LSTNet Model

Model.py

def LSTNetModel(init, input_shape):
    
    # m is the number of time-series
    m = input_shape[2]

    # Get tensor shape except batchsize
    tensor_shape = input_shape[1:]
    
    if K.image_data_format() == 'channels_last':
        ch_axis = 3
    else:
        ch_axis = 1
    
    X = Input(shape = tensor_shape)

    # CNN
    if init.CNNFilters > 0 and init.CNNKernel > 0:
        # Add an extra dimension of size 1 which is the channel dimension in Conv2D
        C = Reshape((input_shape[1], input_shape[2], 1))(X)

	# Apply a Conv2D that will transform it into data of dimensions (batchsize, time, 1, NumofFilters)
        C = Conv2D(filters=init.CNNFilters, kernel_size=(init.CNNKernel, m), kernel_initializer=init.initialiser)(C)
        C = Dropout(init.dropout)(C)

        # Adjust data dimensions by removing axis=2 which is always equal to 1
        c_shape = K.int_shape(C)
        C = Reshape((c_shape[1], c_shape[3]))(C)
    else:
	# If configured not to apply CNN, copy the input
        C = X
    
    # GRU
    # Apply a GRU layer (with activation set to 'relu' as per the paper) and take the returned states as result
    _, R = GRU(init.GRUUnits, activation="relu", return_sequences = False, return_state = True)(C)
    R    = Dropout(init.dropout)(R)
    
    # SkipGRU
    if init.skip > 0:
	# Calculate the number of values to use which is equal to the window divided by how many time values to skip
        pt   = int(init.window / init.skip)

        S    = PreSkipTrans(pt, int((init.window - init.CNNKernel + 1) / pt))(C)
        _, S = GRU(init.SkipGRUUnits, activation="relu", return_sequences = False, return_state = True)(S)
        S    = PostSkipTrans(int((init.window - init.CNNKernel + 1) / pt))([S,X])

	# Concatenate the outputs of GRU and SkipGRU
        R    = Concatenate(axis=1)([R,S])
    
    # Dense layer
    Y = Flatten()(R)
    Y = Dense(m)(Y)
    
    # AR
    if init.highway > 0:
        Z = PreARTrans(init.highway)(X)
        Z = Flatten()(Z)
        Z = Dense(1)(Z)
        Z = PostARTrans(m)([Z,X])

	# Generate output as the summation of the Dense layer output and the AR one
        Y = Add()([Y,Z])
    
    # Generate Model
    model = Model(inputs = X, outputs = Y)
    
    return model