Feature_Critic

Demo code for 'Feature-Critic Networks for Heterogeneous Domain Generalisation', including codes for heterogeneous DG (VD) and homogeneous DG (PACS). This paper is located at https://arxiv.org/abs/1901.11448 and will appear in the forthcoming ICML 2019.

Yiying Li, Yongxin Yang, Wei Zhou, Timothy M. Hospedales. Feature-Critic Networks for Heterogeneous Domain Generalisation[C]. ICML 2019.

Introduction

Feature_Critic aims to address the domain generalisation problem with a particular focus on the heterogeneous case, by meta-learning a regulariser to help train a feature extractor to be domain invariant. The resulting feature extractor outperforms alternatives for general purpose use as a fixed downstream image encoding. Evaluated on Visual Decathlon -- the largest DG evaluation thus far -- this suggests that Feature_Critic trained feature extractors could be of wide potential value in diverse applications. Furthermore Feature_Critic also performs favourably compared to state-of-the-art in the homogeneous DG setting, such as on PACS dataset.

Citing Feature_Critic

If you find Feature_Critic useful in your research, please consider citing:

@inproceedings{Li2019ICML,
   Author={Li, Yiying and Yang, Yongxin and Zhou, Wei and Hospedales, Timothy},
   Title={Feature-Critic Networks for Heterogeneous Domain Generalisation},
   Booktitle={The Thirty-sixth International Conference on Machine Learning},
   Year={2019}
   }

Download datasets and models

Preparation

We provide two ways to download datasets and trained models on our MEGA network disk:

(i) Download directly from the link and put them under the corresponding project dir:

PACS dataset is on https://mega.nz/#F!jBllFAaI!gOXRx97YHx-zorH5wvS6uw. pacs_data and pacs_label can be put under <home_dir>/data/PACS/.

All trained models of VD and PACS are on https://mega.nz/#F!rRkgzawL!qoGX4bT3sif88Ho1Ke8j1Q, and they can be put under <home_dir>/model_output/. If you want to use the trained Feature_Critic for encoding to extract your features, you can download and use the torch models that under the <Feature_Critic> folder.

VD dataset download should follow the Download VD Dataset instructions below.

(ii) Install the network disk command line tool first and then use our script for downloading.

(1) Download the soure code of MEGA tool.
git clone https://github.com/meganz/MEGAcmd.git
cd MEGAcmd
git submodule update --init --recursive
(2) Install the tool
apt install libcurl4-openssl-dev libc-ares-dev libssl-dev libcrypto++-dev zlib1g-dev libsqlite3-dev libfreeimage-dev
apt install autoconf automake libtool libreadline6-dev
sh autogen.sh
./configure
make
sudo make install
sudo ldconfig

Download VD Dataset

From the official website(https://www.robots.ox.ac.uk/%7Evgg/decathlon/), please download the following files:

(1) Annotations and code. The devkit [22MB] contains the annotation files as well as example MATLAB code for evaluation. You can put under `<home_dir>/data/VD/`.
(2) Images. The following archives contain the preprocessed images for each dataset, and they can be put under `<home_dir>/data/`:
            Preprocessed images [406MB]. Images from all datasets except ImageNet ILSVRC.
            Preprocessed ILSVRC images [6.1GB]. In order to download the data, the attendees are required to register an ImageNet (http://image-net.org/signup) account first. Images for the ImageNet ILSVRC dataset (this is shipped separately due to copyright issues).

Download PACS Dataset and trained models

Make sure to run script to download the PACS dataset and trained models from the MEGA network disk.

bash get_model_dataset.sh

Installation

Install Anaconda:

curl -o /tmp/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-latest-MacOSX-x86_64.sh
bash /tmp/miniconda.sh
conda create -n FC python=2.7.12
source activate FC

Install necessary Python packages:

pip install torchvision pycocotools torch

Running and Results

Experiments on VD

Baseline(AGG) Launch the entry script of baseline method:

python main_baseline.py --dataset=VD

Parameters (e.g., learning_rate, batch_size...) and flags can be found and tuned in main_baseline.py. Turn on the is_train to train the baseline model. Experiment data is saved in <home_dir>/logs/VD/baseline/. You can achieve 19.56%, 36.49%, 58.04%, 46.98% on the four target domains (Aircraft, D.Textures, VGG-Flowers, UCF101) with the average 40.27%. (cf Table 1 in the paper)

Feature_Critic Load the VD baseline model to <home_dir>/model_output/VD/baseline/

Launch the entry script of Feature_Critic method, parameters and flags can also be tuned by yourself:

python main_Feature_Critic.py --dataset=VD

Experiment data is saved in <home_dir>/logs/VD/Feature_Critic/. You can achieve 20.94%, 38.88%, 58.53%, 50.82% on the four target domains (Aircraft, D.Textures, VGG-Flowers, UCF101) with the average 42.29%. (cf Table 1 in the paper)

Experiments on PACS

Experiments need to be performed four times in the leave-one-domain-out way. Take the "leave-A-domain-out" as the example, and you can change the target domain (unseen_index) as in the main file.

For baseline, you can achieve 63.3%, 66.3%, 88.6%, 56.5% when setting A, C, P, S as the target domain，respectively, and get the average 68.7%. (cf Table 5 in the paper)

For Feature_Critic, you can achieve 64.4%, 68.6%, 90.1%, 58.4% when setting A, C, P, S as the target domain，respectively, and get the average 70.4%. (cf Table 5 in the paper)

Baseline(AGG) Launch the entry script of baseline method:

python main_baseline.py --dataset=PACS

Feature_Critic Load the PACS baseline model (A) to <home_dir>/model_output/PACS/baseline/A/