Question

我正在学习GAN，并希望在R中编写一个程序，使用来自MNIST数据集的数据训练GAN模型生成手写数字。

以下是我的代码：

datasets <- tf$contrib$learn$datasets
mnist <- datasets$mnist$read_data_sets("MNIST-data")


discriminator <- function(images, reuse = FALSE){
  if (reuse) {
    tf$get_variable_scope()$reuse_variables()
  }

  # First convolutional layer
  d_w1 = tf$get_variable(name = "d_w1", dtype = tf$float32, shape = list(5L, 5L, 1L, 32L), initializer = tf$truncated_normal_initializer(stddev = 0.2))
  d_b1 = tf$get_variable(name = "d_b1", dtype = tf$float32, shape = list(32L), initializer = tf$truncated_normal_initializer(stddev = 0.2))
  d1 = tf$nn$conv2d(input=images, filter=d_w1, strides=list(1L, 1L, 1L, 1L), padding='SAME')
  d1 = d1 + d_b1
  d1 = tf$nn$relu(d1)
  d1 = tf$nn$avg_pool(d1, ksize=list(1L, 2L, 2L, 1L), strides=list(1L, 2L, 2L, 1L), padding='SAME')

  # Second convolutional layer
  d_w2 = tf$get_variable(name = "d_w2", shape = list(5L, 5L, 32L, 64L), initializer=tf$truncated_normal_initializer(stddev=0.02))
  d_b2 = tf$get_variable(name = "d_b2", shape = list(64L), initializer=tf$constant_initializer(0))
  d2 = tf$nn$conv2d(input=d1, filter=d_w2, strides=list(1L, 1L, 1L, 1L), padding='SAME')
  d2 = d2 + d_b2
  d2 = tf$nn$relu(d2)
  d2 = tf$nn$avg_pool(d2, ksize=list(1L, 2L, 2L, 1L), strides=list(1L, 2L, 2L, 1L), padding='SAME')

  # First fully connected layer
  d_w3 = tf$get_variable(name = "d_w3", shape = list(7L * 7L * 128L, 1024L), initializer=tf$truncated_normal_initializer(stddev=0.02))
  d_b3 = tf$get_variable(name = "d_b3", shape = list(1024L), initializer=tf$constant_initializer(0))
  d3 = tf$reshape(d2, shape = list(-1L, as.integer(7*7*64)))
  d3 = tf$matmul(d3, d_w3)
  d3 = d3 + d_b3
  d3 = tf$nn$relu(d3)

  # Second fully connected layer
  d_w4 = tf$get_variable('d_w4', shape = list(1024L, 1L), initializer=tf$truncated_normal_initializer(stddev=0.02))
  d_b4 = tf$get_variable('d_b4', shape = list(1L), initializer=tf$constant_initializer(0))
  d4 = tf$matmul(d3, d_w4) + d_b4
  d4 = tf$sigmoid(d4) 
  # d4 contains unscaled values
  return(d4)
}


# Define the generator network
generator <- function(batch_size, z_dim, reuse = FALSE){
  if (reuse){
    tf$get_variable_scope()$reuse_variables()
  }

  z = tf$random_normal(shape = list(batch_size, z_dim), mean=0, stddev=1, name='z')
  g_w1 = tf$get_variable('g_w1', shape = list(z_dim, 3136L), dtype=tf$float32, initializer=tf$truncated_normal_initializer(stddev=0.02))
  g_b1 = tf$get_variable('g_b1', shape = list(3136L), initializer=tf$truncated_normal_initializer(stddev=0.02))
  g1 = tf$matmul(z, g_w1) + g_b1
  g1 = tf$reshape(g1, list(-1L, 56L, 56L, 1L))
  g1 = tf$nn$relu(g1)

  # Generate 50 features
  g_w2 = tf$get_variable('g_w2', shape = list(3L, 3L, 1L, as.integer(z_dim/2)), dtype=tf$float32, initializer=tf$truncated_normal_initializer(stddev=0.02))
  g_b2 = tf$get_variable('g_b2', shape = list(as.integer(z_dim/2)), initializer=tf$truncated_normal_initializer(stddev=0.02))
  g2 = tf$nn$conv2d(g1, g_w2, strides=list(1L, 2L, 2L, 1L), padding='SAME')
  g2 = g2 + g_b2
  g2 = tf$nn$relu(g2)
  g2 = tf$image$resize_images(g2, list(56L, 56L))

  # Generate 25 features
  g_w3 = tf$get_variable('g_w3', shape = list(3L, 3L, as.integer(z_dim/2), as.integer(z_dim/4)), dtype=tf$float32, initializer=tf$truncated_normal_initializer(stddev=0.02))
  g_b3 = tf$get_variable('g_b3', shape = list(as.integer(z_dim/4)), initializer=tf$truncated_normal_initializer(stddev=0.02))
  g3 = tf$nn$conv2d(g2, g_w3, strides=list(1L, 2L, 2L, 1L), padding='SAME')
  g3 = g3 + g_b3
  g3 = tf$nn$relu(g3)
  g3 = tf$image$resize_images(g3, list(56L, 56L))

  # Final convolution with one output channel
  g_w4 = tf$get_variable('g_w4', shape = list(1L, 1L, as.integer(z_dim/4), 1L), dtype=tf$float32, initializer=tf$truncated_normal_initializer(stddev=0.02))
  g_b4 = tf$get_variable('g_b4', shape = list(1L), initializer=tf$truncated_normal_initializer(stddev=0.02))
  g4 = tf$nn$conv2d(g3, g_w4, strides=list(1L, 2L, 2L, 1L), padding='SAME')
  g4 = g4 + g_b4
  g4 = tf$sigmoid(g4)

  # Dimensions of g4: batch_size x 28 x 28 x 1
  return(g4)
}

# Declare the batch size and the dimension of the generate output
z_dimensions = 100L
batch_size = 50L

# Create a placeholder for input
x_placeholder = tf$placeholder(tf$float32, shape = list(NULL,28L,28L,1L), name='x_placeholder')

# Gz holds the generated image
Gz = generator(batch_size, z_dimensions)

我收到了这个错误：

Error in py_call_impl(callable, dots$args, dots$keywords) : 
  ValueError: Variable g_w1 does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?

我试图修正错误：

1）我使用tf$get_variable()而不是tf$Variable()。错误消失了，但根据这个 discussion，我应该使用tf$get_variable()而不是tf$Variable，因为只有tf$get_variable()创建的变量才会被变量范围识别。

2）我试图将Gz的赋值包含在这样的范围内：

with(tf$variable_scope(tf$get_variable_scope(), reuse = FALSE),{
      Gz = generator(batch_size, z_dimensions)
   })

但仍然存在同样的错误。

我很感激有关如何解决此错误的任何建议。非常感谢你！

Tensorflow R：变量不存在，或者不是用tf.get_variable（）创建的

0 个答案: