spacelib is a small collection of tools for off-policy reinforcement learning, particularly with recurrent agents and particularly for learning directly from pixel observations.
The package makes pretty heavy use of pytorch tensors and the openai gym APIs.
Features:
- efficiently store completed episodes on disk in memory-mapped numpy arrays
- transparently handle hierarchically structured action/observation spaces
- simple interface for batch sampling of sequences from experience history
- store hidden states for recurrent models
For a usage example, see the example notebook.
- stop one-hot encoding discrete spaces
- test/add an example for hidden state storage
This project is licensed under the terms of the MIT license Please let me know if you find it useful!