TomMakkink / transformers-for-rl

Investigating the use of Transformers as memory mechanisms in reinforcement learning.
MIT License
5 stars 0 forks source link

Universal Transformer #6

Closed TomMakkink closed 3 years ago

TomMakkink commented 4 years ago

Implement the Universal Transformer model from the following paper: https://arxiv.org/pdf/2006.04768.pdf Official Implementation: https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/models/research/universal_transformer.py Opensource PyTorch implementation: https://github.com/andreamad8/Universal-Transformer-Pytorch

The PyTorch implementation looks fairly modular but there has been some trouble reproducing results from the official paper (which may come down to hyperparameter tuning). Compare code to the official repo which is in tf.