Update distributed trainings - Githubissues

DeepX-inc / machina

Control section: Deep Reinforcement Learning framework

MIT License

279 stars 45 forks source link

Update distributed trainings #247

Closed mmisono closed 5 years ago

mmisono commented 5 years ago

This PR introduces distributed training using ray and some refactoring.

Major change:

Refactor DistributedEpiSampler
- Reset DB state on initialization. This is needed when the previous processes are killed during training.
- Refactor synchronization code.
Add ray sampler (This is faster than DistributedEpiSampler)
Remove dp_run (see run_ppo_distributed.py for multi-GPU training example)