jeong-tae / Deep-COOC.pytorch

Implementation of Deep-COOC pytorch version

Deep-COOC.pytorch

Implementation of Deep-COOC pytorch version

Requirements

• python3
• Pytorch 0.4
• torchvision
• numpy
• tensorboard, tensorboard only would be enough

TODO

• Reproduce same results
• Combine many modules
• Try with other dataset

Current status

• Model can't reproduce the results that described in the paper.
• I think there is a lot of details like weights freezing, learning rate, decays, but I can't follows all the stuff.
• There is a lots of things you can do to improve. Do tuning!

Performance

• Sometime Deep-cooc model well perform than benchmarks, it depends on hyperparams.
• They saids that they've got 73.3% accuracy for ResNet-152 but I've got 83.14% after carefully tuning the model.
• For now, I've got 83.4% accuracy for Deep-cooc model with some trials. Still about 2% improvements is needed to reproduce

Usage

• For training a deepcooc model, do ./deepcooc_CUB2011.sh
• You can modify some hyperparam in deepcooc_CUB2011.sh or you can also add new params. To see more details, refer train.py