Add force_data_batch_size to ReplayPlugin for manual assignment of data-mem ratio

[HamedHemati]

No description provided.

[ContinualAI-bot]

Oh no! It seems there are some PEP8 errors! 😕
Don't worry, you can fix them! 💪
Here's a report about the errors and where you can find them:

avalanche/benchmarks/utils/data_loader.py:265:81: E501 line too long (83 > 80 characters)
1       E501 line too long (83 > 80 characters)

[coveralls]

Pull Request Test Coverage Report for Build 1482890049

• 1 of 5 (20.0%) changed or added relevant lines in 2 files are covered.
• 3 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-0.02%) to 79.921%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
avalanche/benchmarks/utils/data_loader.py	0	4	0.0%

Files with Coverage Reduction	New Missed Lines	%
avalanche/benchmarks/utils/data_loader.py	1	89.68%
avalanche/benchmarks/scenarios/generic_definitions.py	2	85.37%

Totals
Change from base Build 1481778571:	-0.02%
Covered Lines:	10962
Relevant Lines:	13716

---

💛 - Coveralls

[AntonioCarta]

I see that GroupBalancedDataLoader and ReplayDataLoader have different behaviors regarding the batch_size. The first one creates num_groups dataloaders of batch_size samples, obtaining a final mini-batch of size num_groups * batch_size. The second one uses batch_size as the final size. I think we should make the consistent. Either GroupBalancedDataLoader uses batch_size as the total batch size, or ReplayDataLoader creates a minibatch of size 2 * batch_size. Maybe the first option is slightly more intuitive.

What do you think?

[HamedHemati]

What do you think?

@AntonioCarta I agree with you. I think keeping the batch size equal to a specified batch_size defined by the user makes it more intuitive because that's what one would normally expect (fixed instead of growing). Therefore, I think it's better to set the batch size in GroupBalancedDataLoader equal to batch_size / num_groups for each group; and obviously it should raise an error when batch_size < num_groups.

[HamedHemati]

Changes are applied to GroupBalancedDataloader to have a fixed batch size. It just checks whether batch_size > len(datasets) to ensure that each dataset gets at least one slot.

[AntonioCarta]

@HamedHemati I fixed the test that was not passing so that I can merge the PR.