[rllib] A3C Refactoring #1166
[rllib] A3C Refactoring #1166
Conversation
|
Merged build finished. Test FAILed. |
|
Test FAILed. |
|
Merged build finished. Test PASSed. |
|
Test PASSed. |
|
Merged build finished. Test PASSed. |
|
Test PASSed. |
|
Merged build finished. Test PASSed. |
|
Test PASSed. |
python/ray/rllib/a3c/a3c.py
Outdated
| @@ -24,76 +22,11 @@ | |||
| "use_lstm": True, | |||
| "model": {"grayscale": True, | |||
| "zero_mean": False, | |||
| "dim": 42} | |||
| "dim": 42, | |||
| "pytorch": True} | |||
There was a problem hiding this comment.
Should this be a top level option since it affects more than the model?
| from ray.rllib.a3c.policy import Policy | ||
|
|
||
|
|
||
| class TFPolicy(Policy): |
There was a problem hiding this comment.
Eventually we will want to move this to the top level RLlib dir but I guess we need to do more refactoring for that.
python/ray/rllib/models/catalog.py
Outdated
| @@ -21,7 +21,8 @@ | |||
| "extra_frameskip", # (int) for number of frames to skip | |||
| "fcnet_activation", # Nonlinearity for fully connected net (tanh, relu) | |||
| "fcnet_hiddens", # Number of hidden layers for fully connected net | |||
| "free_log_std" # Documented in ray.rllib.models.Model | |||
| "free_log_std", # Documented in ray.rllib.models.Model | |||
| "pytorch", # Pytorch images need to be channel-major | |||
There was a problem hiding this comment.
rename this option to "channel_major" instead?
|
Merged build finished. Test PASSed. |
|
Test PASSed. |
get_gradientstocompute_gradientsmodel_updatetoapply_gradientscompute_actionstocompute_action, as we only compute one action at once. There is zero-indexing in the code since the TF nodes can actually take variable number of inputs and hence will return a list of items.a3c.py@ericl @pcmoritz