[Feature] Support VAN

[okotaku]

Motivation

Visual Attention Network
official repo

Result

CUB200
bs8×1GPU

model	top1 acc
van-l	89.54

Config

_base_ = [
    '../_base_/models/van/van_large.py', '../_base_/datasets/cub_bs8_384.py',
    '../_base_/schedules/cub_bs64.py', '../_base_/default_runtime.py'
]

checkpoint = 'van_large.pth'  # converted weight from official repo

model = dict(
    type='ImageClassifier',
    backbone=dict(
        init_cfg=dict(
            type='Pretrained', checkpoint=checkpoint, prefix='backbone')),
    head=dict(num_classes=200, ))

paramwise_cfg = dict(
    norm_decay_mult=0.0,
    bias_decay_mult=0.0,
    custom_keys={
        '.absolute_pos_embed': dict(decay_mult=0.0),
        '.relative_position_bias_table': dict(decay_mult=0.0)
    })

optimizer = dict(
    _delete_=True,
    type='AdamW',
    lr=5e-5,
    weight_decay=0.0005,
    eps=1e-8,
    betas=(0.9, 0.999),
    paramwise_cfg=paramwise_cfg)
optimizer_config = dict(grad_clip=dict(max_norm=5.0), _delete_=True)

log_config = dict(interval=20)  # log every 20 intervals

checkpoint_config = dict(
    interval=1, max_keep_ckpts=3)  # save last three checkpoints
fp16 = dict(loss_scale='dynamic')

Checklist

Before PR:

• Pre-commit or other linting tools are used to fix the potential lint issues.
• Bug fixes are fully covered by unit tests, the case that causes the bug should be added in the unit tests.
• The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
• The documentation has been modified accordingly, like docstring or example tutorials.

After PR:

• If the modification has potential influence on downstream or other related projects, this PR should be tested with those projects, like MMDet or MMSeg.
• CLA has been signed and all committers have signed the CLA in this PR.

[codecov]

Codecov Report

Merging #739 (8e11e3b) into dev (6beac50) will increase coverage by 0.21%.
The diff coverage is 96.85%.

@@            Coverage Diff             @@
##              dev     #739      +/-   ##
==========================================
+ Coverage   86.68%   86.90%   +0.21%     
==========================================
  Files         128      129       +1     
  Lines        8255     8414     +159     
  Branches     1422     1446      +24     
==========================================
+ Hits         7156     7312     +156     
- Misses        885      886       +1     
- Partials      214      216       +2     

Flag	Coverage Δ
unittests	86.81% <96.85%> (+0.19%)	⬆️

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

Impacted Files	Coverage Δ
mmcls/models/backbones/van.py	96.83% <96.83%> (ø)
mmcls/models/backbones/__init__.py	100.00% <100.00%> (ø)
mmcls/datasets/builder.py	91.04% <0.00%> (+1.49%)	⬆️
mmcls/utils/setup_env.py	100.00% <0.00%> (+4.54%)	⬆️

---

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 6beac50...8e11e3b. Read the comment docs.

[Ezra-Yu]

Thank you for your great contribution, please fix the conflicts!

Could you provide us with the converted checkpoints for imagenet-1k dataset? you can refer to the convert-scripts like twins2mmcls. The weights can be uploaded to Google Driver or TmpLink or just provide the convert script.

[okotaku]

@Ezra-Yu
http://tmp.link/f/624a4df33256d
http://tmp.link/f/624a4e02a47eb
http://tmp.link/f/624a4e0b28ba8
http://tmp.link/f/624a4e0dbbfc1

[Ezra-Yu]

1. Please fix the conficts
2. try to use mmcv config init

i will complete the checkpoints and docs including metafile later.

[CLAassistant]

All committers have signed the CLA.

[ZwwWayne]

Would be great if the PR message could also show the example config in downstream repos like mmseg/mmdet

[Ezra-Yu]

LGTM.