ADAPT

Awesome Domain Adaptation Python Toolbox

ADAPT is a python library which provides several domain adaptation methods implemented with Tensorflow and Scikit-learn.

Documentation Website

Find the details of all implemented methods as well as illustrative examples here: ADAPT Documentation Website

Installation

This package is available on Pypi and can be installed with the following command line:

pip install adapt

The following dependencies are required and will be installed with the library:

• numpy
• scipy
• tensorflow (>= 2.0)
• scikit-learn
• cvxopt

If for some reason, these packages failed to install, you can do it manually with:

pip install numpy scipy tensorflow scikit-learn cvxopt

Finally import the module in your python scripts with:

import adapt

Reference

If you use this library in your research, please cite ADAPT using the following reference: https://arxiv.org/pdf/2107.03049.pdf

@article{de2021adapt,
	  title={ADAPT: Awesome Domain Adaptation Python Toolbox},
	  author={de Mathelin, Antoine and Deheeger, Fran{\c{c}}ois and Richard, Guillaume and Mougeot, Mathilde and Vayatis, Nicolas},
	  journal={arXiv preprint arXiv:2107.03049},
	  year={2021}
	}

Quick Start

import numpy as np
from adapt.feature_based import DANN
np.random.seed(0)

# Xs and Xt are shifted along the second feature.
Xs = np.concatenate((np.random.random((100, 1)),
                     np.zeros((100, 1))), 1)
Xt = np.concatenate((np.random.random((100, 1)),
                     np.ones((100, 1))), 1)
ys = 0.2 * Xs[:, 0]
yt = 0.2 * Xt[:, 0]

# With lambda set to zero, no adaptation is performed.
model = DANN(lambda_=0., random_state=0)
model.fit(Xs, ys, Xt=Xt, epochs=100, verbose=0)
print(model.evaluate(Xt, yt)) # This gives the target score at the last training epoch.
>>> 0.0231

# With lambda set to 0.1, the shift is corrected, the target score is then improved.
model = DANN(lambda_=0.1, random_state=0)
model.fit(Xs, ys, Xt=Xt, epochs=100, verbose=0)
model.evaluate(Xt, yt)
>>> 0.0011

Examples

Two Moons	Classification	Regression

Sample Bias	Multi-Fidelity	Rotation

Content

ADAPT package is divided in three sub-modules containing the following domain adaptation methods:

Feature-based methods

• FE (Frustratingly Easy Domain Adaptation) [paper]
• mSDA (marginalized Stacked Denoising Autoencoder) [paper]
• DANN (Discriminative Adversarial Neural Network) [paper]
• ADDA (Adversarial Discriminative Domain Adaptation) [paper]
• CORAL (CORrelation ALignment) [paper]
• DeepCORAL (Deep CORrelation ALignment) [paper]
• MCD (Maximum Classifier Discrepancy) [paper]
• MDD (Margin Disparity Discrepancy) [paper]
• WDGRL (Wasserstein Distance Guided Representation Learning) [paper]

Instance-based methods

• KMM (Kernel Mean Matching) [paper]
• KLIEP (Kullback–Leibler Importance Estimation Procedure) [paper]
• TrAdaBoost (Transfer AdaBoost) [paper]
• TrAdaBoostR2 (Transfer AdaBoost for Regression) [paper]
• TwoStageTrAdaBoostR2 (Two Stage Transfer AdaBoost for Regression) [paper]

Parameter-based methods

• RegularTransferLR (Regular Transfer with Linear Regression) [paper]
• RegularTransferLC (Regular Transfer with Linear Classification) [paper]
• RegularTransferNN (Regular Transfer with Neural Network) [paper]

Acknowledgement

This work has been funded by Michelin and the Industrial Data Analytics and Machine Learning chair from ENS Paris-Saclay, Borelli center.