Ways to optimize Deeplearning4j model, FP32 to FP16 Quantization

QuantizationDL4J.java

// Tip 1: Do not save the updater if you do not plan to continue training your model
// Set false for saveUpdater flag.
ModelSerializer.writeModel(model, modelFilename, false);
// Results: Model size drop almost 40%.

// Tip 2: Convert FP32 to FP16 floating point precision(Quantization), currently DL4J only supports float. {DOUBLE, FLOAT, HALF}
// Convert the parameters just as below:
model = model.convertDataType(DataType.HALF);
// Results: Model size drop by 50%, half of its original size. Accuracy did not drop.

// You can check data type by:
System.out.println(model.params().dataType());