Added GaussianBlur transform

[tejank10]

This PR addresses issue #2635. It adds GaussianBlur image transformation. It takes a range of sigma as input. It picks a sigma for Gaussian kernel at random and applies the kernel to the image.

• Implement the transform
• Add tests
• Add docs

Let me know your thoughts @vfdev-5, @yaox12 on where could it be possibly improved.

[oke-aditya]

Small linting errors.

./torchvision/transforms/functional_tensor.py:473:16: E225 missing whitespace around operator
./torchvision/transforms/functional_tensor.py:482:23: E221 multiple spaces before operator
./torchvision/transforms/functional_tensor.py:483:22: E226 missing whitespace around arithmetic operator
./torchvision/transforms/functional_tensor.py:1001:31: E226 missing whitespace around arithmetic operator
./torchvision/transforms/functional_tensor.py:1002:19: E221 multiple spaces before operator

[vfdev-5]

@tejank10 thanks for the PR and sorry for delay.

Finally I think the API we would like is with kernel_size and sigma:

# functional.py
def gaussian_blur(img: Tensor, kernel_size: int, sigma: float) -> Tensor:

The main problem is with PIL which can apply kernels larger than 5x5. Maybe, a solution can be to reuse torch implementation for that : PIL -> np.array -> torch.Tensor -> F_t.gaussian_blur -> np.array -> PIL

About your implementation, I'll comment the things where I have questions.

[vfdev-5]

@tejank10 OK, let's work on this PR the following way:

• implement F_t.gaussian_blur(img: Tensor, kernel_size: int, sigma: float) which should be close to opencv's implementation : kernel creation, padding, etc (maybe, this is how it is done in opencv).
• implement tests

I propose you to work on F_t.gaussian_blur if you do not mind and I'll work on tests and commit my modifications directly to this PR. How does it sound ?

EDIT: I'll also push some modifications of F.gaussian_blur on how to dispatch between PIL and torch Tensor.

[tejank10]

Sounds good, will update the API.

[vfdev-5]

@yaox12 how about this API ?

[codecov]

Codecov Report

Merging #2658 into master will increase coverage by 0.15%.
The diff coverage is 86.86%.

@@            Coverage Diff             @@
##           master    #2658      +/-   ##
==========================================
+ Coverage   72.75%   72.90%   +0.15%     
==========================================
  Files          96       96              
  Lines        8309     8396      +87     
  Branches     1292     1315      +23     
==========================================
+ Hits         6045     6121      +76     
- Misses       1871     1878       +7     
- Partials      393      397       +4     

Impacted Files	Coverage Δ
torchvision/transforms/transforms.py	81.25% <81.48%> (+0.32%)	⬆️
torchvision/transforms/functional.py	80.44% <81.81%> (+0.07%)	⬆️
torchvision/transforms/functional_tensor.py	73.91% <94.73%> (+1.36%)	⬆️
torchvision/transforms/functional_pil.py	66.19% <100.00%> (ø)

---

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 754c954...94477dc. Read the comment docs.

[vfdev-5]

@tejank10 we have to improve coverage by executing all input checkings and asserting that they raise correct exception. Can you work on that too ?

[vfdev-5]

@tejank10 do you plan to work on this PR this week ? Such that we do not do the same things without syncing. Otherwise, I'll rework everything tomorrow.

[tejank10]

Yeah, I'll be working on this PR this week.

[vfdev-5]

@tejank10 could you please push all your work today if any. I'll finalize today / tomorrow this PR.

[vfdev-5]

@tejank10 I'll fix broken CI and merge conflict today. Let me know if you want to push some other updates. Thanks

[fmassa]

Thanks a lot @tejank10 and @vfdev-5 !