Add different types of visual encoder (nature cnn/resnet) #2289
Conversation
|
Noticed this doesn't actually hook the different encoder types up to any config. Is that intended to be a separate PR / planned at all? BTW, also please make sure to make a descriptive commit message and PR description (they can be the same) for this. |
Added config option for encoder type in trainer config |
| @@ -302,7 +434,9 @@ def create_discrete_action_masking_layer(all_logits, action_masks, action_size): | |||
| ), | |||
| ) | |||
|
|
|||
| def create_observation_streams(self, num_streams, h_size, num_layers): | |||
| def create_observation_streams( | |||
There was a problem hiding this comment.
Please add type annotations here.
| @@ -401,13 +559,15 @@ def create_value_heads(self, stream_names, hidden_input): | |||
| self.value_heads[name] = value | |||
| self.value = tf.reduce_mean(list(self.value_heads.values()), 0) | |||
|
|
|||
| def create_cc_actor_critic(self, h_size, num_layers): | |||
| def create_cc_actor_critic(self, h_size, num_layers, vis_encode_type): | |||
| @@ -44,6 +44,7 @@ def __init__(self, seed, brain, trainer_params, is_training, load): | |||
| m_size=self.m_size, | |||
| seed=seed, | |||
| stream_names=list(reward_signal_configs.keys()), | |||
| vis_encode_type=trainer_params["vis_encode_type"], | |||
There was a problem hiding this comment.
Convert to EncoderType enum here and pass that through instead of the string type. Handle a missing value with .get(). For example
vis_encode_type = EncoderType(trainer_params.get("vis_encode_type", "default"))
|
@dongruoping can you make sure to address the comments here from @chriselion in a following PR? |
Add different types of visual encoder (nature cnn/resnet)