TD-VAE

TD-VAE implementation in PyTorch 1.0.

This code implements the ideas presented in the paper Temporal Difference Variational Auto-Encoder (Gregor et al). This implementation includes configurable number of stochastic layers as well as the specific multilayer RNN design proposed in the paper.

NOTE: This implementation also makes use of pylego, which is a minimal library to write easily extendable experimental machine learning code.

Replication

To replicate our results:

1. For model-free, run python main.py --model conditional.tdvae --name tdqvae
2. For the DRQN baseline, run python main.py --model conditional.drqn --name dqn
3. For model-based, run python main.py --model conditional.modeltdvae --tdvae_weight 1 --rl_weight 10 --mpc --eps_decay_end 1 --name mpc