Add gaussian noise transform #6192 #6233

parth-shastri · 2022-07-03T09:24:32Z

Adds GaussianNoise transform functionality to the torchvision.transform.

-Basic GaussianNoise augmentation technique

facebook-github-bot · 2022-07-03T09:24:36Z

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 [email protected]. Thanks!

parth-shastri · 2022-07-03T09:38:28Z

This PR addresses the issue #6192.

facebook-github-bot · 2022-07-03T10:55:43Z

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

oke-aditya

Hi @parth-shastri thanks a lot for the PR.

While I haven't gone through thorough details and verified the implementation fully. I had few code style and comments.

Also,

torchvision/_C.pyd

file is unnecessary. Is it because you are using windows and this file is generated while compiling torchvision from source?

I am sure @vfdev-5 is champion in this domain and will have a deep look 😄

torchvision/transforms/transforms.py

oke-aditya · 2022-07-03T18:34:48Z

torchvision/transforms/transforms.py

+        Returns:
+            PIL Image or Tensor: Image added with gaussian noise.
+        """
+        if not isinstance(image, torch.Tensor):


I'm bit hesitant about this implementation. Why not create F.gaussian_noise that would work well on PIL Images and tensors which we can then use here?

See https://github.com/pytorch/vision/blob/main/torchvision/transforms/functional.py#L1338

I agree, let's create a functional op F.gaussian_noise and at first iteration we can have tensor implementation. If input is PIL, we can convert it to tensor, apply the op and get back PIL image type for the output.

@parth-shastri please create a functional op F.gaussian_noise

@parth-shastri please address this comment

Yes working on that

test/test_transforms.py

torchvision/transforms/transforms.py

datumbox · 2022-07-04T14:03:11Z

@parth-shastri Thanks for the contribution.

As noted at #6192 (comment), we are currently reviewing the API of Transforms so ideally we don't want to introduce new ones while working on it. Perhaps we could add this directly to the new API as a means of offering incentives to the users to migrate. @vfdev-5 wdyt?

vfdev-5 · 2022-07-04T14:09:06Z

@parth-shastri Thanks for the contribution.

As noted at #6192 (comment), we are currently reviewing the API of Transforms so ideally we don't want to introduce new ones while working on it. Perhaps we could add this directly to the new API as a means of offering incentives to the users to migrate. @vfdev-5 wdyt?

I think it is OK, this transformation is applied on images and meanwhile we are coding/reviewing new API, we can accept small transforms like this one.

Co-authored-by: Aditya Oke <[email protected]>

vfdev-5 · 2022-07-07T07:46:33Z

@parth-shastri can you please fix failing job "ci/circleci: lint_python_and_config"

vfdev-5 · 2022-07-07T07:52:45Z

gallery/plot_transforms.py

+noisy = T.GaussianNoise(mean=0.0, sigma=(0.1, 5))
+noisy_imgs = [noisy(orig_img) for _ in range(4)]
+plot(noisy_imgs)


This gives the following result: https://output.circle-artifacts.com/output/job/34f9c3d8-231e-4748-81b9-cda1724a9f16/artifacts/0/docs/auto_examples/plot_transforms.html#gaussiannoise which does not make sense. Let's reduce the number of transformed images and propose acceptable range of sigmas

I think we should not keep an accepted range of sigmas, just let the user decide how much distortion does he want ?

Fixed the documentation images in the documentation.

…th-shastri/vision into add-gaussian-noise-transform

gallery/plot_transforms.py

test/test_transforms.py

torchvision/transforms/transforms.py

vfdev-5 · 2022-07-25T15:01:57Z

torchvision/transforms/transforms.py

+        Returns:
+            PIL Image or Tensor: Image added with gaussian noise.
+        """
+        if not isinstance(image, torch.Tensor):


@parth-shastri please address this comment

addition of type hints Co-authored-by: Aditya Oke <[email protected]>

Co-authored-by: vfdev <[email protected]>

added isinstance instead of F._is_pil_image()

parth-shastri · 2022-08-21T14:41:46Z

Are we stuck on this? @parth-shastri do you need help?
Hey, thanks I am not able to decide on a valid sigma to display
and also is there a way that we can visualize the results in the documentation before the commit?
apart from the deciding of sigma is there anything else that I'm supposed to do?

oke-aditya · 2022-08-21T15:43:15Z

You can visualise the results by building the docs.
Also I think you need to refactor the code to create a functional API.

You can just try doing a grid search (or just try a few values and post the images here)

cc @vfdev-5 any suggestions for Sigma ?

parth-shastri · 2022-09-07T12:22:30Z

Hi, tried out some sigmas and found that the range [0.1, 0.5] works.

vfdev-5

Thanks for the update @parth-shastri
I left few other comment to improve the PR

torchvision/transforms/functional_tensor.py

vfdev-5 · 2022-09-09T13:54:40Z

torchvision/transforms/functional_tensor.py

+        ],
+    )
+    # add the gaussian noise with the given mean and sigma.
+    normalize_img = img / 255.0


We can't make this assumption that img range is [0, 255].
By default, we assume: [0, 1] for float images and [0, 255] for uint8 RGB images.
Let's not rescale image range and let user pick appropriate mean and sigma according to their data range.

vfdev-5 · 2022-09-09T13:56:10Z

torchvision/transforms/functional_tensor.py

@@ -748,8 +748,6 @@ def gaussian_blur(img: Tensor, kernel_size: List[int], sigma: List[float]) -> Te
            kernel.dtype,
        ],
    )
-
-    # padding = (left, right, top, bottom)


Let's keep this comment

vfdev-5 · 2022-09-09T13:58:06Z

torchvision/transforms/functional.py

+    if sigma is None:
+        raise ValueError("The value of sigma cannot be None.")
+
+    if sigma is not None and not isinstance(sigma, (int, float, list, tuple)):
+        raise TypeError(f"sigma should be either float or sequence of floats. Got {type(sigma)}")


Something is unclear here about whether sigma is Optional or not. Docstring and typehint says optional, but you raise an error if sigma is None. Let's remove optional and assume that sigma is not None

@parth-shastri check above comment and the code. Why do you check sigma is None if sigma is not intended to be Optional.

Co-authored-by: vfdev <[email protected]>

parth-shastri · 2022-10-27T13:14:48Z

Hi @vfdev-5 Anything more that you want me to do
Did not get the build error resolved can you help?

torchvision/transforms/functional.py

parth-shastri · 2022-12-31T19:09:20Z

@vfdev-5 is there anything more to do?

parth-shastri · 2023-02-16T07:56:39Z

@oke-aditya @vfdev-5 can you please highlight any changes I need to make in order to merge this pull request?

javirk · 2023-11-28T13:15:32Z

Hi. I think it would be valuable to have Gaussian Noise as a transformation. What is the state of this at the moment? Should I take over this PR in order to finally merge it?

vfdev-5 · 2023-11-28T13:41:30Z

The problem with this PR right now is that it should add code compatible to the new v2 API.
I think it would be easier to start from a new PR and add new transform and all functional part. @javirk if you are interested in working on that, feel free to provide another PR. You may cherry-pick some commits on the gallery to acknowledge the work of @parth-shastri.

parth-shastri · 2024-02-08T17:22:43Z

The problem with this PR right now is that it should add code compatible to the new v2 API. I think it would be easier to start from a new PR and add new transform and all functional part. @javirk if you are interested in working on that, feel free to provide another PR. You may cherry-pick some commits on the gallery to acknowledge the work of @parth-shastri.

Hi I can take this up, do you want me to create another pull request compatible with the new v2 API or shall I continue here?

NicolasHug · 2024-05-31T10:26:59Z

Thanks a lot for your PR @parth-shastri . We just merged #8381 which adds gaussian noise, but thanks again for the original implementation!

parth-shastri added 2 commits July 3, 2022 12:31

adds gaussian noise transform

9be5e13

adds gaussian noise transform

42952f6

facebook-github-bot added the cla signed label Jul 3, 2022

oke-aditya reviewed Jul 3, 2022

View reviewed changes

parth-shastri and others added 7 commits July 4, 2022 20:20

Update torchvision/transforms/transforms.py

db9756a

Co-authored-by: Aditya Oke <[email protected]>

Update torchvision/transforms/transforms.py

05a52af

Co-authored-by: Aditya Oke <[email protected]>

Delete _C.pyd

c380281

added GaussianNoise transform

431a7e0

adds the GaussianNoise transform

5fc7c85

adds GaussianNoise transform

179908b

Merge branch 'main' into add-gaussian-noise-transform

396abba

vfdev-5 reviewed Jul 7, 2022

View reviewed changes

parth-shastri and others added 6 commits July 12, 2022 22:06

fixes on the lint tests and the plot_transforms

5d8b0f7

Merge branch 'main' into add-gaussian-noise-transform

98e4e98

test

aa9b2e7

test

223074f

Merge branch 'add-gaussian-noise-transform' of https://github.com/par…

a26ed67

…th-shastri/vision into add-gaussian-noise-transform

fixes the plot_transforms bug

6d57443

vfdev-5 reviewed Jul 25, 2022

View reviewed changes

vfdev-5 self-assigned this Jul 25, 2022

parth-shastri and others added 4 commits July 27, 2022 13:54

Update torchvision/transforms/transforms.py

1d1fbcd

addition of type hints Co-authored-by: Aditya Oke <[email protected]>

Merge branch 'main' into add-gaussian-noise-transform

ff80571

Update gallery/plot_transforms.py

533e76f

Co-authored-by: vfdev <[email protected]>

Update test_transforms.py

b8d98d7

added isinstance instead of F._is_pil_image()

Merge branch 'main' into add-gaussian-noise-transform

7fc04aa

parth-shastri added 2 commits September 7, 2022 17:56

adds functional transforms, fixed sigma

8a500fe

updated lint, adds functional transforms, fixed sigma

8b560fb

vfdev-5 reviewed Sep 9, 2022

View reviewed changes

parth-shastri and others added 7 commits October 5, 2022 18:52

Merge branch 'main' into add-gaussian-noise-transform

ac15585

Update torchvision/transforms/functional_tensor.py

3a85c34

Co-authored-by: vfdev <[email protected]>

Merge branch 'main' into add-gaussian-noise-transform

5892695

suggested changes

e6b4e45

update

58d525d

fixed docs

021ecba

Merge branch 'main' into add-gaussian-noise-transform

5956088

Merge branch 'main' into add-gaussian-noise-transform

92d024f

vfdev-5 reviewed Oct 28, 2022

View reviewed changes

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

parth-shastri and others added 4 commits November 9, 2022 20:22

Merge branch 'main' into add-gaussian-noise-transform

bfde863

Merge branch 'main' into add-gaussian-noise-transform

dbc3e1a

Merge branch 'main' into add-gaussian-noise-transform

d18195b

fixes for random calls in functional transforms

a8d8137

Merge branch 'main' into add-gaussian-noise-transform

2f7f558

Richienb mentioned this pull request Apr 17, 2024

Add GaussianNoise #8381

Merged

NicolasHug closed this in #8381 May 31, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add gaussian noise transform #6192 #6233

Add gaussian noise transform #6192 #6233

parth-shastri commented Jul 3, 2022 •

edited by vfdev-5

Loading

facebook-github-bot commented Jul 3, 2022

parth-shastri commented Jul 3, 2022

facebook-github-bot commented Jul 3, 2022

oke-aditya left a comment

oke-aditya Jul 3, 2022 •

edited

Loading

vfdev-5 Jul 4, 2022

vfdev-5 Jul 7, 2022

vfdev-5 Jul 25, 2022

parth-shastri Jul 28, 2022

datumbox commented Jul 4, 2022

vfdev-5 commented Jul 4, 2022

vfdev-5 commented Jul 7, 2022

vfdev-5 Jul 7, 2022

parth-shastri Jul 13, 2022

parth-shastri Jul 13, 2022

vfdev-5 Jul 25, 2022

parth-shastri commented Aug 21, 2022

oke-aditya commented Aug 21, 2022

parth-shastri commented Sep 7, 2022

vfdev-5 left a comment

vfdev-5 Sep 9, 2022

vfdev-5 Sep 9, 2022

vfdev-5 Sep 9, 2022

vfdev-5 Oct 28, 2022

parth-shastri commented Oct 27, 2022

parth-shastri commented Dec 31, 2022

parth-shastri commented Feb 16, 2023

javirk commented Nov 28, 2023

vfdev-5 commented Nov 28, 2023

parth-shastri commented Feb 8, 2024

NicolasHug commented May 31, 2024

Add gaussian noise transform #6192 #6233

Add gaussian noise transform #6192 #6233

Conversation

parth-shastri commented Jul 3, 2022 • edited by vfdev-5 Loading

facebook-github-bot commented Jul 3, 2022

Action Required

Process

parth-shastri commented Jul 3, 2022

facebook-github-bot commented Jul 3, 2022

oke-aditya left a comment

Choose a reason for hiding this comment

oke-aditya Jul 3, 2022 • edited Loading

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

datumbox commented Jul 4, 2022

vfdev-5 commented Jul 4, 2022

vfdev-5 commented Jul 7, 2022

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

parth-shastri commented Aug 21, 2022

oke-aditya commented Aug 21, 2022

parth-shastri commented Sep 7, 2022

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

parth-shastri commented Oct 27, 2022

parth-shastri commented Dec 31, 2022

parth-shastri commented Feb 16, 2023

javirk commented Nov 28, 2023

vfdev-5 commented Nov 28, 2023

parth-shastri commented Feb 8, 2024

NicolasHug commented May 31, 2024

parth-shastri commented Jul 3, 2022 •

edited by vfdev-5

Loading

oke-aditya Jul 3, 2022 •

edited

Loading