Setting states for rnn forecasting

[JohnVorwald]

Below is a simple rnn that runs on some input to generate the output. The input and output are split into train / test. The model is fitted with the train data. When the test data is applied, the results are incorrect due to the state not be initialized. The two questions are: 1) how to save the final state from training, and 2) how to set the initial state for testing.

def main():
    input_data = pd.DataFrame(data={
        "A": [1, 1, -19, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 4],
        "B": [1, 1, -18, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 4]
    })

    print(input_data.values.sum())

    depth = input_data.shape[1]

    x_t = Input(shape=(depth,))  # The input to the RNN at time t
    y_tm1 = Input(shape=(depth,))  # Previous output
    h_tm1 = Input(shape=(depth,))

    # Compute new hidden state
    x_t_and_y_tm1 = Concatenate()([x_t, y_tm1])
    h_t = Dense(depth, kernel_initializer='ones')(x_t_and_y_tm1)

    # Build the RNN
    rnn = recurrentshop.RecurrentModel(input=x_t, initial_states=[h_tm1], output=h_t, final_states=[h_t],
                                       readout_input=y_tm1, return_sequences=True)# , stateful=True)

    # Build a Keras Model using our RNN layer
    n_time_steps = input_data.shape[0]
    x = Input(shape=(None, depth))
    y = rnn(x)
    model = Model(x, y)

    # Run the RNN over a sequence
    expanded_input_data = np.expand_dims(input_data, axis=0)
    out = model.predict(expanded_input_data)

    print(f"expanded_input_data: {expanded_input_data}")
    print(f"out: {out}")

    # now train over in/out
    n_training = int(0.8*n_time_steps)

    training_in, testing_in = expanded_input_data[:, :n_training, :], expanded_input_data[:, n_training:, :]
    training_out, testing_out = out[:, :n_training, :], out[:, n_training:, :]

    # reinitialize weights

    for layer in model.layers:
        if hasattr(layer, 'kernel_initializer'):
            layer.kernel.initializer.run()
        if hasattr(layer, 'bias_initializer'):
            layer.bias.initializer.run()

    model.compile(loss='mse', optimizer='adam')
    model.fit(training_in, training_out, validation_data=(training_in, training_out))

    print(f"model.states: {model.summary()}")
    # K.set_value(lstm.states[i], val) or K.get_value(lstm.states[i])

    predict_out = model.predict(testing_in)

    print(f"testing_out: {testing_out}")
    print(f"predict_out: {predict_out}")

if __name__ == '__main__':
    main()

The expected output is

testing_out: [[[-9.6468736e+07 -9.6468736e+07]
[-1.9293747e+08 -1.9293747e+08]
[-3.8587494e+08 -3.8587494e+08]
[-7.7174989e+08 -7.7174989e+08]
[-1.5434998e+09 -1.5434998e+09]
[-3.0869996e+09 -3.0869996e+09]]]

The predict output is

predict_out: [[[  2.   2.]
  [  6.   6.]
  [ 14.  14.]
  [ 30.  30.]
  [ 62.  62.]
  [132. 132.]]]