Scriptable Resize Added #1666
Scriptable Resize Added #1666
Conversation
Codecov Report
@@ Coverage Diff @@
## master #1666 +/- ##
==========================================
- Coverage 66.02% 0% -66.03%
==========================================
Files 92 92
Lines 7331 7348 +17
Branches 1107 1111 +4
==========================================
- Hits 4840 0 -4840
- Misses 2175 7348 +5173
+ Partials 316 0 -316
Continue to review full report at Codecov.
|
|
@fmassa could you review this please. |
| else: | ||
| out_img = Fn.interpolate(img.unsqueeze(0), size=size, mode=interpolation) | ||
|
|
||
| return(out_img.clamp(min=0, max=255).squeeze(0)) |
There was a problem hiding this comment.
I'm not sure we should be clamping the input to 0-255 in the function. Although this makes it match the behavior of the Pillow implementation, this only works if the input is in 0-255, which is rarely the case for floating point values.
My take is that we should remove the clamp in this function.
Also, can you remove the parenthesis and add a space after the return?
| if w < h: | ||
| ow = size | ||
| oh = int(size * h / w) | ||
| out_img = Fn.interpolate(img.unsqueeze(0), size=(oh, ow), mode=interpolation) |
There was a problem hiding this comment.
This assumes that the input is not batched (but it could potentially be, like for videos).
Instead, I think we should unsqueeze only if it's a 3d tensor, and in this case we should squeeze back.
| ow = int(size * w / h) | ||
| out_img = Fn.interpolate(img.unsqueeze(0), size=(oh, ow), mode=interpolation) | ||
| else: | ||
| out_img = Fn.interpolate(img.unsqueeze(0), size=size, mode=interpolation) |
There was a problem hiding this comment.
This is fine as is, but I think we can simplify a bit the code so that we only need to call interpolate in a single codepath.
A solution would be to create the size in the int case to be a tuple. Something like
if isinstance(size, int):
w, h = img.shape[2], img.shape[1]
if (w <= h and w == size) or (h <= w and h == size):
return img
if w < h:
ow = size
oh = int(size * h / w)
size = (oh, ow)
else:
oh = size
ow = int(size * w / h)
size = (oh, ow)
out_img = Fn.interpolate(..., size=size, mode=interpolation)| @@ -219,3 +220,40 @@ def ten_crop(img, size, vertical_flip=False): | |||
| def _blend(img1, img2, ratio): | |||
| bound = 1 if img1.dtype.is_floating_point else 255 | |||
| return (ratio * img1 + (1 - ratio) * img2).clamp(0, bound).to(img1.dtype) | |||
|
|
|||
|
|
|||
| def resize(img, size, interpolation="bilinear"): | |||
There was a problem hiding this comment.
This is something that I still need to figure out: should we use the old PIL-based interpolation with int, or the torch-based one? I'm not sure.
| @@ -138,6 +138,36 @@ def test_ten_crop(self): | |||
| (transforms.ToTensor()(cropped_pil_image[9]) * 255).to(torch.uint8))) | |||
| self.assertTrue(torch.equal(img_tensor, img_tensor_clone)) | |||
|
|
|||
| def test_resize(self): | |||
There was a problem hiding this comment.
Can you also add a test for scriptability?
|
@fmassa, I am getting error while scripting nn.functional.interpolate `RuntimeError: isIntList() INTERNAL ASSERT FAILED at /pytorch/aten/src/ATen/core/ivalue_inl.h:544, please report a bug to PyTorch. Expected IntList but got DoubleList @torch.jit.script |
|
@surgan12 you might just need to specify that |
|
@fmassa I am using 1.3.0. |
|
having both size and |
|
Sure @fmassa , will just update and get back. |
|
This PR can be closed as the feature is already implemented (#2394) |
This PR adds torchscriptable resize as a part of #1375
cc : @fmassa