A Bit More Bayesian: Domain-Invariant Learning with Uncertainty

Code for paper "A Bit More Bayesian: Domain-Invariant Learning with Uncertainty" submitted to ICML 2021.

Prerequisites

• Python 3.6.9
• Pytorch 1.1.0

Components

• ../kfold/: Directory of images
• ../files/: Directory of the train/validation/test split txt files
• pacs_main.py: script to run classification experiments on PACS
• pacs_model.py: the model used in pacs_main.py
• pacs_datas.py: script to load data from PACS for the experiments
• pacs_test.py: script to evaluate the trained model
• ./logs/: folder to store the trained model
• augs.py: data augmentation functions for the experiments
• utils.py: assorted functions to support the repository

Setup

The code is for the PACS dataset. Download the datasets from the following link (or the link in the main paper), extract the compressed file, and place the images in ../kfold/ directory and the train/validation/test split txt files in the ../files/ directory.

[Google Drive]

Training

For training the model run the following:

python pacs_main.py --test_domain cartoon --log_dir model_name

Change the cartoon after --test_domain to art_painting, photo or sketch to change the target domain. Use --classifier SGP/NO to choose Bayesian invariant classifier or deterministic classifier. The default value is --classifier SGP. Use --feature bayes/no to choose Bayesian invariant feature extractor or deterministic one. The default value is --feature bayes. Use --classifier NO --feature no to train the baseline method. The trained model and logs will be stored in ./logs/model_name/

Evaluation

For evaluation of the trained model run the following:

python pacs_test.py --test_domain cartoon --log_dir cartoon_model

Change the target domain as the training phase. Change the cartoon_model to other names to evaluate other trained models