This is the implementation of the paper Lower Dimensional Kernels for Video Discriminators based on MoCoGAN.
Along adding ldvd models to the code, I have fixed TF2, scipy and python3 compatibility problems but some parts still needs improvement.
To use the Factorized video disciminator in the paper use:
--video_discriminator FactorizedPatchVideoDiscriminator
option when specifying the video disciminator.
- This implementation does not reduce number of parameters as much as the original paper claims.
- Time shifting is still not implemented
- Frechet Inception distance and Inception score should be evaluated
- Saving images, videos have been changed and might be sub-optimal.
For more details on training and generating videos, refer to the MoCoGAN repo.