Home

Awesome

Predictive Domain Adaptation with AdaGraph

This is the official PyTorch code of AdaGraph: Unifying Predictive and ContinuousDomain Adaptation through Graphs. alt text

This version has been tested on:

Installation

To install all the dependencies, please run:

pip install -r requirements.txt

Datasets

For setting up the datasets, please download CompCars and A Century of Portraits, unpack them and move them into the data folder. This folder contains also the file data/car-images.txt which contains the list of images and metadata+class annotations for CompCars.

Pretrained architectures

As pretrained models we use the default PyTorch ResNet and the the converted Caffe AlexNet for experiments with decaf. You can find the network at the following link. Please download the network and either place it on pretrained or modify the relative path in models/networks.py.

Predictive DA experiments

Our code allows to perform predictive DA experiments, producing the results for the following methods:

For reproducing the results on CompCars, please run:

python3 main.py --network decaf --suffix test_comp_alexnet --dataset compcars

for the experiments with the decaf features and:

python3 main.py --network resnet --suffix test_comp_resnet --dataset compcars

for the experiments with ResNet-18.

For Portraits, please run:

python3 main.py --network resnet --suffix portraits_test --dataset portraits --skip regions

for the experiments across regions (i.e. source and target of different regions) and:

python3 main.py --network resnet --suffix portraits_test --dataset portraits --skip decades

for the experiments across decades (i.e. source and target of different decades).

The full results of all the models will be stored in the ./results folder, as .npy.

References

If you find this code useful, please cite:

@inProceedings{mancini2019adagraph,
author = {Mancini, Massimiliano and Rota Bul\`o, Samuel and Caputo, Barbara and Ricci, Elisa},
title  = {AdaGraph: Unifying Predictive and Continuous Domain Adaptation through Graphs},
booktitle = {Computer Vision and Pattern Recognition (CVPR)},
year      = {2019},
month     = {June}
}