Unified input for resize op #2394

vfdev-5 · 2020-07-06T08:46:10Z

Related to #2292

Description:

Unified resize operation: T.Resize, F.resize and F_t.resize
Added tests

Remark:

Only certain PIL interpolations can be supported
Tests of pixel values for resize between PIL and Tensors are done for interpolation > nearest as MAE.

…-6/issue-2292-resize

codecov · 2020-07-06T09:04:41Z

Codecov Report

Merging #2394 into master will decrease coverage by 0.15%.
The diff coverage is 66.66%.

@@            Coverage Diff             @@
##           master    #2394      +/-   ##
==========================================
- Coverage   70.69%   70.53%   -0.16%     
==========================================
  Files          94       94              
  Lines        7839     7895      +56     
  Branches     1221     1241      +20     
==========================================
+ Hits         5542     5569      +27     
- Misses       1925     1943      +18     
- Partials      372      383      +11

Impacted Files	Coverage Δ
torchvision/transforms/transforms.py	`76.02% <50.00%> (-0.53%)`	⬇️
torchvision/transforms/functional_tensor.py	`64.58% <64.44%> (-0.04%)`	⬇️
torchvision/transforms/functional_pil.py	`64.33% <71.42%> (+0.92%)`	⬆️
torchvision/transforms/functional.py	`79.82% <100.00%> (+0.93%)`	⬆️
torchvision/extension.py	`48.64% <0.00%> (-21.63%)`	⬇️
torchvision/models/detection/rpn.py	`92.70% <0.00%> (-0.43%)`	⬇️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 75f5b57...8f48a02. Read the comment docs.

fmassa

Thanks a lot for the PR!

I have a couple of minor comments, otherwise the PR looks good to merge

fmassa · 2020-07-06T11:43:54Z

torchvision/transforms/functional_tensor.py

@@ -1,3 +1,5 @@
+from PIL.Image import NEAREST, BILINEAR, BICUBIC


nit: this doesn't seem to be used?

Correct. Will remove it

fmassa · 2020-07-06T11:47:02Z

torchvision/transforms/functional_tensor.py

+                         "{} element tuple/list".format(len(size)))
+
+    if interpolation not in [0, 1, 2, 3, 4]:
+        raise ValueError("Interpolation mode should be either constant, edge, reflect or symmetric")


error message seems off?

fmassa · 2020-07-06T11:47:11Z

torchvision/transforms/functional_tensor.py

+        size = list(size)
+
+    if isinstance(size, list) and len(size) not in [1, 2]:
+        raise ValueError("Padding must be an int or a 1 or 2 element tuple/list, not a " +


error message seems off?

fmassa · 2020-07-06T12:02:10Z

test/test_functional_tensor.py

+                        # E.g. resized_tensor  = [[a, a, b, c, d, d, e, ...]]
+                        # E.g. resized_pil_img = [[a, b, c, c, d, e, f, ...]]


It's a pity that the behavior of nearest interpolate is different between implementations, I would say it could be worth opening an issue in PyTorch to mention this. I also believe that PIL and OpenCV are consistent, which would make for a case to maybe change the implementation in PyTorch to make this more consistent.

Let me check that between PIL and OpenCV and then we decide about PyTorch.

fmassa · 2020-07-06T12:05:31Z

torchvision/transforms/functional_tensor.py

+        raise ValueError("Padding must be an int or a 1 or 2 element tuple/list, not a " +
+                         "{} element tuple/list".format(len(size)))
+
+    if interpolation not in [0, 1, 2, 3, 4]:


Looks like we don't support mode 1 in the _interpolation_modes?

fmassa

Thanks a lot!

Let's merge this now, and then discuss if we should open an issue in PyTorch to propose to change the behavior of nearest interpolate.

fmassa · 2020-07-06T12:50:36Z

torchvision/transforms/functional_tensor.py

+
+    if need_cast:
+        if mode == "bicubic":
+            img = img.clamp(min=0, max=255)


nit: for the future, we might want to change the 255 value to be dtype-dependent. But this can be done in a future, maybe using convert_image_dtype before and after interpolate is called.

* [WIP] F.resize with tensor * Adapted T.Resize and F.resize with a test * According to the review, fixed copy-pasted messages and unused imports

vfdev-5 added 2 commits July 3, 2020 14:31

[WIP] F.resize with tensor

f5a9776

Merge branch 'master' of https://github.com/pytorch/vision into vfdev…

7936258

…-6/issue-2292-resize

vfdev-5 mentioned this pull request Jul 6, 2020

Unify Tensor and PIL transforms #2292

Closed

16 tasks

Adapted T.Resize and F.resize with a test

965ad6b

vfdev-5 force-pushed the vfdev-5/issue-2292-resize branch from d69e4b1 to 965ad6b Compare July 6, 2020 09:31

vfdev-5 mentioned this pull request Jul 6, 2020

Unified input for resized crop op #2396

Merged

fmassa reviewed Jul 6, 2020

View reviewed changes

According to the review, fixed copy-pasted messages and unused imports

8f48a02

fmassa approved these changes Jul 6, 2020

View reviewed changes

fmassa merged commit e212cc8 into pytorch:master Jul 6, 2020

vfdev-5 deleted the vfdev-5/issue-2292-resize branch July 6, 2020 13:14

vfdev-5 mentioned this pull request Oct 22, 2020

Scriptable Resize Added #1666

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Unified input for resize op #2394

Unified input for resize op #2394

vfdev-5 commented Jul 6, 2020 •

edited

Loading

codecov bot commented Jul 6, 2020 •

edited

Loading

fmassa left a comment

fmassa Jul 6, 2020

vfdev-5 Jul 6, 2020

fmassa Jul 6, 2020

fmassa Jul 6, 2020

vfdev-5 Jul 6, 2020

fmassa Jul 6, 2020

vfdev-5 Jul 6, 2020

fmassa Jul 6, 2020

fmassa left a comment

fmassa Jul 6, 2020

		@@ -1,3 +1,5 @@
		from PIL.Image import NEAREST, BILINEAR, BICUBIC

		# E.g. resized_tensor = [[a, a, b, c, d, d, e, ...]]
		# E.g. resized_pil_img = [[a, b, c, c, d, e, f, ...]]

Unified input for resize op #2394

Unified input for resize op #2394

Conversation

vfdev-5 commented Jul 6, 2020 • edited Loading

codecov bot commented Jul 6, 2020 • edited Loading

Codecov Report

fmassa left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

fmassa left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vfdev-5 commented Jul 6, 2020 •

edited

Loading

codecov bot commented Jul 6, 2020 •

edited

Loading