Develop remove past action communication #2913
Conversation
UnitySDK/Assets/ML-Agents/Scripts/InferenceBrain/GeneratorImpl.cs
Outdated
Show resolved
Hide resolved
| @@ -172,7 +178,7 @@ def make_empty_memory(self, num_agents): | |||
| :param num_agents: Number of agents. | |||
| :return: Numpy array of zeros. | |||
| """ | |||
| return np.zeros((num_agents, self.m_size)) | |||
| return np.zeros((num_agents, self.m_size), dtype=np.float) | |||
There was a problem hiding this comment.
Out of scope for this PR, but should we try to make sure we never create float64 nparrays?
| @@ -194,6 +200,34 @@ def remove_memories(self, agent_ids): | |||
| if agent_id in self.memory_dict: | |||
| self.memory_dict.pop(agent_id) | |||
|
|
|||
| def make_empty_previous_action(self, num_agents): | |||
There was a problem hiding this comment.
Is there any difference between how we handle memories and previous actions besides the the dtype? Worth combining into a small utility class?
There was a problem hiding this comment.
I think there is going to be a bigger refactor of this part coming later. I think it is fine for now. What do you think @ervteng ?
There was a problem hiding this comment.
Might be worth combining - but then again, there will only be two things we store between steps. Also, the "when" they're added is different - memories can be added after the get_action, previous_actions must be added before.
I think there is no way around it, this logic will still need to live in the Policy for Python inference to work. The only other place I can think of is the env_manager, but that would be messy.
|
Just to clarify, existing models will keep working as before? The only thing that would break when upgrading is existing demo files? |
|
Models will still work as before but the demo files will not. |
Will we need to re-record all of them or just the Hallway one? |
I reconverted them already |
No description provided.