[Improvement] Improve digit_version & use it for version_checking

[kennymckormick]

No description provided.

[ZwwWayne]

need to resolve comments and fix CI

[codecov]

Codecov Report

Merging #1185 (cfa417a) into master (6659c38) will increase coverage by 0.28%.
The diff coverage is 66.12%.

@@            Coverage Diff             @@
##           master    #1185      +/-   ##
==========================================
+ Coverage   68.00%   68.28%   +0.28%     
==========================================
  Files         160      160              
  Lines       10442    10519      +77     
  Branches     1902     1918      +16     
==========================================
+ Hits         7101     7183      +82     
+ Misses       2968     2958      -10     
- Partials      373      378       +5     

Flag	Coverage Δ
unittests	68.28% <66.12%> (+0.28%)	⬆️

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

Impacted Files	Coverage Δ
mmcv/ops/saconv.py	83.33% <20.00%> (-0.25%)	⬇️
mmcv/parallel/distributed.py	20.83% <20.00%> (-1.62%)	⬇️
mmcv/runner/fp16_utils.py	58.12% <33.33%> (-0.27%)	⬇️
mmcv/parallel/distributed_deprecated.py	66.66% <50.00%> (-0.73%)	⬇️
mmcv/runner/dist_utils.py	50.00% <50.00%> (-0.50%)	⬇️
mmcv/runner/hooks/logger/tensorboard.py	34.37% <50.00%> (-1.99%)	⬇️
mmcv/utils/parrots_wrapper.py	59.37% <50.00%> (ø)
mmcv/utils/trace.py	54.54% <50.00%> (ø)
mmcv/version.py	82.35% <78.57%> (-9.96%)	⬇️
mmcv/utils/version_utils.py	68.08% <86.95%> (+6.32%)	⬆️
... and 9 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 6659c38...cfa417a. Read the comment docs.

[kennymckormick]

@ZwwWayne CI passed.

[zhouzaida]

mmcv/mmcv/utils/trace.py

Line 8 in 366c628

if LooseVersion(torch.__version__) >= LooseVersion('1.6.0'):

This line should also be replaced with digit_version

[kennymckormick]

mmcv/mmcv/utils/trace.py

Line 8 in 366c628

if LooseVersion(torch.__version__) >= LooseVersion('1.6.0'):

This line should also be replaced with digit_version

resolved

[zhouzaida]

LGTM

[kennymckormick]

@zhouzaida

[kennymckormick]

@ZwwWayne

[innerlee]

My sole interest is the test cases. They should be as complete as possible. Here are some idea sources:

https://github.com/JuliaInterop/VersionParsing.jl/blob/master/test/runtests.jl
https://github.com/eieste/VersionParser/blob/master/version_parser/test.py
https://github.com/pypa/packaging/blob/main/tests/test_version.py

[kennymckormick]

My sole interest is the test cases. They should be as complete as possible. Here are some idea sources:

https://github.com/JuliaInterop/VersionParsing.jl/blob/master/test/runtests.jl
https://github.com/eieste/VersionParser/blob/master/version_parser/test.py
https://github.com/pypa/packaging/blob/main/tests/test_version.py

Since we use the parse_version API from pkg_resources, the capability of digit_version is not unlimited. For example, it can not parse sth. like 'vM1m2b3' and will treat '0.99.1+1' like '0.99.1'. However, I think can parse versions of most python packages correctly (it is mainly used to parse the version of pytorch or MM codebases). The test cases added define the boundary of its capability.

[zhiqi-li]

Can mmcv and mmegine improve the ability of digit_version function to handle custom compiled torch version?

[kennymckormick]

Can mmcv and mmegine improve the ability of digit_version function to handle custom compiled torch version?

That might be difficult because you need the entire commit history to make it possible. Besides, there is no guarantee that all commits are in linear (actually they can be parallel, it's not possible to determine the ranking in that case).