transnorm's Introduction

TransNorm

Code release for "Transferable Normalization: Towards Improving Transferability of Deep Neural Networks" (NeurIPS 2019)

Prerequisites

PyTorch >= 0.4.0 (with suitable CUDA and CuDNN version)
torchvision >= 0.2.1
Python3
Numpy
argparse
PIL

Training

Office-31

pythonn train_image.py --gpu_id id --net ResNet50 --dset office --test_interval 500 --s_dset_path ../data/office/amazon_list.txt --t_dset_path ../data/office/webcam_list.txt

Office-Home

pythonn train_image.py --gpu_id id --net ResNet50 --dset office-home --test_interval 2000 --s_dset_path ../data/office-home/Art.txt --t_dset_path ../data/office-home/Clipart.txt

VisDA 2017

pythonn train_image.py --gpu_id id --net ResNet50 --dset visda --test_interval 5000 --s_dset_path ../data/visda-2017/train_list.txt --t_dset_path ../data/visda-2017/validation_list.txt

Image-clef

pythonn train_image.py --gpu_id id --net ResNet50 --dset image-clef --test_interval 500 --s_dset_path ../data/image-clef/b_list.txt --t_dset_path ../data/image-clef/i_list.txt

Acknowledgement

This code is implemented based on the published code of CDAN and BatchNorm, and it is our pleasure to acknowledge their contributions. CDAN (Conditional Adversarial Domain Adaptation) BatchNorm (Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift)

Citation

If you use this code for your research, please consider citing:

@inproceedings{Wang19TransNorm,
    title = {Transferable Normalization: Towards Improving Transferability of Deep Neural Networks},
    author = {Wang, Ximei and Jin, Ying and Long, Mingsheng and Wang, Jianmin and Jordan, Michael I},
    booktitle = {Advances in Neural Information Processing Systems 32},
    year = {2019}
}