What happens under the hood?

The visual editor is built on top of Decthings model system - a model created with the visual editor is just a regular model, but created with a helpful web-based editor. It might be useful to know how this works in detail!

As described in the section intro to models, a model in decthings is defined by code and state. What is the code and state of a model created with the visual editor? Well, when you create a model using the visual editor, what really happens is that you create a model that is based on a snapshot of the model "Visual editor" (a model created by Decthings). This means that your new model will actually use the code that is defined in that model. As for the state, this is what actually contains the neural network you have created, and all the corresponding weights and biases. When you make an edit in the visual editor, a new state which stored the updated configuration is created.

To make it possible to use your model, you must compile it. Compiling means that all the weights and biases will be initialized to their default values. What really happens is that the model "Visual editor" will be executed and the function "createModelState" will be called in the model. The model will then return a new state which contains all the weights and biases.

When training or evaluating your model, because the code has access to the state, it will be able to access the graph structure that you have defined, as well as all the weights and biases of all the nodes in your network.

If you edit the input and/or output nodes of your model, and then apply that as the active state of the model, the editor will automatically update the parameter definitions of the model. The parameter definitions can be seen by going to the model's page and clicking the input/output button. As you can see, the parameters there will change to reflect the input and output types that you have defined in your graph.