This is a third party implementation of RA-CNN in pytorch. I am still working on reproducing a same performance written in paper

• python3
• Pytorch 0.4 >
• torchvision
• numpy
• tensorflow, optional

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

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

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

For training, use following command.

$ python trainer.py

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.

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