Faster R-CNN (WIP)

[fmassa]

This WIP PR implements Faster R-CNN based on https://github.com/rbgirshick/py-faster-rcnn .
There are lots of copy-paste from Girshick's repo, and a few files were copied as is.
I decided to fuse a number of operations together, but I'm not sure anymore it's the best way to go.

A few comments:

• There are two main classes: FasterRCNN and RPN. RPN implements the region proposal network, and faster r-cnn wraps almost all the detection logic (which might hurt flexibility at some point).
• Both classes behave differently if a ground-truth is provided, avoiding to have to write different network definitions for train/test.
• I tried to avoid a global cfg dict containing all the parameters, and instead pass the parameters as the constructors of the classes. But different train/test parameters still need to be defined in the model.
• There is currently no handling on the image scaling (originally present in im_info[2]), necessary for properly pruning small boxes in the RPN.
• I'll add optimized implementations for ROIPooling/nms/etc later on using cffi.

I'm opening this PR to get some feed-back on the general structure of the code.
Overall, I'm starting to think that it might be better to use the same structure as the one in py-faster-rcnn from Girshick repo.

Code cleanup and organization is required.

[colesbury]

This looks good. I think the most important thing is to get training (and evaluation) working and then worry more about design and cleaning up the code.

Some thoughts:

• I think copying from Ross's Faster-RCNN repo makes sense where appropriate. Keeping them in separate files, unchanged if possible, would be best
• There's a trade-off between making everything part of the model (nn.Container) and keeping it as part of the training function. For example, FasterRCNN.forward is a bit complicated because it combines training and evaluation logic: it has two possible return types and two possible input types.
• The forward() method in modules should take in Variables where appropriate; the wrapping of tensors in Variables should happen outside the module
• Try to match PyTorch style (PEP 8) for stuff that's not from RBG's repo. (4 space indent, etc.)
• Use Python argparse to configure options. I like how you're not passing a global cfg around. If you want to support a config file as well, I think you can do so with ConfigParser

[apaszke]

Don't we have to include the license file if we're including the original code?

[apaszke]

Also, there are a lot of commented statements that should be removed before merging

[fmassa]

@colesbury @apaszke thanks for your comments. I think the ConfigParser might be a nice way of addressing tons of arguments.

I initially wanted to keep the number of files small (as it is an example code), so I fused a number of things together, but that's probably a poor design choice.

I will validate that the basic code is working as expected by performing a training/evaluation and then I'll focus on getting a refactoring of this PR.

I'll get back to it on Monday, I've a trip to do in Rio tomorrow :)

[apaszke]

Ugh nvm, for some reason I haven't noticed that it's WIP...

Have a nice trip! 😃

[bhack]

Have you tested https://github.com/longcw/faster_rcnn_pytorch?
/cc @longcw

[fmassa]

@bhack I haven't, and I didn't have the time to finish this properly.
Given that there are already a number of pytorch implementations of object detection algorithms in pytorch, I'll close this one for the time being.
If I find some time to finish this up with a simple interface, I'll send a new PR.

[bhack]

@KaimingHe In the plan of releasing Mask r-cnn there will be also a faster-rcnn pytorch implementation merged in this repository?

[bhack]

A TF WIP Mask R-CNN effort was started at https://github.com/CharlesShang/FastMaskRCNN. Actually there is still no public reference implementation of the paper so we will see what kind of accurancy can be reproduced.

[bhack]

/cc @soumith for https://discuss.pytorch.org/t/deep-sharp-mask-or-mask-r-cnn/1469/4