lagom: A PyTorch infrastructure for rapid prototyping of reinforcement learning algorithms.
JIT-enabled LayerNormLSTM, it is much faster than the raw implementation !
Sync examples/mdn
and examples/VAE
to the latest API design
Added instruction about how to train dm_control
environments with minimal modifications.
Remove AtariPreprocessing
because it's merged as a PR in gym officially.
Improve scripts & CI: use conda as much as possible for MKL optimization (like numpy/scipy) etc.
Research-friendly (easy to read & modify) RL baselines
It contains following algorithms for now:
Much easier and cleaner API
Major high-level designs and major APIs converge to be stable in this release. Most of modules are well-tested.
This is a preview release of lagom. For this version, see Basics in README for quick start or directly play around with examples. A full documentation is available online at http://lagom.readthedocs.io/.