Add drop_last option to method fit of PyTorchClassifier

[beat-buesser]

Description

This pull request adds drop_last option to method fit of PyTorchClassifier.

Fixes #1723

Type of change

Please check all relevant options.

• Improvement (non-breaking)
• Bug fix (non-breaking)
• New feature (non-breaking)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

[codecov-commenter]

Codecov Report

Merging #1883 (fc12c05) into dev_1.12.2 (3e3a438) will decrease coverage by 0.16%.
The diff coverage is 48.83%.

@@              Coverage Diff               @@
##           dev_1.12.2    #1883      +/-   ##
==============================================
- Coverage       85.90%   85.73%   -0.17%     
==============================================
  Files             248      248              
  Lines           23312    23491     +179     
  Branches         4212     4280      +68     
==============================================
+ Hits            20027    20141     +114     
- Misses           2224     2273      +49     
- Partials         1061     1077      +16     

Impacted Files	Coverage Δ
...rs/certification/derandomized_smoothing/pytorch.py	79.31% <40.00%> (-8.69%)	⬇️
art/estimators/classification/pytorch.py	86.79% <50.00%> (-1.56%)	⬇️
...tors/certification/randomized_smoothing/pytorch.py	84.09% <52.94%> (-8.02%)	⬇️
...poison_mitigation/neural_cleanse/neural_cleanse.py	84.21% <0.00%> (-6.15%)	⬇️
art/attacks/evasion/boundary.py	92.77% <0.00%> (-1.21%)	⬇️
art/attacks/poisoning/sleeper_agent_attack.py	41.00% <0.00%> (-0.36%)	⬇️
art/metrics/verification_decisions_trees.py	92.83% <0.00%> (+1.16%)	⬆️

[kieranfraser]

Hi Beat,
This looks good - I've tested locally and the code makes sense.
Just need to fix the style-checks, line 377 is over 120 characters.

Would it be worth wrapping the model prediction in a try-except block and warning the user that a drop_last may be required if this error is caught (as the default is False)?
Something like:

# Perform prediction
try:
    model_outputs = self._model(i_batch)
except ValueError as e:
    if 'Expected more than 1 value per channel when training' in str(e):
        logger.exception("Try dropping the last incomplete batch by setting drop_last=True.")
    raise e

[kieranfraser]

mypy and pylint failing in CI / Style checks. The drop_last option should be included in PyTorchDeRandomizedSmoothing and PyTorchRandomizedSmoothing fit method too?

[beat-buesser]

I think the try-except block is an interesting idea. I'm running a few tests if it would slow down the model training.

[lgtm-com]

This pull request introduces 4 alerts and fixes 1 when merging 6e05685 into 3e3a438 - view on LGTM.com

new alerts:

• 4 for Unused import

fixed alerts:

• 1 for Module is imported more than once

[kieranfraser]

Looks good to me!