plot(history, metrics = "loss")

现在让我们看看该模型捕捉训练集特征的效果如何。

pred_train <- model %>%
    predict(
        X_train,
        batch_size = FLAGS$batch_size) %>%
    .[, , 1]

# Retransform values to original scale
pred_train <- (pred_train * scale_history + center_history) ^2
compare_train <- df %>%
    filter(key == "training")

# build a dataframe that has both actual and predicted values
for (i in 1:nrow(pred_train))
{
    varname <- paste0("pred_train", i)
    compare_train <- mutate(
        compare_train,
        !!varname := c(
            rep(NA, FLAGS$n_timesteps + i - 1),
            pred_train[i,],
            rep(NA,
                nrow(compare_train) - FLAGS$n_timesteps * 2 - i + 1)))
}

我们计算所有预测序列的平均 RSME。

coln <- colnames(compare_train)[4:ncol(compare_train)]
cols <- map(coln, quo(sym(.)))
rsme_train <- map_dbl(
    cols,
    function(col)
    {
        rmse(
            compare_train,
            truth = value,
            estimate = !!col,
            na.rm = TRUE)
    }) %>%
    mean()

rsme_train

21.01495

这些预测看起来如何？由于所有预测序列的可视化看起来会非常拥挤，我们会间隔地选择起始点。

ggplot(
    compare_train,
    aes(x = index, y = value)) +
    geom_line() +
    geom_line(aes(y = pred_train1), color = "cyan") +
    geom_line(aes(y = pred_train50), color = "red") +
    geom_line(aes(y = pred_train100), color = "green") +
    geom_line(aes(y = pred_train150), color = "violet") +
    geom_line(aes(y = pred_train200), color = "cyan") +
    geom_line(aes(y = pred_train250), color = "red") +
    geom_line(aes(y = pred_train300), color = "red") +
    geom_line(aes(y = pred_train350), color = "green") +
    geom_line(aes(y = pred_train400), color = "cyan") +
    geom_line(aes(y = pred_train450), color = "red") +
    geom_line(aes(y = pred_train500), color = "green") +
    geom_line(aes(y = pred_train550), color = "violet") +
    geom_line(aes(y = pred_train600), color = "cyan") +
    geom_line(aes(y = pred_train650), color = "red") +
    geom_line(aes(y = pred_train700), color = "red") +
    geom_line(aes(y = pred_train750), color = "green") +
    ggtitle("Predictions on the training set")

看起来像当好。对于验证集，我们并不认为有同样好的结果。

pred_test <- model %>%
    predict(
        X_test,
        batch_size = FLAGS$batch_size) %>%
    .[, , 1]

# Retransform values to original scale
pred_test <- (pred_test * scale_history + center_history) ^2
pred_test[1:10, 1:5] %>% print()
compare_test <- df %>% filter(key == "testing")

# build a dataframe that has both actual and predicted values
for (i in 1:nrow(pred_test))
{
    varname <- paste0("pred_test", i)
    compare_test <-mutate(
        compare_test,
        !!varname := c(
            rep(NA, FLAGS$n_timesteps + i - 1),
            pred_test[i,],
            rep(NA,
                nrow(compare_test) - FLAGS$n_timesteps * 2 - i + 1)))
}

compare_test %>% write_csv(
    str_replace(model_path, ".hdf5", ".test.csv"))
compare_test[FLAGS$n_timesteps:(FLAGS$n_timesteps + 10), c(2, 4:8)] %>%
    print()

coln <- colnames(compare_test)[4:ncol(compare_test)]
cols <- map(coln, quo(sym(.)))
rsme_test <- map_dbl(
    cols,
    function(col)
    {
        rmse(
            compare_test,
            truth = value,
            estimate = !!col,
            na.rm = TRUE)
    }) %>%
    mean()

rsme_test

31.31616