COT-GAN: Generating Sequential Data via Causal Optimal Transport

Authors: Tianlin Xu, Li K. Wenliang, Michael Munn, Beatrice Acciaio

COT-GAN is an adversarial algorithm to train implicit generative models optimized for producing sequential data. The loss function of this algorithm is formulated using ideas from Causal Optimal Transport (COT), which combines classic optimal transport methods with an additional temporal causality constraint.

This repository contains an implementation and further details of COT-GAN.

Reference: Tianlin Xu, Li K. Wenliang, Michael Munn, Beatrice Acciaio, "COT-GAN: Generating Sequential Data via Causal Optimal Transport," Neural Information Processing Systems (NeurIPS), 2020.

Paper Link: https://arxiv.org/abs/2006.08571

Contact: tianlin.xu1@gmail.com

Setup

Begin by installing pip and setting up virtualenv.

$ curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
$ python get-pip.py
$ python3 -m pip install virtualenv

Clone the Github repository and install requirements

$ git clone https://github.com/tianlinxu312/cot-gan.git
$ cd cot-gan/

# Create a virtual environment called 'venv'
$ virtualenv venv 
$ source venv/bin/activate    # Activate virtual environment
$ python3 -m pip install -r requirements.txt

Data

The .tfrecord files used in the COT-GAN experiments can be downloaded here: https://drive.google.com/drive/folders/1ja9OlAyObPTDIp8bNl8rDT1RgpEt0qO-?usp=sharing To run the code in this repository as is, download the .tfrecord files to the corresponding subfolder inside the data folder.

Note: If you have the Google Cloud SDK installed, you can copy the files using gsutil from this publicly accessible bucket

# Download the data from Google Cloud Storage
$ gsutil -m cp -r gs://munn-sandbox/cwgan/data .

Training COT-GAN

We trained COT-GAN on synthetic low-dimensional datasets as well as two high-dimensional video datasets: a human action dataset and an animated Sprites dataset

For training on low-dimensional datasets, use a flag to specify either synthetic time series sine data (SineImage), auto-regressive data of order one (AROne), or EEG data (eeg). For example, to train on AR-1 data:

# Train COTGAN on AR-1 data
$ python3 -m toy_train \
    --dname="AROne"

See the code for how to modify the default values of other training parameters or hyperparameters.

Similarly, for training on video datasets, specify either the human action or animated Sprites dataset; either human_action or animation, resp. For example,

# Train COTGAN on human action dataset
$ python3 -m video_train \
    --dname="human_action" \
    --path="./data/human_action/*.tfrecord"

# Train COTGAN on animated sprites dataset
$ python3 -m video_train \
    --dname="animation" \
    --path="./data/animation/*.tfrecord"

See the code for how to modify the default values of other training parameters or hyperparameters.

Results

Baseline models chosen for the video datasets are MoCoGAN (S. Tulyakov et al.) and direct minimization of the mixed Sinkhorn divergence. The evaluation metrics we use to assess model performance are the Fréchet Inception Distance (FID) which compares individual frames, the Fréchet Video Distance (FVD) which compares the video sequences as a whole by mapping samples into features via pretrained 3D convolutional networks, and their kernel counterparts (KID, KVD). Previous studies suggest that FVD correlates better with human judgement than KVD for videos, whereas KID correlates better than FID on images. Generated samples are provided below.

Animated Sprites

	FVD	FID	KVD	KID
MoCoGAN	1,108.2	280.25	146.8	0.34
direct minimization	498.8	81.56	83.2	0.078
COT-GAN	458.0	84.6	66.1	0.081

Human Actions

	FVD	FID	KVD	KID
MoCoGAN	1,034.3	151.3	89.0	0.26
direct minimization	507.6	120.7	34.3	0.23
COT-GAN	462.8	58.9	43.7	0.13

A minimum PyTorch implementation please see here.

luchungi/cot-gan