Skip to content

This is implementation of convolutional variational autoencoder in TensorFlow library and it will be used for video generation.

License

Notifications You must be signed in to change notification settings

Nemzy/video_generator

Repository files navigation

video_generator

This is implementation of convolutional variational autoencoder in TensorFlow library and it was used for video generation.

Also this is my code for the Siraj Raval's coding challenge "How to Generate Images - Intro to Deep Learning #14".

You can find more about it here

Overview

For more info about video generation see jupyter notebook and for implementation details of conv-variational autoencoder see conv_vae.py. Dataset was created by using ffmpeg tool.

Dependencies

  • TensorFlow r1.0
  • pillow
  • matplotlib
  • numpy
  • sklearn

All dependencies can be installed by pip.

Dataset creation steps

  • Resize video: ffmpeg -i input.mp4 -s 64x64 -c:a copy output.mp4
  • Create frames: ffmpeg -i output.mp4 -r NUM-OF-FRAMES-PER-SEC -f image2 SOME/PATH/%05d.png

Credits

Credits go to these guys:

  • Kevin Frans url

  • Arthur Juliani url

About

This is implementation of convolutional variational autoencoder in TensorFlow library and it will be used for video generation.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published