[NeurIPS2016] Domain Separation Network

Notes

This is a Pytorch implementation of the Domain Separation Network Arxiv. There have been multiple implementations of this article on GitHub, but most of them are based on TensorFlow with poor readability. The few PyTorch-based implementations have extremely brief documentation, making it difficult to start training and reproduce the metrics in the paper. This project provides a PyTorch implementation with clear steps, facilitating everyone to reproduce the results.

Get Started

1. Please follow these steps to create environment.

a. Create a conda virtual environment and activate it.

conda create -n dsn python=3.7 -y
conda activate dsn

b. Install other packets as followings.

• numpy 1.21.5
• protobuf 3.20.0
• scikit-image 0.19.3
• tensorboard 1.15.0
• tensorflow 1.15.0
• tensorflow-estimator 1.15.1
• torch 1.12.1
• torchvision 0.13.1

The main purpose of installing TensorFlow here is to leverage its interfaces for downloading and loading the MNIST dataset, while the network training and testing are completely implemented using PyTorch. It should also be noted that the versions of TensorFlow and Protobuf need to be configured as mentioned above, otherwise these two packages are prone to conflicting errors.

2. Download BSR_bsds500.tgz and place it in the main directory.

3. Generate training and testing datasets.

python create_mnistm.py

The script will automatically create a data directory and generate mnist_data_label.hkl, mnist_data.hkl, and mnistm_data.hkl in this directory for training and testing.

|---data
|   |---mnist_data_label.hkl        # MNIST labels（training + test, one-hot, shared with MNIST-M）
|   |---mnist_data.hkl              # MNIST images（training + test）
|   |---mnistm_data.hkl             # MNIST-M images（traning + test + valid）

4. Start Training and get the training.log.

python model.py

Note that this model is very sensitive to the batch_size, our implementation cannot perform as perfect as the original paper, so be careful when you tune parameters for other datasets.

Result

We only conduct the experiments from mnist to mnist_m, the target accuracy of our implementation is about 85.8% (original paper ~83.2%).