Fast rotation for right angles #8295

gau-nernst · 2024-03-05T01:43:35Z

Fixes #8281

cc @vfdev-5

pytorch-bot · 2024-03-05T01:43:38Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8295

📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures

As of commit 4260a3b with merge base c7bcfad ():

NEW FAILURES - The following jobs have failed:

Tests / unittests-linux (3.11, linux.12xlarge, cpu) / linux-job (gh)
test/test_internet.py::TestDatasetUtils::test_download_url_dispatch_download_from_google_drive[False]
Tests / unittests-linux (3.9, linux.12xlarge, cpu) / linux-job (gh)
test/test_internet.py::TestDatasetUtils::test_download_url_dispatch_download_from_google_drive[False]
Tests / unittests-macos (3.8, macos-m1-stable) / macos-job (gh)
test/test_transforms_v2.py::TestLinearTransform::test_transform[cpu-dtype1-make_image_tensor]

This comment was automatically generated by Dr. CI and updates every 15 minutes.

facebook-github-bot · 2024-03-05T01:43:41Z

Hi @gau-nernst!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

gau-nernst · 2024-03-05T08:22:59Z

@NicolasHug I signed the CLA but the status did not change. Not sure what went wrong. Do I need to use the same email as my GitHub account? I signed the CLA with a different email. Then I tried again with my GitHub email. Still no changes here.

torchvision/transforms/_functional_pil.py

vfdev-5 · 2024-03-05T13:36:52Z

torchvision/transforms/v2/functional/_geometry.py

+    angle = angle % 360  # shift angle to [0, 360) range
+
+    # fast path: transpose without affine transform
+    if expand or center is None:


why if expand ?

I was reading the NOTE section of v2.RandomRotation, and it says that

setting center has no effect if expand=True

From what I understand, if expand=True, we can ignore center, and thus we can safely use torch.rot90().

I think we can remove expand from the check

facebook-github-bot · 2024-03-05T15:06:17Z

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks!

gau-nernst · 2024-03-06T01:54:28Z

Regarding testing, I think we should test for angles 0, 90, 180, 270. Should I add them here?

vision/test/test_transforms_v2.py

Lines 1496 to 1503 in f3298dc

    
           class TestRotate: 
        
               _EXHAUSTIVE_TYPE_AFFINE_KWARGS = dict( 
        
                   # float, int 
        
                   angle=[-10.9, 18], 
        
                   # None 
        
                   # two-list of float, two-list of int, two-tuple of float, two-tuple of int 
        
                   center=[None, [1.2, 4.9], [-3, 1], (2.5, -4.7), (3, 2)], 
        
               )

My concern is that it will create many more test cases.

The alternative is to write a separate test for angles 0, 90, 180, 270, and we don't use too many values for other parameters (center, expand, fill).

NicolasHug · 2024-03-06T09:46:15Z

Thanks @gau-nernst

The alternative is to write a separate test for angles 0, 90, 180, 270, and we don't use too many values for other parameters (center, expand, fill).

Yes, we can probably create a small separate test to check these specific values, and simply test the output of PIL vs our tensor implementation

NicolasHug

Thanks a lot @gau-nernst for the PR and @vfdev-5 for the reviews. This is good to go on my side, let's just wait for @vfdev-5 's input before merging

vfdev-5

lgtm as well, thanks @gau-nernst !

NicolasHug · 2024-03-13T11:25:49Z

torchvision/transforms/v2/functional/_geometry.py

+    angle = angle % 360  # shift angle to [0, 360) range
+
+    # fast path: transpose without affine transform
+    if center is None:


The tests are failing and I think it's because we should set expand to True. When expand is False we're not supposed to be changing the shape of the output (which is what rot90 does!).

Suggested change

if center is None:

if expand and center is None:

Here is what does PIL here:
https://github.com/python-pillow/Pillow/blob/8f63748e50378424628155994efd7e0739a4d1d1/src/PIL/Image.py#L2287-L2296

Yes, I agree that it is incorrect to omit expand for image with h != w, sorry for an incorrect suggestion previously.

Should we also check for square input then? So even if expand is false, rot90 is still valid.

Following their code they: 1) accept angle 0 and 180 whatever provided expand value and 2) for rotations 90 or 270 they do if angle in (90, 270) and (expand or self.width == self.height).
I think this is reasonable and we can do the same

I will update accordingly

I have updated the code. Also updated the test cases to cover everything. I just realized the tests didn't fail on my machine because I was running inside an environment with torchvision installed (big mistake!).

No worries @gau-nernst thanks for the update. I can confirm the tests are passing now (tried locally). let's just wait for the CI to be green before merging

github-actions · 2024-03-13T17:17:43Z

Hey @NicolasHug!

You merged this PR, but no labels were added.
The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py

Summary: Co-authored-by: Thien Tran <thien.tran@parallelchain.io> Reviewed By: vmoens Differential Revision: D55062767 fbshipit-source-id: b67a7f7fecab8a33143b95e6233637c336c74cd4

add rotation fast path

3c5b5b2

vfdev-5 reviewed Mar 5, 2024

View reviewed changes

torchvision/transforms/_functional_pil.py Outdated Show resolved Hide resolved

gau-nernst added 3 commits March 5, 2024 21:25

Merge branch 'pytorch:main' into main

8652144

remove PIL fast path

cb4f943

return clone() for angle=0

dbb36a8

vfdev-5 reviewed Mar 5, 2024

View reviewed changes

gau-nernst added 3 commits March 5, 2024 22:08

fix torchscript (and typing)

f4d9e00

remove expand check for fast path

7cad321

Merge branch 'main' into main

dbde073

facebook-github-bot added the cla signed label Mar 5, 2024

gau-nernst added 3 commits March 6, 2024 21:25

add test

872d4c4

Merge branch 'main' into main

10f51bd

Merge branch 'main' into main

1860a1f

gau-nernst requested a review from vfdev-5 March 12, 2024 11:34

NicolasHug approved these changes Mar 13, 2024

View reviewed changes

vfdev-5 approved these changes Mar 13, 2024

View reviewed changes

NicolasHug reviewed Mar 13, 2024

View reviewed changes

NicolasHug mentioned this pull request Mar 13, 2024

Add Jpeg Compression augmentation #8290

Closed

gau-nernst added 3 commits March 14, 2024 00:06

Merge branch 'main' into main

53f6c92

fix fast path

154d715

Merge branch 'main' into main

4260a3b

gau-nernst requested review from NicolasHug and vfdev-5 March 13, 2024 16:40

NicolasHug merged commit ba64d65 into pytorch:main Mar 13, 2024
72 of 75 checks passed

NicolasHug added enhancement module: transforms labels Mar 14, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fast rotation for right angles #8295

Fast rotation for right angles #8295

gau-nernst commented Mar 5, 2024 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Mar 5, 2024 •

edited

Loading

facebook-github-bot commented Mar 5, 2024

gau-nernst commented Mar 5, 2024

vfdev-5 Mar 5, 2024

gau-nernst Mar 5, 2024

vfdev-5 Mar 5, 2024

facebook-github-bot commented Mar 5, 2024

gau-nernst commented Mar 6, 2024

NicolasHug commented Mar 6, 2024

NicolasHug left a comment

vfdev-5 left a comment

NicolasHug Mar 13, 2024

vfdev-5 Mar 13, 2024

gau-nernst Mar 13, 2024

vfdev-5 Mar 13, 2024

gau-nernst Mar 13, 2024

gau-nernst Mar 13, 2024

NicolasHug Mar 13, 2024

github-actions bot commented Mar 13, 2024

Fast rotation for right angles #8295

Fast rotation for right angles #8295

Conversation

gau-nernst commented Mar 5, 2024 • edited by pytorch-bot bot Loading

pytorch-bot bot commented Mar 5, 2024 • edited Loading

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8295

❌ 3 New Failures

facebook-github-bot commented Mar 5, 2024

Action Required

Process

gau-nernst commented Mar 5, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

facebook-github-bot commented Mar 5, 2024

gau-nernst commented Mar 6, 2024

NicolasHug commented Mar 6, 2024

NicolasHug left a comment

Choose a reason for hiding this comment

vfdev-5 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

github-actions bot commented Mar 13, 2024

gau-nernst commented Mar 5, 2024 •

edited by pytorch-bot bot

Loading

pytorch-bot bot commented Mar 5, 2024 •

edited

Loading