Better Practices for Domain Adaptation

UDA

The main branch of this repo covers the UDA setting. Please find SFDA under the sfda branch and TTA under the tta branch.

Requirements

To install requirements:

pip install -r requirements.txt

Set environment variables

Set the following variables in your terminal:

export UDA_ROOT=<path-to-this-directory>

export DATA_ROOT=<path-to-data-directory>
export RESULTS_ROOT=$UDA_ROOT/results

Datasets

If only running the simpler MNIST setup for reproducibility, skip this section and go straight to the Source-only Training section.

To download datasets: Download VisDA2017 manually from here. Download Office-31 manually from here. Download Office-Home manually from here.

To generate MNIST-MR, run:

mkdir ${DATA_ROOT}/bsds500 && cd ${DATA_ROOT}/bsds500 && wget http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/BSR_bsds500.tgz && cd ${UDA_ROOT}
python generate_mnistmr.py

Next we generate the data split files for VisDA2017/Office-31/OfficeHome by running:

python split.py ${DATA_ROOT}/VisDA2017/classification train 0.6 0.2
python split.py ${DATA_ROOT}/VisDA2017/classification validation 0.6 0.2
python split.py ${DATA_ROOT}/VisDA2017/classification test 0.6 0.2

python split.py ${DATA_ROOT}/office31 amazon 0.6 0.2
python split.py ${DATA_ROOT}/office31 dslr 0.6 0.2
python split.py ${DATA_ROOT}/office31 webcam 0.6 0.2

python split.py ${DATA_ROOT}/officehome art 0.6 0.2
python split.py ${DATA_ROOT}/officehome clipart 0.6 0.2
python split.py ${DATA_ROOT}/officehome product 0.6 0.2
python split.py ${DATA_ROOT}/officehome real 0.6 0.2

When all is done, the datasets are expected in the following filestructure:

${DATA_ROOT}/
    mnist_m_r/
        ...
    VisDA2017/
        ...
    office31/
        ...
    officehome/
        ...

Source-only Training

For the training, we will run through a setup on MNIST-M. The other datasets can be used by specifying the right --dataset, --source and --target arguments.

To train 10 source-only models with different hyperparameters and select the checkpoint to use as initialisation for adaptation, run the following:

python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS-ROOT} --dataset mnistm --source mnist --target mnistm --algorithm source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100
python select_best_checkpoint ${RESULTS_ROOT} mnistm mnist mnistm source-only src_val_acc_score

The checkpoint should then exist in the following structure:

${RESULTS_ROOT}/
    mnistm/
        mnist/
            mnistm/
                source-only/
                    best.pt
                    ...

Adaptation

To run all the adaptation algorithms, using the generated source-only checkpoint as initialisation, run the following:

python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm atdoc --init-source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100
python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm bnm --init-source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100
python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm dann --init-source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100
python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm mcc --init-source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100
python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm mcd --init-source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100
python classification_benchmark.py --data-root ${DATA_ROOT} --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm mmd --init-source-only --G-arch mnistG --hpo-num-samples 10 --hpo-validate-freq 5 --hpo-max-epochs 100

Compute Validators

Next, we compute all validators for the algorithm checkpoints by running the following:

python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm source-only
python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm atdoc
python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm bnm
python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm dann
python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm mcc
python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm mcd
python compute_validators.py --results-root ${RESULTS_ROOT} --dataset mnistm --source mnist --target mnistm --algorithm mmd

Generate table

Open the jupyter notebook generate_table.py and run all cells. A latex table similar to table 3 will be generated, containing the results over the datasets used.

Acknowledgements

Our code is built upon the public code of the pytorch-adapt.

linusericsson/better-da