[Feature] Repeated Aug and Sampler Registry

[0x4f5da2]

Motivation

RepAug is a necessary part to re-implement 'ResNet Strikes Back'

Modification

• implement repeated aug

Use cases

Just add sampler = dict(type="RepeatAugSampler") to the config file

[mzr1996]

Can we use a dataset wrapper to implement this function?

[0x4f5da2]

Can we use a dataset wrapper to implement this function?

@mzr1996 Probably not. Repeated Augmentation requires the repeated samples co-exist in the same batch while the Dataset class is epoch agnostic. In this way, we have to use un-shuffled data for training or implement the shuffle logic in the wrapper and use the same shuffled sequence for all epochs.

[codecov]

Codecov Report

Merging #588 (36dc51b) into master (894a82e) will increase coverage by 1.22%.
The diff coverage is 80.82%.

@@            Coverage Diff             @@
##           master     #588      +/-   ##
==========================================
+ Coverage   80.11%   81.33%   +1.22%     
==========================================
  Files         113      116       +3     
  Lines        6526     6720     +194     
  Branches     1120     1156      +36     
==========================================
+ Hits         5228     5466     +238     
+ Misses       1153     1103      -50     
- Partials      145      151       +6     

Flag	Coverage Δ
unittests	81.33% <80.82%> (+1.22%)	⬆️

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

Impacted Files	Coverage Δ
mmcls/datasets/builder.py	46.15% <55.55%> (+3.60%)	⬆️
mmcls/apis/train.py	27.77% <72.72%> (+6.25%)	⬆️
mmcls/datasets/samplers/repeat_aug.py	85.10% <85.10%> (ø)
mmcls/datasets/__init__.py	100.00% <100.00%> (ø)
mmcls/datasets/samplers/__init__.py	100.00% <100.00%> (ø)
mmcls/datasets/samplers/distributed_sampler.py	28.00% <100.00%> (+6.26%)	⬆️
mmcls/models/classifiers/image.py	84.88% <0.00%> (-1.96%)	⬇️
mmcls/models/heads/__init__.py	100.00% <0.00%> (ø)
mmcls/models/backbones/__init__.py	100.00% <0.00%> (ø)
mmcls/models/backbones/deit.py	97.50% <0.00%> (ø)
... and 10 more

---

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 6f25beb...36dc51b. Read the comment docs.

[mzr1996]

LGTM