Awesome

Official implementation for PaCDA

[CLVision Workshop - CVPR 2023] Continual Domain Adaptation Through Pruning-Aided Domain-Specific Weight Modulation

Commands for running

General instructions

Within each training .py file, modify lines enclosed within #TODO and the list variable "names" appropriately before each run

Baseline

Source train

python image_source.py --trte val --da uda --output ./debug_bl_1/ACPR --dset office-home --max_epoch 100

Target adaptation

python image_target_train.py --dset office-home --output ./debug_bl_1/ACPR --output_src ./debug_bl_1/ACPR

Ours

Source train

python image_source.py --trte val --da uda --output ./debug_1/ACPR --dset office-home --max_epoch 100

Source prune and finetune

Set --pf_c to be the desired fraction of params of each layer in source model to be pruned

python image_prune_source_finetune.py --trte val --da uda --output ./debug_1/ACPR --dset office-home --pf_c 0.4 --output_src ./debug_1/ACPR

Target train

python image_prune_target_train.py --dset domain-net --output ./debug_1/ACPR --output_src ./debug_1/ACPR

Target prune and finetune

Set --pf_c to be the desired fraction of params of each layer in train model to be pruned

python image_prune_target_finetune.py --pf_c 0.3 --output ./debug_1/ACPR --output_src ./debug_1/ACPR --dset office-home

Acknowledgement

The code is based on SHOT (ICML, 2020)

Citation

If you find this work useful, please feel free to cite this work.

@InProceedings{B_2023_CVPR,
    author    = {B, Prasanna and Sanyal, Sunandini and Babu, R. Venkatesh},
    title     = {Continual Domain Adaptation Through Pruning-Aided Domain-Specific Weight Modulation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {2456-2462}
}