VectorSensor and StackedSensor #2813
Conversation
| @@ -92,6 +92,8 @@ public void Generate(TensorProxy tensorProxy, int batchSize, IEnumerable<Agent> | |||
| foreach (var agent in agents) | |||
| { | |||
| var info = agent.Info; | |||
| // TODO decide what to do here - should we have 1:1 sensor:tensor, or all sensors write | |||
| // consecutively to the same 1 tensor | |||
There was a problem hiding this comment.
To be future oriented, I would say one sensor --> one tensorbut that will require changes to the trainer code...
There was a problem hiding this comment.
I wrote it to support many sensors to one tensor. We could revisit in the future, but this is pretty tightly coupled with both the trainer and the barracuda import.
|
|
||
| namespace MLAgents.Tests | ||
| { | ||
| public class EditModeTestInternalBrainTensorGenerator | ||
| { | ||
| static IEnumerable<Agent> GetFakeAgentInfos() | ||
| static IEnumerable<Agent> GetFakeAgents() |
There was a problem hiding this comment.
Rename since it returns Agents. This changed a fair amount in order to initialize the agents before returning
UnitySDK/Assets/ML-Agents/Scripts/InferenceBrain/TensorGenerator.cs
Outdated
Show resolved
Hide resolved
| /// <summary> | ||
| /// Allows sensors to write to both TensorProxy and float arrays/lists. | ||
| /// </summary> | ||
| public class WriteAdapter |
There was a problem hiding this comment.
Not sure about how I feel about this write adapter. Is this going to be user facing ? Is this what we want the API to look like eventually or is this an intermediate form?
There was a problem hiding this comment.
It's not user facing, unless they implement a custom ISensor, and even then, they only need to write to it, and even then they can use SensorBase to avoid it altogether.
I took this approach instead of trying to use ArraySegment (which might have had it's own problems) because writes to Barracuda.Tensor have some additional sideeffects (calls PrepareCacheForAccess and sets a dirty flag); I couldn't see another suitable way to abstract the writes.
There was a problem hiding this comment.
Is there a better way to hide it or mark it as "don't use"?
There was a problem hiding this comment.
(talked offline, Vince is OK with this)
chriselion
changed the title
Addresses https://jira.hq.unity3d.com/browse/MLA-231 and https://jira.hq.unity3d.com/browse/MLA-232
Adds VectorSensor and StackedSensor implementations, and uses them for Agent observations. Now all observations being sent to python and barracuda go through the ISensor interface.
The BrainParameters for number of vector observations and stacked observations are still used during Agent setup but are unused after this. They will be removed from the communication protocol in a subsequent PR.