API documentation of pytorch_seed_rl

This package provides an extendable implementation of Google Brains “SEED”.

It is designed to be capsuled to be extended in any way. Good points of customization would be the neural netwok (nets) or loss computation (functional). Latter might need little modification of (Learner), depending of the wanted change.

See also

• “SEED RL: Scalable and Efficient Deep-RL with Accelerated Central Inference” on arXiv by Espeholt et al.

• SEED Github repository

Warning

Missing features in comparison to original implementation:

• Algorithms R2D2 and SAC

• LSTM model

• Multi-Node distributed training

• Tensorboard functionality

Submodules

• Module agents
  • Exposed Classes
    • Actor (agents.Actor)
    • Learner (agents.Learner)
  • Parent classes
    • RpcCallee (agents.RpcCallee)
    • RpcCaller (agents.RpcCaller)
• Module environments
  • Exposed Classes
    • Environment spawner object (environments.EnvSpawner)
  • Unexposed Submodules
    • Utility functions for wrapping (environments.atari_wrappers)
    • Wrappers for OpenAI gym (environments.atari_wrappers)
• Module functional
  • Exposed functions
    • Loss functions (functional.loss)
    • Vtrace (functional.vtrace)
• Module nets
  • Networks
    • Neural network for atari games (nets.AtariNet)
• Module tools
  • Exposed classes
    • Logger (tools.Logger)
    • Trajectory store (tools.Recorder)
    • Trajectory store (tools.TrajectoryStore)
  • Exposed functions (tools.functions)