R 中的 TF 估计器 API

我们在第 2 章中了解了 TensorFlow 估计器 API。在 R 中，此 API 使用 tfestimator R 包实现。

例如，我们提供了 MLP 模型的演练，用于在以下链接中对来自 MNIST 数据集的手写数字进行分类： https://tensorflow.rstudio.com/tfestimators/articles/examples/mnist.html 。

您可以按照 Jupyter R 笔记本中的代码ch-17b_TFE_Ttimator_in_R。

1. 首先，加载库：

library(tensorflow)
library(tfestimators)

1. 定义超参数：

batch_size <- 128
n_classes <- 10
n_steps <- 100

1. 准备数据：

# initialize data directory
data_dir <- "~/datasets/mnist"
dir.create(data_dir, recursive = TRUE, showWarnings = FALSE)

# download the MNIST data sets, and read them into R
sources <- list(
  train = list(
    x = "https://storage.googleapis.com/cvdf-datasets/mnist/train-images-idx3-ubyte.gz",
    y = "https://storage.googleapis.com/cvdf-datasets/mnist/train-labels-idx1-ubyte.gz"
  ),
  test = list(
    x = "https://storage.googleapis.com/cvdf-datasets/mnist/t10k-images-idx3-ubyte.gz",
    y = "https://storage.googleapis.com/cvdf-datasets/mnist/t10k-labels-idx1-ubyte.gz"
  )
)

# read an MNIST file (encoded in IDX format)
read_idx <- function(file) {

  # create binary connection to file
  conn <- gzfile(file, open = "rb")
  on.exit(close(conn), add = TRUE)

  # read the magic number as sequence of 4 bytes
  magic <- readBin(conn, what="raw", n=4, endian="big")
  ndims <- as.integer(magic[[4]])

  # read the dimensions (32-bit integers)
  dims <- readBin(conn,what="integer",n=ndims,endian="big")

  # read the rest in as a raw vector
  data <- readBin(conn,what="raw",n=prod(dims),endian="big")

  # convert to an integer vecto
  converted <- as.integer(data)

  # return plain vector for 1-dim array
  if (length(dims) == 1)
    return(converted)

  # wrap 3D data into matrix
  matrix(converted,nrow=dims[1],ncol=prod(dims[-1]),byrow=TRUE)
}

mnist <- rapply(sources,classes="character",how ="list",function(url) {
  # download + extract the file at the URL
  target <- file.path(data_dir, basename(url))
  if (!file.exists(target))
    download.file(url, target)

  # read the IDX file
  read_idx(target)
})

# convert training data intensities to 0-1 range
mnist$train$x <- mnist$train$x / 255
mnist$test$x <- mnist$test$x / 255

从下载的 gzip 文件中读取数据，然后归一化以落入[0,1]范围。

1. 定义模型：

# construct a linear classifier
classifier <- linear_classifier(
  feature_columns = feature_columns(
    column_numeric("x", shape = shape(784L))
  ),
  n_classes = n_classes # 10 digits
)

# construct an input function generator
mnist_input_fn <- function(data, ...) {
  input_fn(
    data,
    response = "y",
    features = "x",
    batch_size = batch_size,
    ...
  )
}

1. 训练模型：

train(classifier,input_fn=mnist_input_fn(mnist$train),steps=n_steps)

1. 评估模型：

evaluate(classifier,input_fn=mnist_input_fn(mnist$test),steps=200)

输出如下：

Evaluation completed after 79 steps but 200 steps was specified

太酷!!

通过以下链接查找 R 中 TF 估计器的更多示例：https://tensorflow.rstudio.com/tfestimators/articles/examples/

有关tensorflow R 包的更多文档可以在以下链接中找到：https://tensorflow.rstudio.com/tfestimators/reference/