests). I suggest to do it after review. |
|
Kudos, SonarCloud Quality Gate passed!
|
| @@ -53,7 +53,7 @@ def brisque(x: torch.Tensor, | |||
| assert kernel_size % 2 == 1, f'Kernel size must be odd, got [{kernel_size}]' | |||
| _validate_input([x, ], dim_range=(4, 4), data_range=(0, data_range)) | |||
|
|
|||
| x = x / data_range * 255 | |||
| x = x / float(data_range) * 255 | |||
There was a problem hiding this comment.
Why is this explicit conversion necessary?
There was a problem hiding this comment.
In case of x and data_range being torch.uint8 and int, the result will be Byte for torch==1.5.1 due to classic division instead of the true_division. It will lead to zeroing the input and possible exceptions from pytorch layers, which could be incompatible with torch.Byte.
There was a problem hiding this comment.
Agh, I see. This is unfortunate because from one hand we don't check user's inputs but from the other we require data_range to be int of float. I think this is a good compromise though.
Closes #245
Proposed Changes
torchtorch=={'1.5.1', '1.8.1'}, torchvision=={'0.6.1', '0,9,1'})torch>=1.5.1