In step 1, nOut() expects the number of image labels that we calculated using FileSplit in an earlier recipe.

In step 2, we have used setInputType() to set the convolutional input type. This will trigger computation/settings of the input neurons and add preprocessors (LocalResponseNormalization) to handle data flow from the convolutional/subsampling layers to the dense layers.

The InputType class is used to track and define the types of activations. This is most useful for automatically adding preprocessors between layers, and automatically setting nIn (number of input neurons) values. That's how we skipped specifying nIn values earlier when configuring the model. The convolutional input type is four-dimensional in shape [miniBatchSize, channels, height, width].