Use Iterate to resume training an experiment from a selected checkpoint, keeping the weights, but with different settings. That is, you keep what the model has learned previously and don’t retrain it from scratch. This can be used with transfer learning.
How to Iterate
Click the Iterate button in the Modeling view while a model is running or when it is done training.
Clicking the Iterate button will open up a dialog where you can select one of:
Use Continue training if you have a good model and want to keep experimenting from a specific checkpoint (epoch) with some new settings.
When you click Run, you will start the new experiment straight away.
Reuse part of model
The feature Reuse part of model creates a new experiment, including one single block. We call it a user block. This custom block contains the model you just trained with weights from a specific checkpoint (epoch).
It is useful to reuse part of a model when you want to build a new model around your current model.
The reused part of the model will show up on the Modeling canvas as a new block when you click Create. The new block will, by default, be named based on the parent experiment, but it’s a good idea to name it on yourself.
The shape of the block’s input and the output will be the same as the parent experiment.
Use when you have an autoencoder, and you want to reuse just the encoding part of this model as a single block inside a new model.
When you have a model with more than one input dataset. In this case, it’s a good idea to create an experiment with one type of input and then optimize the model for this input. Then add an additional dataset as we do in the tutorial Predict real estate prices.
Other ways to continue experimenting
Iterate will keep an experiment’s weights. If you want to continue with an experiment but without the weights, you can:
Use model adjustments that the platform suggests to see if you can potentially improve the performance.
When tuning, you copy the complete project without weights. That is, you do not keep what the model has learned previously.
Duplicate the experiment and make your own changes to the model.
When duplicating, you copy the complete experiment without weights.