domainadaptation / salad

A toolbox for domain adaptation and semi-supervised learning. Contributions welcome.

🥗 salad

Semi-supervised Adaptive Learning Across Domains

salad is a library to easily setup experiments using the current state-of-the art techniques in domain adaptation. It features several of recent approaches, with the goal of being able to run fair comparisons between algorithms and transfer them to real-world use cases. The toolbox is under active development and will extended when new approaches are published.

Contribute and explore the code on Github. For commonly asked questions, head to our FAQ.

Check out robusta, our new library for domain adaptation and robustness evaluation on ImageNet scale: https://github.com/bethgelab/robustness

📊 Benchmarking Results

One of salad's purposes is to constantly track the state of the art of a variety of domain adaptation algorithms. The latest results can be reproduced by the files in the scripts/ directory.

Code for reproducing these results can be found in the scripts/ directory. Usage is outlined below.

💻 Installation

Requirements can be found in requirement.txt and can be installed via

pip install -r requirements.txt

Install the package (recommended) via

pip install torch-salad

For the latest development version, install via

pip install git+https://github.com/domainadaptation/salad

📚 Using this library

Along with the implementation of domain adaptation routines, this library comprises code to easily set up deep learning experiments in general.

The toolbox currently implements the following techniques (in salad.solver) that can be easily run with the provided example script.

• VADA (VADASolver), arxiv:1802.08735

  $ python scripts/train_digits.py --source svhn --target mnist  --vada

• Domain Adversarial Training (DANNSolver), jmlr:v17/15-239.html

  $ python scripts/train_digits.py --source svhn --target mnist  --dann

• Associative Domain Adaptation (AssociativeSolver), arxiv:1708.00938

  $ python scripts/train_digits.py --source svhn --target mnist  --assoc

• Deep Correlation Alignment

  $ python scripts/train_digits.py --source svhn --target mnist  --coral

• Self-Ensembling for Visual Domain Adaptation (SelfEnsemblingSolver) arxiv:1706.05208

  $ python scripts/train_digits.py --source svhn --target mnist    --teach

• Adversarial Dropout Regularization (AdversarialDropoutSolver), arxiv.org:1711.01575

  $ python scripts/train_digits.py --source svhn --target mnist  --adv

Examples (already refer to the examples/ subfolder) soon to be added for:

• Generalizing Across Domains via Cross-Gradient Training (CrossGradSolver), arxiv:1708.00938 Example coming soon!
• DIRT-T (DIRTTSolver), arxiv:1802.08735

Implements the following features (in salad.layers):

• Weights Ensembling using Exponential Moving Averages or Stored Weights
• WalkerLoss and Visit Loss (arxiv:1708.00938)
• Virtual Adversarial Training (arxiv:1704.03976)

Coming soon:

• Deep Joint Optimal Transport (DJDOTSolver), arxiv:1803.10081
• Translation based approaches

Quick Start

To get started, the scripts/ directory contains several python scripts for both running replication studies on digit benchmarks and studies on a different dataset (toy example: adaptation to noisy images).

$ cd scripts
$ python train_digits.py --log ./log --teach --source svhn --target mnist

Refer to the help pages for all options:

usage: train_digits.py [-h] [--gpu GPU] [--cpu] [--njobs NJOBS] [--log LOG]
                    [--epochs EPOCHS] [--checkpoint CHECKPOINT]
                    [--learningrate LEARNINGRATE] [--dryrun]
                    [--source {mnist,svhn,usps,synth,synth-small}]
                    [--target {mnist,svhn,usps,synth,synth-small}]
                    [--sourcebatch SOURCEBATCH] [--targetbatch TARGETBATCH]
                    [--seed SEED] [--print] [--null] [--adv] [--vada]
                    [--dann] [--assoc] [--coral] [--teach]

Domain Adaptation Comparision and Reproduction Study

optional arguments:
-h, --help            show this help message and exit
--gpu GPU             Specify GPU
--cpu                 Use CPU Training
--njobs NJOBS         Number of processes per dataloader
--log LOG             Log directory. Will be created if non-existing
--epochs EPOCHS       Number of Epochs (Full passes through the unsupervised
                        training set)
--checkpoint CHECKPOINT
                        Checkpoint path
--learningrate LEARNINGRATE
                        Learning rate for Adam. Defaults to Karpathy's
                        constant ;-)
--dryrun              Perform a test run, without actually training a
                        network.
--source {mnist,svhn,usps,synth,synth-small}
                        Source Dataset. Choose mnist or svhn
--target {mnist,svhn,usps,synth,synth-small}
                        Target Dataset. Choose mnist or svhn
--sourcebatch SOURCEBATCH
                        Batch size of Source
--targetbatch TARGETBATCH
                        Batch size of Target
--seed SEED           Random Seed
--print
--null
--adv                 Train a model with Adversarial Domain Regularization
--vada                Train a model with Virtual Adversarial Domain
                        Adaptation
--dann                Train a model with Domain Adversarial Training
--assoc               Train a model with Associative Domain Adaptation
--coral               Train a model with Deep Correlation Alignment
--teach               Train a model with Self-Ensembling

Reasons for using solver abstractions

The chosen abstraction style organizes experiments into a subclass of Solver.

Quickstart: MNIST Experiment

As a quick MNIST experiment:

from salad.solvers import Solver

class MNISTSolver(Solver):

    def __init__(self, model, dataset, **kwargs):

        self.model = model
        super().__init__(dataset, **kwargs)

    def _init_optims(self, lr = 1e-4, **kwargs):
        super()._init_optims(**kwargs)

        opt = torch.optim.Adam(self.model.parameters(), lr = lr)
        self.register_optimizer(opt)

    def _init_losses(self):
        pass

For a simple tasks as MNIST, the code is quite long compared to other PyTorch examples TODO.

💡 Domain Adaptation Problems

Legend: Implemented (✓), Under Construction (🚧)

📷 Vision

• Digits: MNIST ↔ SVHN ↔ USPS ↔ SYNTH (✓)
• VisDA 2018 Openset and Detection (✓)
• Synthetic (GAN) ↔ Real (🚧)
• CIFAR ↔ STL (🚧)
• ImageNet to iCubWorld (🚧)

🎤 Audio

• Mozilla Common Voice Dataset (🚧)

፨ Neuroscience

• White Noise ↔ Gratings ↔ Natural Images (🚧)
• Deep Lab Cut Tracking (🚧)

🔗 References

If you use salad in your publications, please cite

@misc{schneider2018salad,
   title={Salad: A Toolbox for Semi-supervised Adaptive Learning Across Domains},
   author={Schneider, Steffen and Ecker, Alexander S. and Macke, Jakob H. and Bethge, Matthias},
   year={2018},
   url={https://openreview.net/forum?id=S1lTifykqm}
}

along with the references to the original papers that are implemented here.

Part of the code in this repository is inspired or borrowed from original implementations, especially:

• https://github.com/Britefury/self-ensemble-visual-domain-adapt
• https://github.com/Britefury/self-ensemble-visual-domain-adapt-photo/
• https://github.com/RuiShu/dirt-t
• https://github.com/gpascualg/CrossGrad
• https://github.com/stes/torch-associative
• https://github.com/haeusser/learning_by_association
• https://mil-tokyo.github.io/adr_da/

Excellent list of domain adaptation ressources:

• https://github.com/artix41/awesome-transfer-learning

Further transfer learning ressources:

• http://transferlearning.xyz

👤 Contact

Maintained by Steffen Schneider. Work is part of my thesis project at the Bethge Lab. This README is also available as a webpage at salad.domainadaptation.org. We welcome issues and pull requests to the official github repository.