In step 1, the parameters included are as follows:

• parentPathLabelGenerator—created during the data extraction stage (see the Extracting images from disk recipe in this chapter).
• channels—The number of color channels (default = 3 for RGB).
• ImageRecordReader(imageHeight, imageWidth, channels, parentPathLabelGenerator)—resize the actual image to the specified size (imageHeight, imageWidth) to reduce the data loading effort.
• The null attribute in the initialize() method is to indicate that we are not training transformed images.

In step 3, we use ImagePreProcessingScaler for min-max normalization. Note that we need to use both fit() and setPreProcessor() to apply normalization to the data.

For GPU-accelerated environments, we can use PerformanceListener instead of ScoreIterationListener in step 4 to optimize the training process further. PerformanceListener tracks the time spent on training per iteration, while ScoreIterationListener reports the score of the network every N iterations during training. Make sure that GPU dependencies are added as per step 7.

In step 5, we have trained the model again with the image transformations that were created earlier. Make sure to apply normalization to the transformed images as well.