1061136002 / secure-muda

Code for SMUDA

Code for Secure Domain Adaptation with Multiple Sources

We attach the code for Secure Domain Adaptation with Multiple Sources. The code has been tested with PyTorch 1.7.1+cu110 and Python 3.8.

Dataset preparation

The repository comes with pre-trained features already made available in the data folder. If you prefer to generate these from scratch, please refer to the README in the data folder.

Additional datasets

If you wish to evaluate the code on a new dataset, you would need to add it to the data folder following the schematic described there, and changing the config.py and config_populate.py files in the code/train_code folder accordingly.

Configuring the run

Before running the model, runtime parameters can be configured via the config.py file.

Running the code

The code assumes the presence of a GPU. To run the code on a specific task pertaining to one of the four supported datasets, go to code/train_code and call the runner.py script. For example, to run the DW_A task for office-31, call

	python runner.py --dataset office-31 --task DW_A --gpu_id 0 --num_exp 1

The above command will run the specified task once on GPU 0. To re-run the task multiple times, simply specify a larger number of experiments via the num_exp flag.

After finishing a run, results are stored in the summaries and weights folders in the repository root. Each experiment will have a results.txt file, whose output format is described in the evaluate function in main.py (the score based on SWD aggregation is the second one in the file).

Multi-gpu training

It is possible to run several tasks in parallel if more than 1 GPUs are present on the system. You will need to use the run_pipeline.py script in the code folder. For example, running all the tasks for office-31 exactly once on 3 GPUs (1 task on each GPU), simply run

	python run_pipeline.py --dataset office-31 --num_exp 1 --gpu_id -1

Acknowledgements

We thank the authors of "Your Classifier can Secretly Suffice Multi-Source Domain Adaptation" for making their codebase available, as we use it as backbone for our codebase. We use the implementation if Sliced Wasserstein Distance from "Sliced Wasserstein Autoencoder". Finally, the code for generating pre-trained features is adapted from "Transfer Learning", a repository containing many useful reseach tools for this area.