This repository contains the code of our SDM20 paper "Fisher Deep Domain Adaptation".
- PyTorch >= 0.4.0
- Python3
Two domain adaptation benchmarks, Office-31 and Office-Home, are used.
Each benchmark has several domains. For each domain, please create a txt file that lists the image file names and their labels. Each line in the txt file is structured as "image_file_path label".
An example of the Office-31 dataset is provided in the data/office folder. You may replace the [local_path] in the txt files with the path where you save the dataset.
Three transfer criterion, Maximum Mean Discrepancy (MMD), deep CORrelation ALignment (CORAL), and domain adversarial loss, are used in the experiment. Use the train_ada.py script if you use the domain adversarial loss. Use the train_feature.py if you use the MMD or CORAL loss.
An example of training with the MMD loss on the Amazon -> DSLR task is given in the following.
python train_feature.py --gpu_id <gpu_id> \
--net ResNet50 \
--dset office \
--test_interval 500 \
--snapshot_interval 10000 \
--ly_type cosine \
--loss_type mmd \
--fisher_loss_type tr \
--output_dir <output_path> \
--s_dset_path ../data/office31/amazon_31_list.txt \
--t_dset_path ../data/office31/dslr_31_list.txt \
--em_loss_coef 0.1 \
--inter_loss_coef 1. \
--intra_loss_coef 1. \
--trade_off 1.
The weight of the transfer criteria is controlled by the trade_off argument. If you do not want to use a transfer criteria, please set the trade_off argument to 0.
The Fisher loss has two instantiations, Trace Ratio and Trace Difference. The instantiation is determined by the fisher_loss_type argument. The within-class and between-class penalty are controlled by the intra_loss_coef and inter_loss_coef, respectively.
To evaluate the model, the following command can be used.
python eval_da.py --gpu_id <gpu_id> \
--net ResNet50 \
--dset office \
--ly_type cosine \
--ckpt_path <path of the saved model> \
--s_dset_path ../data/office31/amazon_31_list.txt \
--t_dset_path ../data/office31/dslr_31_list.txt
If you use this code, please consider citing our paper:
Yinghua Zhang, Yu Zhang, Ying Wei, Kun Bai, Yangqiu Song, Qiang Yang. Fisher Deep Domain Adaptation. In: Proceedings of SIAM International Conference on Data Mining (SDM), Cincinnati, Ohio, USA, 2020.