TAST

This code is for the paper "Test-time Adaptation via Self-training with Nearest Neighbor information (TAST)", accetped to ICLR'23. We use the publicly released code "https://github.com/matsuolab/T3A". You can follow the descriptions about installation and experiments in "https://github.com/matsuolab/T3A".

Dependencies

Python 3.7.11 PyTorch 1.8.0 Torchvision 0.9.0 CUDA 10.2 CUDNN 7605 NumPy 1.2 PIL 8.4.0

Data

You can download the domain generalization benchmarks, namely VLCS, PACS, OfficeHome, and TerraIncognita by following the procedure. e.g.) python -m domainbed.scripts.download --data_dir=/my/datasets/path --dataset pacs

You can change pacs to vlcs, office_home, terra_incognita to download other datasets.

Train

You can train a model on training domains. e.g.) python -m domainbed.scripts.train
--data_dir /my/datasets/path
--output_dir /my/pretrain/path
--algorithm ERM
--dataset PACS
--hparams "{"backbone": "resnet18-BN"}"

You can use backbone networks such as resnet50-BN, resnet50 which are presented in the train.py file. The trained network and information about the training are recorded in "/my/pretrain/path"

Test-time adaptation

While testing, we adapt trained classifiers. e.g.) python -m domainbed.scripts.unsupervised_adaptation
--input_dir=/my/pretrain/path
--adapt_algorithm=TAST

You can use the test-time adaptation algorithms such as T3A, TAST, and TAST_bn which are presented in the adapt_algorithms.py file.

Then, the test reulsts will be recorded in "/my/pretrain/path/out_TAST.txt" and "/my/pretrain/path/results_TAST.jsonl"