how can I continue training an existing model in unity ml?

[zxh188482825]

When loading the existing model in training mode, training begins from scratch. I tried very low and the same learning rates in the second training round.

Here are the Hyperparameters for the first training round

After the first training round, I got a trained model and added a model for the second training round. And in the second training round, I changed the learning rate to 0. However, training begins from scratch instead of continuing training an existing model.

[zxh188482825]

The purpose of training an existing model is to query some data for the existing model. For example, in this model, what is the intrinsic reward of the agent after each step.

[hk1ll3r]

The model alone is not enough for continuing training AFAIK.

You need to pass the --initialize-from flag to mlagents-learn as described here:

Unity-Technologies/ml-agents/blob/main/docs/Training-ML-Agents.md#loading-an-existing-model

The doc is vague on what things can change and what things cannot. You network shape (hidden_layers, etc.) shouldn’t change.