Skip to content


Repository files navigation

PyPi version

Why use PyTorch Adapt?

PyTorch Adapt provides tools for domain adaptation, a type of machine learning algorithm that repurposes existing models to work in new domains. This library is:

1. Fully featured

Build a complete train/val domain adaptation pipeline in a few lines of code.

2. Modular

Use just the parts that suit your needs, whether it's the algorithms, loss functions, or validation methods.

3. Highly customizable

Customize and combine complex algorithms with ease.

4. Compatible with frameworks

Add additional functionality to your code by using one of the framework wrappers. Converting an algorithm into a PyTorch Lightning module is as simple as wrapping it with Lightning.



See the examples folder for notebooks you can download or run on Google Colab.

How to...

Use in vanilla PyTorch

from pytorch_adapt.hooks import DANNHook
from pytorch_adapt.utils.common_functions import batch_to_device

# Assuming that models, optimizers, and dataloader are already created.
hook = DANNHook(optimizers)
for data in tqdm(dataloader):
    data = batch_to_device(data, device)
    # Optimization is done inside the hook.
    # The returned loss is for logging.
    _, loss = hook({**models, **data})

Build complex algorithms

Let's customize DANNHook with:

  • minimum class confusion
  • virtual adversarial training
from pytorch_adapt.hooks import MCCHook, VATHook

# G and C are the Generator and Classifier models
G, C = models["G"], models["C"]
misc = {"combined_model": torch.nn.Sequential(G, C)}
hook = DANNHook(optimizers, post_g=[MCCHook(), VATHook()])
for data in tqdm(dataloader):
    data = batch_to_device(data, device)
    _, loss = hook({**models, **data, **misc})

Wrap with your favorite PyTorch framework

First, set up the adapter and dataloaders:

from pytorch_adapt.adapters import DANN
from pytorch_adapt.containers import Models
from pytorch_adapt.datasets import DataloaderCreator

models_cont = Models(models)
adapter = DANN(models=models_cont)
dc = DataloaderCreator(num_workers=2)
dataloaders = dc(**datasets)

Then use a framework wrapper:

PyTorch Lightning

import pytorch_lightning as pl
from pytorch_adapt.frameworks.lightning import Lightning

L_adapter = Lightning(adapter)
trainer = pl.Trainer(gpus=1, max_epochs=1), dataloaders["train"])

PyTorch Ignite

trainer = Ignite(adapter), dataloader_creator=dc)

Check your model's performance

You can do this in vanilla PyTorch:

from pytorch_adapt.validators import SNDValidator

# Assuming predictions have been collected
target_train = {"preds": preds}
validator = SNDValidator()
score = validator(target_train=target_train)

You can also do this during training with a framework wrapper:

PyTorch Lightning

from pytorch_adapt.frameworks.utils import filter_datasets

validator = SNDValidator()
dataloaders = dc(**filter_datasets(datasets, validator))
train_loader = dataloaders.pop("train")

L_adapter = Lightning(adapter, validator=validator)
trainer = pl.Trainer(gpus=1, max_epochs=1), train_loader, list(dataloaders.values()))

Pytorch Ignite

from pytorch_adapt.validators import ScoreHistory

validator = ScoreHistory(SNDValidator())
trainer = Ignite(adapter, validator=validator), dataloader_creator=dc)

Run the above examples

See this notebook and the examples page for other notebooks.



pip install pytorch-adapt

To get the latest dev version:

pip install pytorch-adapt --pre

To use pytorch_adapt.frameworks.lightning:

pip install pytorch-adapt[lightning]

To use pytorch_adapt.frameworks.ignite:

pip install pytorch-adapt[ignite]


Coming soon...





Thanks to the contributors who made pull requests!

Contributor Highlights
deepseek-eoghan Improved the TargetDataset class


Thank you to Ser-Nam Lim, and my research advisor, Professor Serge Belongie.


Thanks to Jeff Musgrave for designing the logo.

Citing this library

If you'd like to cite pytorch-adapt in your paper, you can refer to this paper by copy-pasting this bibtex reference:

  title={PyTorch Adapt},
  author={Kevin Musgrave and Serge J. Belongie and Ser Nam Lim},

Code references (in no particular order)