Clockwork VAEs in JAX/Flax

Implementation of experiments in the paper Clockwork Variational Autoencoders (project website) using JAX and Flax, ported from the official TensorFlow implementation.

Running on a single TPU v3, training is 10x faster than reported in the paper (60h -> 6h on minerl).



Clockwork VAEs are deep generative model that learn long-term dependencies in video by leveraging hierarchies of representations that progress at different clock speeds. In contrast to prior video prediction methods that typically focus on predicting sharp but short sequences in the future, Clockwork VAEs can accurately predict high-level content, such as object positions and identities, for 1000 frames.

Clockwork VAEs build upon the Recurrent State Space Model (RSSM), so each state contains a deterministic component for long-term memory and a stochastic component for sampling diverse plausible futures. Clockwork VAEs are trained end-to-end to optimize the evidence lower bound (ELBO) that consists of a reconstruction term for each image and a KL regularizer for each stochastic variable in the model.


This repository contains the code for training the Clockwork VAE model on the datasets minerl, mazes, and mmnist.

The datasets will automatically be downloaded into the --datadir directory.

python3 --logdir /path/to/logdir --datadir /path/to/datasets --config configs/<dataset>.yml 

The evaluation script writes open-loop video predictions in both PNG and NPZ format and plots of PSNR and SSIM to the data directory.

python3 --logdir /path/to/logdir

Known differences from the original

  • Flax' default kernel initializer, layer precision and GRU implementation (avoiding redundant biases) are used.
  • For some configuration parameters, only the defaults are implemented.
  • Training metrics and videos are logged with wandb.
  • The base configuration is in

Added features:

  • This implementation runs on TPU out-of-the-box.
  • Apart from the config file, configuration can be done via command line and wandb.
  • Matching the seed of a previous run will exactly repeat it.

Things to watch out for

Replication of paper results for the mazes dataset has not been confirmed yet.

Getting evaluation metrics is a memory bottleneck during training, due to the large eval_seq_len.
If you run out of device memory, consider lowering it during training, for example to 100.
Remember to pass in the original value to to get unchanged results.


Thanks to Vaibhav Saxena and Danijar Hafner for helpful discussions and to Jamie Townsend for reviewing code.