unsupervised domain adaptation using optimal transport

Paper

This repository contains the code for Domain Adaptation using Optimal Transport for Invariant Learning on Histopathology Datasets. Licensing details can be found in the license.txt file.

Requirements

This source has been tested on Ubuntu 18.04.4
CPU or NVIDIA GPU
Python 3.7.1
CUDA 10.1
PyTorch 1.3

Python Dependencies

numpy
torch
torchvision
geomloss

Dataset

The data used for training is camelyon17-wilds. For installation use:

pip install wilds

For downloading cameyon17 use:

from wilds import get_dataset
dataset = get_dataset(dataset="camelyon17", download=True)

More information can be found here

Training and evaluation

To train OT model:

python train_OT.py --exp 0 --Lambda .1 --batch_size 128

To train DANN model:

python train_DANN.py --exp 0  --batch_size 128