R 包“keras3”错误：OP_REQUIRES 在 reshape_op.h:65 处失败：INVALID_ARGUMENT：只有一个输入大小可以为 -1，不能同时为 0 和 1

Question

我正在按照“使用 R 进行深度学习，第二版”一书第 10 章中的示例，使用 R 包

keras3

。出现错误：W tensorflow/core/framework/op_kernel.cc:1841] OP_REQUIRES failed at reshape_op.h:65 : INVALID_ARGUMENT: 只有一个输入大小可以为 -1，不能同时为 0 和 1。

我是 R 深度学习框架的新手，我想知道如何修复这个错误？非常感谢！

本书在此链接中提供了数据集。

library(keras3)
library(dplyr)

# Preparing the Data 
full_df <- readr::read_csv("jena_climate_2009_2016.csv")

full_df$`Date Time` %<>%                     
  as.POSIXct(tz = "Etc/GMT+1", format = "%d.%m.%Y %H:%M:%S")

num_train_samples <- round(nrow(full_df) * 0.5)
num_val_samples <- round(nrow(full_df) * 0.25)
num_test_samples <- nrow(full_df) - num_train_samples - num_val_samples

train_df <- full_df[seq(num_train_samples), ]
val_df <- full_df[seq(from = nrow(train_df) + 1,
                  length.out = num_val_samples), ]
test_df <- full_df[seq(to = nrow(full_df),
                   length.out = num_test_samples), ]

## Normalize the data
input_data_colnames <- names(full_df) %>%
  setdiff(c("Date Time"))                                  

normalization_values <-
  zip_lists(mean = lapply(train_df[input_data_colnames], mean),
            sd = lapply(train_df[input_data_colnames], sd))
str(normalization_values)

normalize_input_data <- function(df) {
  normalize <- function(x, center, scale)
    (x - center) / scale

  for(col_nm in input_data_colnames) {
    col_nv <- normalization_values[[col_nm]]
    df[[col_nm]] %<>% normalize(., col_nv$mean, col_nv$sd)
  }
  df
}

## Instantiating datasets for training, validation, and testing
sampling_rate <- 6
sequence_length <- 120
delay <- sampling_rate * (sequence_length + 24 - 1)
batch_size <- 256

df_to_inputs_and_targets <- function(df) {
  inputs <- df[input_data_colnames] %>%
    normalize_input_data() %>%
    as.matrix()

  targets <- as.array(df$`T (degC)`)

  list(head(inputs, -delay), tail(targets, -delay))           
}

make_dataset <- function(df) {
  c(inputs, targets) %<-% df_to_inputs_and_targets(df)       

  timeseries_dataset_from_array(                                              
    inputs, targets,
    sampling_rate = sampling_rate,
    sequence_length = sequence_length,
    shuffle = TRUE,
    batch_size = batch_size
  )
}

train_dataset <- make_dataset(train_df)
val_dataset <- make_dataset(val_df)
test_dataset <- make_dataset(test_df)


## A densely connected model
ncol_input_data <- length(input_data_colnames)

inputs <- layer_input(shape = c(sequence_length, ncol_input_data))

outputs <- inputs %>%
  layer_flatten() %>%
  layer_dense(16, activation = "relu") %>%
  layer_dense(1)

model <- keras_model(inputs, outputs)

callbacks = list(
  callback_model_checkpoint("jena_dense.keras",
                            save_best_only = TRUE)
)

model %>%
  compile(optimizer = "rmsprop",
          loss = "mse",
          metrics = "mae")

history <- model %>%
  fit(train_dataset,
      epochs = 10,
      validation_data = val_dataset,
      callbacks = callbacks)

model <- load_model_tf("jena_dense.keras")
sprintf("Test MAE: %.2f", evaluate(model, test_dataset)["mae"])

Answer 1

这似乎是一个“已知问题”，在使用 keras2 的代码运行 keras3 时会发生。解决这个问题的简单方法是您应该使用 keras 而不是像书中原始代码中那样的

keras3

库library(keras)