ml-agents train an already trained model

I might be saying something really stupid but is it possible to train an already trained model instead of starting from scratch.
If not is it possible to increase the number of steps it takes before it stops training?

 

I like to know this as well. p.s. I do know you can increase the training steps in the config file: trainer_config.yaml
(or make a copy and change that)
Then changing the setting of the corresponding name you are using like default:

and there change:

max_steps: 5.0e5

to for example:

max_steps: 8.0e5

or if your if you have the time:
max_steps: 5.0e7

 

Thank you that helps out a lot. Are there any plans to implement this in the future?

 

If you can't train a model starting from another model what is the point of that command line option:
--load: If set, the training code loads an already trained model to initialize the neural network before training.

I haven't try it yet but I though it did exactly that.

 

--load option is deprecated and replaced with --resume since release 1.
https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Migrating.md#important-changes-3

 

Pretty sure now you can use --initialize-from=MoveToGoal --run-id=MoveToGoal2 where it uses data from MoveTopGoal to build upon previous learning. But, im not sure if you need to pass the new id in each time? for example :

run one
--initialize-from=MoveToGoal --run-id=MoveToGoal2
run two
--initialize-from=MoveToGoal2 --run-id=MoveToGoal3

or whether you just do
--initialize-from=MoveToGoal --run-id=MoveToGoal3

anyone know?

 

May 21,2023:
try to use:

Code (CSharp):

1. >mlagents-learn --run-id=yourID --resume

 

Just some little note from the little that i found.
This option is to simply resume a run as it is. This has the added advantage that we preserve all the training history and stats.

So make sure to not delete the folder and files of the run.
So unless one edit the content to conform to what we need (dragons ahead), then we should just use it as intended.

This option is the one that we probably should use if we want to resume the training from any earlier model and at any step. Note that we will only have the model and that the earlier history will be lost.

But to make it work, ther are some important caveats.
We needs a 'checkpoint.pt' file to initialize the training. 'checkpoint.pt' seem to basically be the <modelName>.pt file that is in the run folder beside the <model>.onnx one. Just change its name and extension.
I think that the .pt file is the model in Pytorch format.
So if you only have the onnx file, then you might be able to convert it to pytorch before changing its name.
To know where to put that 'checkpoint.pt' file, just check any run and you will see the folder structure and its place in it.

 

How do you restore a specific checkpoint? Is it even possible?