WeightsUpgradeOnTraining.py

from keras.callbacks import Callback
import numpy

class WeightsUpgradeOnTraining(Callback):

    def __init__(self, particleWeights, numberOfNeurons):
        self.PWeights = particleWeights
        self.nNeurons = numberOfNeurons

    #THIS FUNCTION IS IMPORTANT TO UNDERSTAND THE CONCEPT OF WEIGHTS ON LSTM
    def on_train_begin(self, logs={}):

        #Ref: https://stackoverflow.com/questions/42861460/how-to-interpret-weights-in-a-lstm-layer-in-keras
        #REF CORRECT ORDER OF WEIGHTS ON LSTM LAYERS (i,f,c,o) --> https://stackoverflow.com/questions/47661105/order-of-lstm-weights-in-keras

        for layer in self.model.layers: #UPDATING THE WEIGHTS OF EACH MODEL LAYER

            if "LSTM".lower() in layer.name :
                W = layer.get_weights()[0]
                U = layer.get_weights()[1]

                #EXTRACT ALL MATRICES OF U AND V
                W_i = W[:, :self.nNeurons]
                W_f = W[:, self.nNeurons: self.nNeurons * 2]
                W_c = W[:, self.nNeurons * 2: self.nNeurons * 3]
                W_o = W[:, self.nNeurons * 3:]

                U_i = U[:, :self.nNeurons]
                U_f = U[:, self.nNeurons: self.nNeurons * 2]
                U_c = U[:, self.nNeurons * 2: self.nNeurons * 3]
                U_o = U[:, self.nNeurons * 3:]

                '''
                    BIAS --> IN THIS EXAMPLE I DON'T EXPLORE OPTIMIZATION OF BIAS, BUT I PUT HERE THE MATRICES OF BIAS
                    b_i = b[:units]
                    b_f = b[units: units * 2]
                    b_c = b[units * 2: units * 3]
                    b_o = b[units * 3:]
                '''

                print(W_i)
                print(W_f)
                print(W_c)
                print(W_o)
                print(U_i)
                print(U_f)
                print(U_c)
                print(U_o)

    def on_train_end(self, logs={}):
        #print(logs)
        return None

    def on_batch_end(self, batch, logs=None):
        return None #there is not anything left, if there is a need to update the weights in each batch it may be necessary to override this function, however I think it only makes sense to apply the update of the weights at the beginning of the training and not of each batch