Skip to content

A More Robust Domain Feature Decoupling Network. <Huawei 2020 2nd Artificial Intelligence Innovation Competition>

Notifications You must be signed in to change notification settings

Complicateddd/R-DFDN

Repository files navigation

Introduction

This is a pytorch repository of method ‘A more robust domain feature decoupling network ’(R-DFDN), which is heavily followed by https://github.com/salesforce/corr_based_prediction。

Main contributer:Zechen. Zhao 、Shijie.Li 、Tian.Tian

Our model is more robust and stable compared with Corr-Prediction:

Model

Usage && Implement

Requirement:

Python3.6

Pytorch1.2

Numpy1.8

Tqdm

Run:

1、Generate Colored MNIST:

python gen_color_mnist.py

2、Run Correlation based regularization:

python main.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --save_dir corr --beta 0.1

3、Run existing regularization methods:

Maximum Likelihood Estimate (MLE):

python existing_methods.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --lr 0.0001 --bs 128 --save_dir mle

Adaptive Batch Normalization (AdaBN):

python existing_methods.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --lr 0.0001 --bs 32 --save_dir adabn --bn --bn_eval

Clean Logit Pairing (CLP):

python existing_methods.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --lr 0.0001 --save_dir clp --clp --beta 0.5

Projected Gradient Descent (PGD) based adversarial training:

python existing_methods.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --lr 0.0001 --save_dir pgd --pgd --nsteps 20 --stepsz 2 --epsilon 8

Variational Information Bottleneck (VIB):

python existing_methods.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --lr 0.001 --save_dir inp --inp_noise 0.2

4、Run our R-DFDN:

Commands to run R-DFDN:

python train_RDFDN.py --use tf_board True --epochs 300 

See the result in tensorboardX:

tensorboard

5、Run ADDA \ RevGrad \ WDGRL:

The same as a implement in https://github.com/jvanvugt/pytorch-domain-adaptation

Some result:

Evaluation on best model forom c-mnist

CMNIST(B) MNIST MNIST-M SVHN Absolute Gain
Corr-Prediction 96.88 85.94 79.69 43.75 +0.00
MLE 11.72 7.81 8.59 11.72 \
ADABN 12.5 16.41 12.5 9.38 \
CLP 22.66 7.81 9.38 13.28 \
PGD 18.75 13.28 8.59 7.03 \
VIB 16.41 13.28 11.72 11.72 \
R-DFDN(ours) 96.09 97.65 85.16 60.94 \
-0.78 +11.71 +5.47 +17.19 +33.58

About

A More Robust Domain Feature Decoupling Network. <Huawei 2020 2nd Artificial Intelligence Innovation Competition>

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages