iamaaditya / repr-ortho

Repr training and inter-filter ranking

repr-ortho

Repr training and inter-filter ranking

Data

https://drive.google.com/open?id=1q-Hfm7vKMi-7M0NUaS3VmJ1yqzlZrP-d Checksum: 3914026746

To train with repr: python train.py To train with standard: python train.py --standard.

Notes

1. Baseline training accuracy overfits, this is because RePr was designed to improve upon overfitting of small and vanilla conv-nets.
2. Baseline test accuracy might appear low. It is because of two reasons: (i) There is no augmentation. RePr was designed to show overfitting of even small network (ii) Only uses 90% of the training data. Other is used for validation.