Home

Awesome

mocogan-chainer

license arXiv Tag

Chainer implementation of MoCoGAN

This repository contains an Chainer implementation of MoCoGAN.

Paper: MoCoGAN: Decomposing Motion and Content for Video Generation by Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, Jan Kautz.

Result

Generated samples (300 epoch)

MUG Facial Expression Database

normal ( not using label )infogan (during the experiment)

Summary

Dataset

MoCoGAN is one of GAN architectures for video generation. I used MUG Facial Expression Database to train the model. The dataset consisted of 86 subjects(51 male and 34 female) performing various facial expressins. Each video conststed of 50 to 160 frames.

In my experiment, the dataset converted to short video clips like shape of(video_len, channel, height, width) = (16, 3, 64, 64).

Architecture

The model architecture is shown below.

MoCoGAN Architecture

Loss Function

Requirements

Plese refer requirements.txt for required python packages.

pip install -r requirements.txt

Getting started

TODO

<!-- ### 0. Download dataset - MUG Facial Expression Database: [https://mug.ee.auth.gr/fed/](https://mug.ee.auth.gr/fed/) ### 1. Preprocessing ``` python preprocess.py <dataset_name> <dataset_path> <save_path> ``` Supported `<dataset_name>`s for now are - `mnist` - `mug` ### 2. Training For example, you train the normal model using mug dataset, ``` python train.py --model normal --dataset_type mug --dataset data/dataset/mug --max_epoch 300 --batchsize 35 --save_name mug_normal ``` Training result like log, snapshot of each network is saved at `result/<save_name>`. ### 3. Monitoring with Tensorboard Loss graph and generation results during training can be seen from Tensorboard. The data is saved at `runs/<save_name>`. ``` tensorboard --logdir=runs ``` ### 4. Generating samples using trained model ``` python generate_samples.py <generater_snapshot> <save_path> ``` -->