minDG (Coming Soon)

A minimal DomainBed implementation, aims for fast experiment results.

What's up with DomainBed?

DomainBed is a great library for domain generalization research. However, the training and hyperparameter tuning processes are simply too time-consuming and expensive. This repository aims to provide a minimal implementation of DomainBed, which is used for fast-track experiments in cases when researchers simply want to test new methods.

What are the new features?

We aim to provide the following features:

• DomainBed dataloaders: dataloaders from DomainBed are directly ported to this repository, which means that the data loading process is exactly the same as in DomainBed.
• timm models: we provide a wrapper for timm (PyTorch-Image-Models) models, which means that you can use any model from timm with minimal effort.
• DomainBed algorithms backward compatibility: users can directly use the Alrgorithm classes in DomainBed and minDG interchangably without any additional modification.
• Training-domain model selection: we offer the Training-domain model selection method like DomainBed.
• Extendable CLI: users can add CLI arguments more freely.
• Single environment or Iterative-Single environment: users can choose to perform DG on a single environment, or can run DG on all environments iteratively like DomainBed via --train_all. For example, on PACS with 4 environments, when --train_all is set to False (default) and --test_envs 0, only one run will be performed on the first environment. When --train_all is set to True, 4 runs will be performed with --test_envs 0, test_envs, 1,... and so on.
• wandb and webhook support: we support wandb logging and webhook notification via environment variables.

What did we remove?

• No automatic hyperparameter tuning
• No other model selection methods except Training-Domain
• Only supports model saving (by now), not args or other metadata
• Only ERM is ported from DomainBed, but users are free to add other algorithms

Benchmark (Coming Soon)