lstm_model.py

def model_lstm(input_shape):
    # The shape was explained above, must have this order
    inp = Input(shape=(input_shape[1], input_shape[2],))
    # This is the LSTM layer
    # Bidirecional implies that the 160 chunks are calculated in both ways, 0 to 159 and 159 to zero
    # although it appear that just 0 to 159 way matter, I have tested with and without, and tha later worked best
    # 128 and 64 are the number of cells used, too many can overfit and too few can underfit
    x = Bidirectional(CuDNNLSTM(128, return_sequences=True))(inp)
#     x = Activation('relu')(x)
#     x = Dropout(0.25)(x)
#     x = BatchNormalization()(x)
    # The second LSTM can give more fire power to the model, but can overfit it too
    x = Bidirectional(CuDNNLSTM(64, return_sequences=True))(x)
#     x = Activation('relu')(x)
#     x = Dropout(0.25)(x)
#     x = BatchNormalization()(x)
    # Attention is a new tecnology that can be applyed to a Recurrent NN to give more meanings to a signal found in the middle
    # of the data, it helps more in longs chains of data. A normal RNN give all the responsibility of detect the signal
    # to the last cell. Google RNN Attention for more information :)
    #x = Dropout(0.2)(x)
    #x = BatchNormalization()(x)
    x = Attention(input_shape[1])(x)
    
    # A intermediate full connected (Dense) can help to deal with nonlinears outputs
    x = Dense(64, activation="relu")(x)
    # A binnary classification as this must finish with shape (1,)
    x = Dense(1, activation="sigmoid")(x)
    model = Model(inputs=inp, outputs=x)
    # Pay attention in the addition of matthews_correlation metric in the compilation, it is a success factor key
    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=[matthews_correlation])
    
    return model

ok