RACNN-pytorch

This is a third party implementation of RA-CNN in pytorch. I am still working on reproducing a same performance written in paper
You can download CUB200 dataset from this page
and un-compress using this command tar -xvf CUB_200_2011.tgz -C data/

Requirements

• python3
• Pytorch 1.2
• torchvision
• numpy
• tensorflow, optional

TODO

• [x] Network building
• [ ] Repactoring for arguments
• [ ] Pre-training a APN
• [ ] Alternative training between APN and ConvNet/Classifier
• [ ] Reproduce and report on README.md
• [ ] Sample visualization
  • Followed this impl
• [ ] Add new approach to improve

Current issue

• Don't know how to pre-train a APN. Need more details
• Rankloss doesn't decrease. Because no pretrain? or bugs?

Results

Current best is 71.68% at scale1 without APN pretraining. It's bad than using just VGG19

Usage

For training, use following command.

$ python trainer.py

or

$ ./train.sh

Currently only cuda available device support.

If you want to see training process,

$ Tensorboard --log='visual/' --port=6666

and go to 'localhost:6666' on webbrowser. You can see the Loss, Acc and so on.

References

• Original code
• Other pytorch implementation
  • with car dataset, I refer the attention crop code from here