- 03/04/2021: Configured training settings for Places_LT
- 01/05/2021: Update finetuning with knowledge transfer
- PyTorch (tested with version == 1.7.1)
-
Download the ImageNet_2014 and Places_365 (256x256 version). Change the
data_rootinmain.pyto your own directory. -
CIFAR dataset will be automatically downloaded when running CIFAR experiment.
- Baseline (with k=1)
python main.py --m_from 1
- cRT (Decoupling)
python main.py --config ./config/ImageNet_LT/stage_1_crt.py --path <folder path> --crt
- LLW:
python main.py --trainable_logits_weight --path <baseline checkpoint directory name> --scaling_logits --config ./config/ImageNet_LT/weight_finetune_resnet10.py --merge_logits
- Knowledge Transfer Loss:
python main.py --second_dotproduct --m_from 1 --scaling_logits --second_head_alpha 0.2 --temperature 1