Add FiLMQConvNet, custom image downsize - Githubissues

kengz / SLM-Lab

Modular Deep Reinforcement Learning framework in PyTorch. Companion library of the book "Foundations of Deep Reinforcement Learning".

https://slm-lab.gitbook.io/slm-lab/

MIT License

1.24k stars 264 forks source link

Add FiLMQConvNet, custom image downsize #423

Closed kengz closed 4 years ago

kengz commented 4 years ago

FiLMQConvNet

Use Feature-wise Linear Modulation applied to the outputs of the last state_fc_model hid_layers. This is mainly meant for SAC's Q network.

custom image downsize

allow for custom image downsize other than the original 84x84. Specify as image_downsize in the env spec

misc. SAC updates

set target_entropy to action_space.n for discrete actions
add more specs for SAC Atari, currently in the process of tuning, result will be uploaded next.