support of float dtypes for draw_segmentation_masks #8150

GsnMithra · 2023-12-10T21:47:33Z

Hey there!

I've implemented support for draw_* methods to seamlessly handle both uint8 and float32 image types. The processed image will now be returned with the same data type as the input. Your feedback is always appreciated.

Thank you!

pytorch-bot · 2023-12-10T21:47:36Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8150

📄 Preview Python docs built from this PR

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

✅ You can merge normally! (2 Unrelated Failures)

As of commit 3dce586 with merge base c35d385 ():

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

Build Linux Conda / pytorch/vision / conda-py3_8-cuda11_8 (gh)
An unexpected error has occurred. Conda has prepared the above report.
Build Linux Conda / pytorch/vision / conda-py3_8-cuda12_1 (gh)
An unexpected error has occurred. Conda has prepared the above report.

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

NicolasHug

Thanks a lot for submitting this PR @GsnMithra

This is promising, but allowing support for float will require a little bit more work than just converting the dtype using .to(), since the convention for float images is that their valuae range is in [0, 1] instead of [0, 255].

Perhaps the easiest way to make progress here will be to focus on draw_segmentation_masks first, and to add a small unit test in https://github.com/pytorch/vision/blob/main/test/test_utils.py, that checks for equality of result for a given image with different dtypes. Something roughly like this:

from torchvision.transforms.v2.functional import to_dtype

img_uint8 = torch.randing(0, 256, (3, 100, 100), dtype=torch.uint8)
img_float = to_dtype(img_uint8, torch.float32, scale=True)

out_uint8 = draw_segmentation_mask(img_uint8, ...)
out_float = draw_segmentation_mask(img_float, ...)

torch.testing.assert_close(out_uint8, to_dtype(out_float, torch.uint8, scale=True))

Does that make sense?

NicolasHug · 2023-12-11T17:37:05Z

torchvision/utils.py

+    elif image.dtype not in {torch.uint8, torch.float32}:
+        raise ValueError(f"The image dtype must be uint8 or float32, got {image.dtype}")


Here and in the other places, let's just check for image.is_floating_point() instead of checking specifically for float32. This way we can also support float64, etc.

Hey, thanks for replying, this will be added with the upcoming commits.

torchvision/utils.py

Co-authored-by: Nicolas Hug <[email protected]>

GsnMithra · 2023-12-12T06:17:08Z

Hey @NicolasHug

Just dropping a note to mention that I've added some assertion checks for floating-point data types in the unit test method for draw_segmentation_mask() with the recent commit titled "test_draw_sementation_mask". I would love to get your thoughts on it.

Thank you for your time.

NicolasHug · 2023-12-13T14:18:37Z

Thanks a lot @GsnMithra . It looks like there are a few things missing for now: the import of to_dtype is incorrect (it's written "to_type"), and some deleted variables are still in use, throwing an error (out_dtype). Make sure to run the tests locally before pushing :) We have instructions here https://github.com/pytorch/vision/blob/main/CONTRIBUTING.md#unit-tests

Also, I think it would be best not to modify the current tests, and instead just create a separate new test for the one suggested in #8150 (review). Thank you!

GsnMithra · 2023-12-14T19:34:45Z

Hey @NicolasHug

I would like to sincerely apologize for the mistakes I made in my previous contributions. I am still in the learning process and appreciate your guidance.

In the latest commit, I have included a new unit test called test_draw_segmentation_masks_dtypes, which has been implemented according to the suggestions provided earlier. I would greatly appreciate your thoughts and feedback on this addition.

Once again, I apologize for any inconvenience caused and thank you for taking the time to review my work.

NicolasHug

Thank you for the follow-up @GsnMithra, and no problems for the first few hicups.

I can see the new test is passing now, good job! In order to move forward with this PR, I would suggest the following:

Address the comments I made below (it should be pretty easy)
Remove the changes that were made to draw_bounding_boxes() and draw_keypoints(), and only keep the changes and tests relating to draw_segmentation_masks(). This way, we'll be able to merge this PR straight away, and then you can send 2 separate PR (one for boxes, one for keypoints), should you wish to.

Does that sound good to you?

test/test_utils.py

.gitignore

NicolasHug · 2023-12-15T11:43:37Z

test/test_utils.py

+import torchvision.transforms.functional as F
+from torchvision.transforms.v2.functional import to_dtype


You'll need to re-order the imports, check out https://github.com/pytorch/vision/actions/runs/7214131414/job/19676695616?pr=8150, or run the pre-commit hooks locally (check our contributing instructions)

test/test_utils.py

NicolasHug · 2023-12-15T11:45:25Z

torchvision/utils.py

+    if image.is_floating_point():
+        image = (image * 255).to(torch.uint8)


Let's use to_dtype here instead, like you did in the test.

NicolasHug · 2023-12-15T11:45:33Z

torchvision/utils.py

+    if original_dtype in {torch.float16, torch.float32, torch.float64}:
+        out = out.float() / 255.0


Same: Let's use to_dtype here instead, like you did in the test.

NicolasHug · 2023-12-15T11:46:10Z

torchvision/utils.py

@@ -315,7 +318,10 @@ def draw_segmentation_masks(
        img_to_draw[:, mask] = color[:, None]

    out = image * (1 - alpha) + img_to_draw * alpha
-    return out.to(out_dtype)
+    if original_dtype in {torch.float16, torch.float32, torch.float64}:


Check for is_floating_point() instead.

Co-authored-by: Nicolas Hug <[email protected]>

GsnMithra · 2023-12-15T15:35:30Z

Hello @NicolasHug

I hope this message finds you well. I wanted to bring up a couple of points regarding the usage of to_dtype in the torchvision.transforms.v2.functional module, particularly within the draw_segmentation_masks() method.

Firstly, when importing to_dtype globally, it leads to a circular import issue due to the _log_api_usage_once(to_dtype) being utilized within the function. One solution could be importing it inside the draw_segmentation_masks() method, or alternatively, manually converting it.

Secondly, I've encountered a situation where, when a float dtype image is passed, converting it back to float using the following code results in mismatched elements for the unit test:

out = image * (1 - alpha) + img_to_draw * alpha
if original_image.is_floating_point():
    out = to_dtype(out, torch.float, scale=True)

It seems to work only when dividing by 255.0:

out = image * (1 - alpha) + img_to_draw * alpha
if original_image.is_floating_point():
    out = to_dtype(out, torch.float) / 255.0

I appreciate your insights and look forward to any further guidance or suggestions you may have.
Thank you for taking the time.

…at32

…o support-for-float32

NicolasHug

Thanks a lot for the follow-up @GsnMithra !

You are right about the need for /= 255. It was because out was a float32 image in [0, 255] when the input was a float image. So even doing dtype(scale=True) would not work as expected, as no conversion was done (since the image was already float32).

I found a simpler way to handle all this by simply converting the color's dtype and scale, instead of converting the input image. I pushed the changes and also reverted some minor things from the other functions.

I'll merge this PR when green, thanks a ton for your work on this!

(Feel free to submit follow-up PRs for the other functions if you wish to. Just giving you a heads-up that I will only be able to review them starting next year in Jan, as I'll be on leaves from tomorrow.)

Thanks again @GsnMithra !

GsnMithra · 2023-12-18T12:22:43Z

Hey @NicolasHug
You're welcome, I'll be submitting PR's for the other functions also.

Wishing you a wonderful break ;)

github-actions · 2023-12-18T13:40:16Z

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: Nicolas Hug <[email protected]> Reviewed By: vmoens Differential Revision: D52539003 fbshipit-source-id: e7b9412a496e88749dc6e9c5afdd1b5cf85b4aa0

support for float32 for draw_*

f36dd2c

facebook-github-bot added the cla signed label Dec 10, 2023

GsnMithra and others added 3 commits December 11, 2023 03:44

cleaning up

f091cc7

method description

5672d8c

Merge branch 'main' into support-for-float32

e06c1f6

NicolasHug reviewed Dec 11, 2023

View reviewed changes

GsnMithra and others added 3 commits December 12, 2023 02:43

Update torchvision/utils.py

301e42b

Co-authored-by: Nicolas Hug <[email protected]>

floating point dtype check

9146ebe

test_draw_segmentation_masks

bb1cd08

unit test for draw_segmentation_masks dtypes

8d486a8

GsnMithra added 3 commits December 15, 2023 01:38

cleaning up other method and checks

baec368

removing .DS_Store files

745e1b2

gitignore updated for ignoring .DS_Store files

6cfa9d8

NicolasHug reviewed Dec 15, 2023

View reviewed changes

GsnMithra and others added 3 commits December 15, 2023 17:28

Update test/test_utils.py

956a7bd

Co-authored-by: Nicolas Hug <[email protected]>

Update .gitignore

3bcd9a3

Co-authored-by: Nicolas Hug <[email protected]>

Update test/test_utils.py

7e3a176

Co-authored-by: Nicolas Hug <[email protected]>

GsnMithra added 4 commits December 15, 2023 21:08

reverting boxes and keypoints methods and unit tests

acdad33

floating_point check

cad307c

cleaning up

90982ff

floating_point check updated

020a848

GsnMithra changed the title support for float32 for draw_* support for float for draw_segmentation_masks Dec 15, 2023

GsnMithra changed the title ~~support for float for draw_segmentation_masks~~ support of float dtypes for draw_segmentation_masks Dec 15, 2023

GsnMithra and others added 2 commits December 16, 2023 00:58

type conversions using to_dtype() with local import

ee048c5

Merge branch 'main' of github.com:pytorch/vision into support-for-flo…

a5e4c54

…at32

NicolasHug added 3 commits December 18, 2023 11:38

Merge branch 'support-for-float32' of github.com:GsnMithra/vision int…

0b82508

…o support-for-float32

Use _parse_colors to handle dtype conversion

adf249d

Revert some changes

3dce586

NicolasHug approved these changes Dec 18, 2023

View reviewed changes

NicolasHug merged commit 6c2e0ae into pytorch:main Dec 18, 2023
62 of 64 checks passed

NicolasHug added enhancement module: utils labels Dec 18, 2023

GsnMithra mentioned this pull request Feb 17, 2024

draw_keypoints() float support #8276

Merged

GeeCastro mentioned this pull request Mar 17, 2024

add float support to utils.draw_bounding_boxes() #8328

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

support of float dtypes for draw_segmentation_masks #8150

support of float dtypes for draw_segmentation_masks #8150

GsnMithra commented Dec 10, 2023

pytorch-bot bot commented Dec 10, 2023 •

edited

Loading

NicolasHug left a comment

NicolasHug Dec 11, 2023

GsnMithra Dec 11, 2023 •

edited

Loading

GsnMithra commented Dec 12, 2023

NicolasHug commented Dec 13, 2023

GsnMithra commented Dec 14, 2023

NicolasHug left a comment

NicolasHug Dec 15, 2023

NicolasHug Dec 15, 2023

NicolasHug Dec 15, 2023

NicolasHug Dec 15, 2023

GsnMithra commented Dec 15, 2023

NicolasHug left a comment

GsnMithra commented Dec 18, 2023

github-actions bot commented Dec 18, 2023

		elif image.dtype not in {torch.uint8, torch.float32}:
		raise ValueError(f"The image dtype must be uint8 or float32, got {image.dtype}")

		import torchvision.transforms.functional as F
		from torchvision.transforms.v2.functional import to_dtype

		if image.is_floating_point():
		image = (image * 255).to(torch.uint8)

		if original_dtype in {torch.float16, torch.float32, torch.float64}:
		out = out.float() / 255.0

support of float dtypes for draw_segmentation_masks #8150

support of float dtypes for draw_segmentation_masks #8150

Conversation

GsnMithra commented Dec 10, 2023

pytorch-bot bot commented Dec 10, 2023 • edited Loading

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8150

✅ You can merge normally! (2 Unrelated Failures)

NicolasHug left a comment

Choose a reason for hiding this comment

NicolasHug Dec 11, 2023

Choose a reason for hiding this comment

GsnMithra Dec 11, 2023 • edited Loading

Choose a reason for hiding this comment

GsnMithra commented Dec 12, 2023

NicolasHug commented Dec 13, 2023

GsnMithra commented Dec 14, 2023

NicolasHug left a comment

Choose a reason for hiding this comment

NicolasHug Dec 15, 2023

Choose a reason for hiding this comment

NicolasHug Dec 15, 2023

Choose a reason for hiding this comment

NicolasHug Dec 15, 2023

Choose a reason for hiding this comment

NicolasHug Dec 15, 2023

Choose a reason for hiding this comment

GsnMithra commented Dec 15, 2023

NicolasHug left a comment

Choose a reason for hiding this comment

GsnMithra commented Dec 18, 2023

github-actions bot commented Dec 18, 2023

pytorch-bot bot commented Dec 10, 2023 •

edited

Loading

GsnMithra Dec 11, 2023 •

edited

Loading