proceduralia / tgan-pytorch

A PyTorch implementation of Temporal Generative Adversarial Nets with Singular Value Clipping

tgan-pytorch

A PyTorch implementation of Temporal Generative Adversarial Nets with Singular Value Clipping.

Implementation details

Although in the original implementation Wasserstein GANs with weight clipping and singular value clipping were used, this version uses the training procedure highlighted in Improved Training of Wasserstein GANs. The model is trained on MovingMNIST.

Requirements

PyTorch
torchvision
PyYAML

Configuration

The configuration options are saved in config.yml. Modify its content accordingly to your needs, for example setting use_cuda to False.

Usage

Download the MovingMNIST dataset in the data/ directory using:

wget http://www.cs.toronto.edu/~nitish/unsupervised_video/mnist_test_seq.npy

To start the training:

python train.py

Savings will be contained in the checkpoints/ directory, while images representing the generated videos are stored in samples/.

Citation of original authors

I am not one of the authors of the original work nor affiliated with them. To reference their work please use:

@inproceedings{TGAN2017,
    author = {Saito, Masaki and Matsumoto, Eiichi and Saito, Shunta},
    title = {Temporal Generative Adversarial Nets with Singular Value Clipping},
    booktitle = {ICCV},
    year = {2017},
}

tgan-pytorch's Issues

Can't get reasonable results

Thanks for sharing the code.

I run your code without modification, but the results are really bad, as the attached image shows.

What's your results when running this code? Can it produce reasonable results?

question about interpolates

the interpolation should be made in latent variable z not the raw image.

interpolates = alpha * real_data + ((1 - alpha) * fake_data)

Temporal Discriminator problem

Hi, May i ask why did you do a permutation x = x.permute(0,2,1,3,4) on the input tensor of size (8,3,16,64,64)? It resulted in a mismatch in tensor size when I tried to run it.