Add dilation option to ResNet

[fmassa]

This PR adds support for replacing 2x2 strides in ResNet with dilation.

This functionality is useful for segmentation models, where high-resolution feature maps are desirable.

Instead of hard-coding the resolution of the feature map, we let the user specify explicitly which blocks they want the 2x2 stride to be replaced by a dilated convolution.
Given that there are only 3 blocks with 2x2 strides (if we don't consider the initial convolution, which also performs downsampling), the user needs to specify a 3-element list of booleans.

Suggestions for a better naming (instead of replace_stride_with_dilation) are welcome.

cc @aadcock

[codecov-io]

Codecov Report

Merging #866 into master will increase coverage by 0.09%.
The diff coverage is 80%.

@@            Coverage Diff             @@
##           master     #866      +/-   ##
==========================================
+ Coverage   54.49%   54.58%   +0.09%     
==========================================
  Files          36       36              
  Lines        3307     3318      +11     
  Branches      542      545       +3     
==========================================
+ Hits         1802     1811       +9     
- Misses       1372     1373       +1     
- Partials      133      134       +1

Impacted Files	Coverage Δ
torchvision/models/resnet.py	83.43% <80%> (-0.12%)	⬇️

---

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 ccbb322...298a5df. Read the comment docs.

[fmassa]

I'm merging this, but let me know if anyone disagrees with something or has a better idea on how to improve this part.

[labor00]

This change is backward compatible? For now I have been using the torch.nn.apply function to directly change the parametrization of the convolution layers, replacing padding, strides and dilation.

[fmassa]

@labor00 yes, this change is backwards-compatible.

Old models and code still work the same as before, because we do not modify the strides / dilations if the parameters are the default (None)

[usdflt]

The repository for the gluoncv is very well documented and have lots of tests. i believe that some of the models could be ported to pytorch without hurdle.

[fmassa]

@usdflt what models do you have in mind? Note that we are adding segmentation and detection models to torchvision soon.

[usdflt]

@fmassa I was thinking of models like pspnet, yolo-v*, ssd. Is there any plans to merge the maskrccn repo within the torch vision?

[fmassa]

@usdflt the plan is to include some simplified version of faster r-cnn and mask r-cnn in torchvision, together with FCN and DeepLabV3