[Feature] Support FarSeg (CVPR'2020)

[MengzhangLI]

Motivation

Support FarSeg, one of the benchmark in Geospatial Object Segmentation.
Official repo: https://github.com/Z-Zheng/FarSeg.
Original paper: https://arxiv.org/abs/2011.09766.

Current results:

FarSeg setting is

lr=0.007, weight_decay=0.0001, min_lr=1e-5

Original setting is MMSeg default setting:

lr=0.01, weight_decay=0.0005, min_lr=1e-4

	backbone	Batch size	loss	Hyper parameters	mIoU
farseg_r50-d8_4x2_896x896_80k_isaid.py	r50-d8	4x2	CE	FarSeg setting	61.04
farseg_r50-d8_4x2_896x896_80k_isaid.py	r50-d8	4x2	Annealing Soft Focal	FarSeg setting	60.65
farseg_r50-d8_4x4_896x896_80k_isaid.py	r50-d8	4x4	CE	FarSeg setting	61.61
pspnet_r50-d8_4x4_896x896_80k_isaid.py (mmseg current model)	r50-d8	4x4	CE	Original setting	65.36
farseg_r50-d32_4x2_896x896_80k_isaid.py	r50-d32	4x2	CE	FarSeg setting	59.21
farseg_r50-d32_4x2_896x896_80k_isaid.py (FarSeg original model)	r50-d32	4x2	Annealing Soft Focal	FarSeg setting	62.72
farseg_r50-d32_4x4_896x896_80k_isaid.py	r50-d32	4x4	CE	FarSeg setting	60.09
pspnet_r50-d32_4x4_896x896_80k_isaid.py	r50-d32	4x4	CE	FarSeg setting	55.08

[MengzhangLI]

The inference has been aligned, which uses pretrained weight from official repo.

[codecov]

Codecov Report

Attention: Patch coverage is 95.20000% with 6 lines in your changes missing coverage. Please review.

Project coverage is 90.49%. Comparing base (17b500f) to head (ddbb1ec).
Report is 186 commits behind head on master.

Files	Patch %	Lines
mmseg/models/necks/far_neck.py	92.45%	3 Missing and 1 partial ⚠️
mmseg/models/decode_heads/far_head.py	97.10%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1378      +/-   ##
==========================================
+ Coverage   90.42%   90.49%   +0.07%     
==========================================
  Files         133      135       +2     
  Lines        7902     8026     +124     
  Branches     1318     1335      +17     
==========================================
+ Hits         7145     7263     +118     
- Misses        534      538       +4     
- Partials      223      225       +2     

Flag	Coverage Δ
unittests	90.49% <95.20%> (+0.07%)	⬆️

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

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.