Add dim to pytorch_lightning.metrics.PSNR

[manipopopo]

What does this PR do?

Fixes #5933.

Suppose we have a batch containing two image pairs, the original PSNR calculates the mean-squared-error of the two pairs and then take the log to get the PSNR of the batch. By setting dim, users can choose to calculate the _PSNR_s of the two pairs separately and then average the two _PSNR_s to get the score of the batch:

preds_image_nchw = torch.rand([2, 3, 512, 512])
target_image_nchw = torch.rand([2, 3, 512, 512])

# (pred_image_1, target_image_1)
# (pred_image_2, target_image_2)
# squared_error_1 = (pred_image_1 - target_image_1) ** 2
# squared_error_2 = (pred_image_2 - target_image_2) ** 2
# mean_squared_error = (squared_error_1 + squared_error_2) / (squared_error_1.numel() + squared_error_2.numel())
# log_mse_of_two_pairs = -10.0 * log(mean_squared_error)
log_mse_of_two_pairs = pl.metrics.functional.psnr(
    preds_image_nchw,
    target_image_nchw,
    data_range=1.0
)

# squared_error_1 = (pred_image_1 - target_image_1) ** 2
# psnr_1 = -10.0 * log(squared_error_1.mean())
# squared_error_2 = (pred_image_2 - target_image_2) ** 2
# psnr_2 = -10.0 * log(squared_error_2.mean())
# average_log_mse = (psnr_1 + psnr_2) / 2
average_log_mse = pl.metrics.functional.psnr(
    preds_image_nchw, 
    target_image_nchw,
    data_range=1.0,
    dim=(1, 2, 3)
)

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified
• Check that target branch and milestone match!

Did you have fun?

Make sure you had fun coding 🙃

[codecov]

Codecov Report

Merging #5957 (d296919) into master (f655f97) will decrease coverage by 58%.
The diff coverage is 12%.

@@           Coverage Diff            @@
##           master   #5957     +/-   ##
========================================
- Coverage      91%     33%    -58%     
========================================
  Files         160     160             
  Lines       11313   11232     -81     
========================================
- Hits        10272    3701   -6571     
- Misses       1041    7531   +6490     

[manipopopo]

@SkafteNicki
I'm not sure which target branch (master or release/1.2-dev) should be used :(. It seems that there are some modifications to the relevant files on the master branch.

[Borda]

@SkafteNicki mind have look?

[SkafteNicki]

could we move all this logic into the _update function?

[manipopopo]

In the original implementation (before this PR) and the updated commits (after the comment), the data_range is passed to functional._psnr_compute, maybe we could move the checks on data_range and dim from functional.psnr and PSNR.__init__ to _compute. But the users of PSNR will get the exception until they call PSNR.compute (or validation_epoch_end is called).

Should we extract the logic to a single place (functional._psnr_compute)?

[tchaton]

LGTM ! Great contribution !

[pep8speaks]

Hello @manipopopo! Thanks for updating this PR.

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-02-17 07:23:02 UTC

[SkafteNicki]

LGTM