Cell

A cell node is similar to a constant, but instead of defining the value in the editor, the value is defined when training. One of the training input parameters will be given the same name as the cell node, and the value given to this parameter will be stored in the state that is created after the training is completed. When evaluating the model, the value provided at the latest training session will be used.

You can change the name, data type and shape of the value to store in the cell by clicking on the node and configuring it on the right panel.

A cell is useful when we want to be able to configure our model when training it. For example, for an image classifier, in order to convert the input labels to numbers we can use the String lookup node, but for this we need access to the available labels (such as cats, dogs, cars, etc.). The problem is that because these labels really depend on what dataset we provide when training, we might not know these labels when creating the model in the editor. To solve this, we can use a cell of type string and connect that to our String lookup node, and then provide the cell value with labels when training the model.