Kaiwen Wang 559656ad26 Add L2 normalization option and serve mean policy in SAC (#300) ...

Summary:
Pull Request resolved: https://github.com/facebookresearch/ReAgent/pull/300

Workflow is
(1) chose one: stochastic action or mean, (2) tanh, (3) l2norm, (4) compute kld
where (3), (4) are optional.

NOTE: kld computed the same (assuming Gaussian), regardless of tanh or normalization. We do this since actions should be <= 1 for most part.
NOTE: log_prob is also computed the same, regardless of l2 normalization.

Reviewed By: czxttkl

Differential Revision: D22909939

fbshipit-source-id: 128e88e0da781a8157c618148061929085e0b65f

2020-08-04 15:26:57 -07:00

..

__init__.py

Add TD3 model manager. (#249)

2020-05-04 20:51:53 -07:00

dirichlet_fully_connected.py

Delete part of old gym and reorganize ModelManager (#266)

2020-05-20 10:01:33 -07:00

fully_connected.py

Delete part of old gym and reorganize ModelManager (#266)

2020-05-20 10:01:33 -07:00

gaussian_fully_connected.py

Add L2 normalization option and serve mean policy in SAC (#300)

2020-08-04 15:26:57 -07:00