Awesome

Latent Neural Differential Equations for Video Generation

Exploring the usage of Neural O/SDEs for the evolution of latent variables in videos.

This code was written while I was in high school, so sorry for the many cleanliness issues that brings. For fear of breaking the code, I have chosen not to refactor the scripts, and instead offer explanations to some of the confusing parts.

Data

To prepare your data first download UCF101, then use the file ucf101/make_ucf101_tgan.py

To use it properly alter:

src_dir = 'D:/Video Datasets/UCF101_min'
src_split_dir = 'D:/Video Datasets/ucf101/ucfTrainTestList'
dst_dir = 'D:/Video Datasets/ucf101_64px_tgan'

Set src_dir and src_split_dir to their corresponding values, then specify your destination directory (dst_dir).

Usage

In the header of each file there is a block of variables that looks like:

epochs = 100000
batch_size = 32
path = 'ucf101/tgan_svc_ode_lin'
start_epoch = 0
conf = "C:/Video Datasets/ucf101_64px/train.json"
dset = "C:/Video Datasets/ucf101_64px/train.h5"

The variables epochs and batch_size are self-explanatory. path specifies a directory relative to the starting file in two folders called checkpoints and video_samples. (checkpoints/{path}/ and video_samples/{path}/) One existing issue is that these directories must be created by the user BEFORE running. conf and dset are hard-coded paths to your data directory.

From there you should be good to go!

If you need to resume a run for whatever reason alter the start_epoch variable and uncomment the lines which handle loading checkpoints.