LSTM models from 1.x will be incompatible with MLA 2.x#5254
Conversation
| @@ -79,6 +79,8 @@ public BuiltInSensorType GetBuiltInSensorType() | |||
| public class ParameterLoaderTest | |||
| { | |||
| const string k_discrete_ONNX_v2 = "Packages/com.unity.ml-agents/Tests/Editor/TestModels/discrete_rank2_vector_v2_0.onnx"; | |||
There was a problem hiding this comment.
Can we rename the old model file to something like "obsolete_lstm_v1.onnx"? And rename the constants and NNModel variables accordingly?
There was a problem hiding this comment.
I renamed to discrete1vis0vec_2_3action_obsolete_recurr_v1_0.onnx. I remade this model with version 15 since it seems this model was actually not a recurrent model.
chriselion
left a comment
chriselion
left a comment
There was a problem hiding this comment.
Looks good. Please update the changelog and migration guide.
Discussed offline, but some code that got added in #5236 can be removed, and probably BarracudaMemoryOutputApplier and BarracudaRecurrentInputGenerator too.
… were removed. All code that checked for (memories + v1.X) have been removed since these will no longer be supported
| need to be retrained. Please see the Migration Guide for more details. (#5256) | ||
| - Models trained with 1.x versions of ML-Agents will no longer work at inference if they were trained using recurrent neural networks (#5254) | ||
|
|
||
| #### ml-agents / ml-agents-envs / gym-unity (Python) |
There was a problem hiding this comment.
Is this intentional? It was in 1.9.1, right?
There was a problem hiding this comment.
No it is not. I think I made a mistake when resolving conflicts in the last commit.
2 participants
Proposed change(s)
Making 1.x models that contain LSTM not work with MLA 2.0.
This is to prepare for barracuda 2.0 that will break compatibility for LSTM.
Useful links (Github issues, JIRA tickets, ML-Agents forum threads etc.)
Types of change(s)
Checklist
Other comments