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：
Python3.6
Pytorch1.2
Numpy1.8
Tqdm
python gen_color_mnist.pypython main.py --dataset fgbg_cmnist_cpr0.5-0.5 --seed 0 --root_dir cmnist --save_dir corr --beta 0.1Maximum 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 mleAdaptive 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_evalClean 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 0.5Projected 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 8Variational 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 0.2Commands to run R-DFDN:
python train_RDFDN.py --use tf_board True --epochs 300 See the result in tensorboardX:
The same as a implement in https://github.com/jvanvugt/pytorch-domain-adaptation
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 |