While things might seem very different at this point, training an LSTM is actually not any different than training a deep neural network on a typical cross-sectional problem:

LAGS=10
df = read_data()
df_train = select_dates(df, start="2017-01-01", end="2017-05-31")
df_test = select_dates(df, start="2017-06-01", end="2017-06-30")
X_train, X_test, y_train, y_test = prep_data(df_train, df_test, lags=LAGS)
model = build_network(sequence_length=LAGS)
callbacks = create_callbacks("lstm_100_100")
model.fit(x=X_train, y=y_train,
          batch_size=100,
          epochs=10,
          callbacks=callbacks)
model.save("lstm_model.h5")

After preparing our data, we instantiate a network with the architecture we've walked through and then call fit on it as expected.

Here I'm using a stateful LSTM. One practical benefit of stateful LSTMs is that they tend to train in fewer epochs than stateless LSTMs. If you were to refactor this as a stateless LSTM, you might need 100 epochs before the network has finished learning, whereas here we are only using 10.