ICCV21: Deep Co-Training with Task Decomposition for Semi-Supervised Domain Adaptation

Requirements

The code is developed under Python 3.6.5 and PyTorch 1.4.0

To install, pip install -r requirements.txt

Prepare dataset

To reproduce the DomainNet results, download DomainNet from http://ai.bu.edu/M3SDA/ following the instructions on the page.

Your dataset root is expected to contain folders named after all the domains, for example:

PATH-TO-DATASET-ROOT/clipart

Train your own model

There are seven adaptation scenarios in the DomainNet experiment.

Specify the Source and Target domain by either

--source X --target Y or --st X_Y_index

Example to reproduce the DomainNet 3-shot result, from Real to Clipart (X_Y_index=1), saving checkpoint:

python main.py --root PATH-TO-DATASET-ROOT/ --st 1 --save_check

Results will be saved at ./record/multi/mico Checkpoint will be saved at ./checkpoint

Eval

Example to evaluate a saved model at iteration 20000 from ./checkpoint, from Real to Clipart:

python main.py --root PATH-TO-DATASET-ROOT/ --st 1 --eval --net_resume Net_iter_model_mico_real_to_clipart_step_20000.pth.tar