经过所有这些准备工作，构建和训练模型的代码就相当简短！让我们首先快速查看“长版本”，这将允许你测试堆叠多个 LSTM 或使用状态 LSTM，然后通过最终的短版本（两者都不做）并对其进行评论。

完整的代码，仅供参考。

model <- keras_model_sequential()

model %>%
    layer_lstm(
        units = FLAGS$n_units,
        batch_input_shape  = c(
            FLAGS$batch_size,
            FLAGS$n_timesteps,
            n_features),
        dropout = FLAGS$dropout,
        recurrent_dropout = FLAGS$recurrent_dropout,
        return_sequences = TRUE,
        stateful = FLAGS$stateful)

if (FLAGS$stack_layers)
{
    model %>%
        layer_lstm(
            units             = FLAGS$n_units,
            dropout           = FLAGS$dropout,
            recurrent_dropout = FLAGS$recurrent_dropout,
            return_sequences  = TRUE,
            stateful          = FLAGS$stateful)
}

model %>% time_distributed(
    layer_dense(units = 1))

model %>%
    compile(
        loss      = FLAGS$loss,
        optimizer = optimizer,
        metrics   = list("mean_squared_error"))

if (!FLAGS$stateful) {
    model %>% fit(
        x          = X_train,
        y          = y_train,
        validation_data = list(X_valid, y_valid),
        batch_size = FLAGS$batch_size,
        epochs     = FLAGS$n_epochs,
        callbacks  = callbacks)

} else
{
    for (i in 1:FLAGS$n_epochs)
    {
        model %>% fit(
            x          = X_train,
            y          = y_train,
            validation_data = list(X_valid, y_valid),
            callbacks  = callbacks,
            batch_size = FLAGS$batch_size,
            epochs     = 1,
            shuffle    = FALSE)
        model %>% reset_states()
    }
}

if (FLAGS$stateful)
    model %>% reset_states()

现在让我们逐步完成下面更简单但更好（或同样）的配置。

# create the model
model <- keras_model_sequential()

# add layers
# we have just two, the LSTM and the time_distributed
model %>%
    layer_lstm(
        units = FLAGS$n_units,
        # the first layer in a model needs to know the shape of the input data
        batch_input_shape  = c(
            FLAGS$batch_size, FLAGS$n_timesteps, n_features),
        dropout = FLAGS$dropout,
        recurrent_dropout = FLAGS$recurrent_dropout,
        # by default, an LSTM just returns the final state
        return_sequences = TRUE) %>%
    time_distributed(layer_dense(units = 1))

model %>%
    compile(
        loss = FLAGS$loss,
        optimizer = optimizer,
        # in addition to the loss, Keras will inform us about current MSE while training
        metrics = list("mean_squared_error"))

history <- model %>% fit(
    x          = X_train,
    y          = y_train,
    validation_data = list(X_valid, y_valid),
    batch_size = FLAGS$batch_size,
    epochs     = FLAGS$n_epochs,
    callbacks  = callbacks)

正如我们所见，训练在约 55 个周期后停止，因为验证集损失不再减少。我们还发现验证集上的表现比训练集上的性能差——通常表明过度拟合。