Enhance ShufflenetV2

[barrh]

Class shufflenetv2 receives stages_repeats and stages_out_channels arguments.

Following discussion in: #889

[ekagra-ranjan]

Both #889 and #892 address different issues as discussed in #889 although the following changes are needed to remove the getpaperParam function completely.

[ekagra-ranjan]

Can you pass the stage_repeats argument as a list of int as suggested by @fmassa in this comment? Doing so would remove the need of the function getpaperParam completely when both #889 and #892 are merged.

[barrh]

I think it would be nice to expose to the user paper's parameters, either by staticmethod, stand-alone function, or documentation.

[ekagra-ranjan]

I don't know whether it would be better. @fmassa could guide you on this. But the above requested changes will make the implementation similar to that of ResNets.

[fmassa]

I think that I prefer the approach that is followed by resnet

vision/torchvision/models/resnet.py

Line 256 in 0564df4

return _resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress,

Paper parameters are specified in the functions (and defines the meaning of 0_5 / 1_0 etc), but the base class should be agnostic to it if possible.

[fmassa]

@barrh this overlaps a bit with #889 . I'll be merging #889 as it addresses all the points I had, can you then rebase on top of master once it's merged?

[codecov-io]

Codecov Report

Merging #892 into master will decrease coverage by 0.24%.
The diff coverage is 73.33%.

@@            Coverage Diff            @@
##           master    #892      +/-   ##
=========================================
- Coverage   56.95%   56.7%   -0.25%     
=========================================
  Files          43      43              
  Lines        3559    3562       +3     
  Branches      544     546       +2     
=========================================
- Hits         2027    2020       -7     
- Misses       1414    1420       +6     
- Partials      118     122       +4

Impacted Files	Coverage Δ
torchvision/models/shufflenetv2.py	86.04% <73.33%> (-4.32%)	⬇️
torchvision/transforms/transforms.py	81.25% <0%> (-1.3%)	⬇️

---

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 dc3ac29...93ab83e. Read the comment docs.

[barrh]

@fmassa sure.

[barrh]

rebased. I didn't test it because python setup.py install behaves funny following the c++ patch.

[fmassa]

@barrh for easy developing Python code which contains C++ extensions, I generally do python setup.py build develop, which installs locally, so that the python changes are applied immediately and there is no need to recompile

[fmassa]

Thanks a lot!