I try to feed the the energy consumption with timestamp dataset and covid dataset into the CNN_M_LSTM model (library Tensorflow and Keras API).

Energy consumption with timestamp dataset has the size (730, 2)

Covid dataset has the size (730, 1)

I want to slice and window both covid dataset, Energy consumption with timestamp ( with batch size = 128 and window size is 96). Finally zip both dataset together.

Here is my code slice and window both the energy consumption and timestamp dataset, the covid dataset:

MAX_LENGTH = 96
BATCH_SIZE = 128 
TRAIN.SHUFFLE_BUFFER_SIZE = 1000

def windowed_dataset(series, window_size=MAX_LENGTH, batch_size=BATCH_SIZE, shuffle_buffer=TRAIN.SHUFFLE_BUFFER_SIZE):
    """
    We create time windows to create X and y features.
    For example, if we choose a window of 30, we will create a dataset formed by 30 points as X
    """
    dataset= tf.data.Dataset.from_tensor_slices(series) 
    dataset = dataset.window(window_size + 1, shift=1) 
    dataset = dataset.flat_map(lambda window: window.batch(window_size + 1)) # the order dataset stays the same
    dataset = dataset.shuffle(1000)
    dataset = dataset.map(lambda window: (window[:-1], window[-1][0])) 
    dataset = dataset.padded_batch(128,drop_remainder=True).cache()

    return dataset

Here how I zip the dataset:

train_energy_dataset = windowed_dataset(TRAIN.PRE_PROCESSED_SERIES)
train_covid_dataset = windowed_dataset(TRAIN.SERIES_COVID)


train_dataset = tf.data.Dataset.zip((train_covid_dataset,train_energy_dataset))

train_dataset = train_dataset.shuffle(buffer_size=1000)
train_dataset = train_dataset.map(lambda data1, data2: ((data1[0], data2[0]), data1[1]))  # Adjust mapping as per your need
train_dataset = train_dataset.batch(128, drop_remainder=True).cache()

My model:

def create_CNN_LSTM_model():
    # Define the inputs
    input1 = tf.keras.layers.Input(shape=(128,1), name="input1")
    input2 = tf.keras.layers.Input(shape=(128,2), name="input2")

    # Define the CNN-LSTM part of the model
    x = tf.keras.layers.Conv1D(filters=128, kernel_size=3, activation='relu', strides=1, padding="causal")(input1)
    x = tf.keras.layers.MaxPooling1D(pool_size=2)(x)
    x = tf.keras.layers.Conv1D(filters=64, kernel_size=3, activation='relu', strides=1, padding="causal")(x)
    x = tf.keras.layers.MaxPooling1D(pool_size=2)(x)
    x = tf.keras.layers.Dropout(0.5)(x)
    x = tf.keras.layers.LSTM(16, return_sequences=True)(x)
    x = tf.keras.layers.LSTM(8, return_sequences=True)(x)
    x = tf.keras.layers.Flatten()(x)
    output_lstm = tf.keras.layers.Dense(1)(x)

    # Define the dense part of the model
    output_dense_1 = tf.keras.layers.Dense(1)(input2)

    # Concatenate the outputs of LSTM and Dense layers
    concatenated = tf.keras.layers.Concatenate(axis=1)([output_dense_1, output_lstm])

    # Add further dense layers
    x = tf.keras.layers.Dense(6, activation=tf.nn.leaky_relu)(concatenated)
    output = tf.keras.layers.Dense(4)(x)
    model_final = tf.keras.Model(inputs=[input1, input2], outputs=output)
    # Define the final model
    return model_final
   

model_cnn_m_lstm = create_CNN_LSTM_model()

    # Compile the model
model_cnn_m_lstm.compile(
        loss=tf.keras.losses.Huber(),
        optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
        metrics=["mse"]
    )

model_cnn_m_lstm.summary()



model_cnn_m_lstm.fit(train_dataset, epochs=100, batch_size=128)

With the trial to window and zip the datasets, I receive error as

ValueError: Exception encountered when calling Functional.call().

Invalid input shape for input Tensor("functional_100_1/Cast_1:0", shape=(128, 128, None, 2), dtype=float32). Expected shape (None, 128, 2), but input has incompatible shape (128, 128, None, 2)

Arguments received by Functional.call():
  • inputs=('tf.Tensor(shape=(128, 128, None, 1), dtype=float64)', 'tf.Tensor(shape=(128, 128, None, 2), dtype=float64)')
  • training=True
  • mask=('None', 'None')

I also fixed by change the input shape of the model to

    input1 = tf.keras.layers.Input(shape=(128,2), name="input1")
    input2 = tf.keras.layers.Input(shape=(128), name="input2")

The error appear as expected 2 dims but received 3 dims. 

My expectation is to feed the zip dataset into my CNN_M_LSTM model.