[Feature] Support batch augmentation through BatchAugSampler

[gaotongxiao]

Motivation

Batch augmentation is a trick that improves the model's generalization ability. It's done by repeating the same data instance for several times and let each of them goes through parallel data augmentations in a single batch.

In MMOCR, the same effect can be achieved by using BatchAugSampler, which generates the same index for num_repeat times consecutively. And the final "dataset length" will be num_repeat times the original dataset length.

That is, a dataset comprising of [0, 1, 2, 3] may be randomly shuffled and sampled by BatchAugSampler and what dataloader finally gets could be [2, 2, 1, 1, 0, 0, 3, 3] when num_repeats=2.

It's also important to make batch size divisible by num_repeats, otherwise the repeated elements are not necessarily placed in the same batch.

Usage

1. Replace sampler in train_dataloader with BatchAugSampler, and specify num_repeats
2. Update batch size in train_dataloader, which should be divisible by num_repeats, e.g.

train_dataloader = dict(
    batch_size=4,
    sampler=dict(type='RepeatAugSampler', shuffle=True, num_repeats=2),
)

[codecov]

Codecov Report

Patch coverage: 75.60% and project coverage change: +2.04 🎉

Comparison is base (62d440f) 88.08% compared to head (7af3618) 90.12%.

❗ Current head 7af3618 differs from pull request most recent head c934e8f. Consider uploading reports for the commit c934e8f to get more accurate results

Additional details and impacted files

@@             Coverage Diff             @@
##           dev-1.x    #1757      +/-   ##
===========================================
+ Coverage    88.08%   90.12%   +2.04%     
===========================================
  Files          176      190      +14     
  Lines        11022    11124     +102     
  Branches      1558     1566       +8     
===========================================
+ Hits          9709    10026     +317     
+ Misses        1022      790     -232     
- Partials       291      308      +17     

Flag	Coverage Δ
unittests	90.12% <75.60%> (+2.04%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
...atasets/preparers/obtainers/naive_data_obtainer.py	16.82% <5.88%> (ø)
...r/datasets/preparers/packers/wildreceipt_packer.py	16.36% <16.36%> (ø)
mmocr/datasets/preparers/gatherers/naf_gatherer.py	31.42% <31.42%> (ø)
mmocr/datasets/preparers/parsers/coco_parser.py	18.00% <33.33%> (+0.69%)	⬆️
mmocr/datasets/preparers/parsers/base.py	80.00% <77.77%> (+2.22%)	⬆️
mmocr/utils/processing.py	79.31% <79.31%> (ø)
mmocr/datasets/preparers/dumpers/base.py	80.00% <80.00%> (ø)
mmocr/datasets/preparers/packers/base.py	81.25% <81.25%> (ø)
...ocr/datasets/preparers/packers/textrecog_packer.py	81.25% <81.25%> (ø)
mmocr/datasets/preparers/data_preparer.py	83.67% <81.81%> (+15.49%)	⬆️
... and 26 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.