polinakirichenko/deep_feature_reweighting

deep_feature_reweighting's Issues

Exact command to obtain base waterbirds model

I'm trying to reproduce the 74.9% worst case and 98.1% average test accuracy on Waterbirds, but could not do so. I'm running the command below.

python3 train_classifier.py --output_dir=<OUTPUT_DIR> --pretrained_model \
  --num_epochs=100 --weight_decay=1e-3 --batch_size=32 --init_lr=1e-3 \
  --eval_freq=1 --data_dir=<WATERBIRDS_DIR> --test_wb_dir=<WATERBIRDS_DIR> \
  --augment_data --seed=<SEED> --num_minority_groups_remove=0

Base Model Checkpoints

Dear authors,

Could you provide the checkpoints (i.e., saved weights) of the base models used in your paper? I run your commands on CelebA & Waterbirds (for 5 random seeds), and the performance of base models & DFR on top of these base models is slightly worse than that reported in your paper. Thus, I want to request your trained base models for an exact reproduction & further comparison. I would highly appreciate it if you could provide a downloadable link to a Dropbox/Box/Google Drive folder containing your trained models. Thanks!

@andrewgordonwilson @PolinaKirichenko @izmailovpavel