Project of Computer Vision 2018 NYU course.
I mainly implemented GrouPy and Tund part of this project. Work done with Yu Cao: https://github.com/Yucao42/Galaxy_Zoo
We achieve test score of 0.07520 with Resnet18 and 0.07669 with GrouPy. Averaged model achieved 0.07484, beyond SOTA performance of 0.07491.
The original kaggle challenage is at: https://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge/leaderboard
This branch is trying to implement a Group Equivariant CNN with the idea from this paper: http://proceedings.mlr.press/v48/cohenc16.pdf
The PyTorch version of GrouPy is from: https://github.com/adambielski/GrouPy
Environment requirements except GrouPy are included in requirements.yaml. GrouPy's setup process is provided in the last link above.
To run training module of our work, use train*.sh in ./shell:
source activate galaxy1 // activate conda env
bash train*.sh // pick the model with your expectation, and remember the MODEL setting
To run evaluation module, use eval.sh in ./shell:
source activate galaxy1
bash eval.sh // remember the MODEL setting corresponding to your training model