CNN可学习参数的数量--Python/TensorFlow

Question

在TensorFlow中,我能做些什么来找出网络中学习参数的数量？

Answer 1

没有我知道的功能,但你仍然可以使用for循环来计算自己 tf.trainable_variables():

total_parameters = 0
for variable in tf.trainable_variables():
    variable_parameters = 1
    for dim in variable.get_shape():
        variable_parameters *= dim.value
    total_parameters += variable_parameters

print("Total number of trainable parameters: %d" % total_parameters)

Answer 2

你可以用一句简单的话来做到这一点：

np.sum([np.prod(v.get_shape().as_list()) for v in tf.trainable_variables()])

如果您需要更多详细信息，这里是我用来查看所有可训练参数的辅助函数：

def show_params():
  total = 0
  for v in tf.trainable_variables():
    dims = v.get_shape().as_list()
    num  = int(np.prod(dims))
    total += num
    print('  %s \t\t Num: %d \t\t Shape %s ' % (v.name, num, dims))
  print('\nTotal number of params: %d' % total)

它会打印如下信息：

  params/weights/W1:0        Num: 34992      Shape [3, 3, 18, 216] 
  params/weights/W2:0        Num: 839808     Shape [3, 3, 216, 432] 
  params/weights/W3:0        Num: 839808     Shape [3, 3, 432, 216] 
  params/weights/W4:0        Num: 57856      Shape [226, 256] 
  params/weights/W5:0        Num: 32768      Shape [256, 128] 
  params/weights/W6:0        Num: 8192       Shape [128, 64] 
  params/weights/W7:0        Num: 64         Shape [64, 1] 
  params/biases/b1:0         Num: 216        Shape [216] 
  params/biases/b2:0         Num: 432        Shape [432] 
  params/biases/b3:0         Num: 216        Shape [216] 
  params/biases/b4:0         Num: 256        Shape [256] 
  params/biases/b5:0         Num: 128        Shape [128] 
  params/biases/b6:0         Num: 64         Shape [64] 
  params/biases/b7:0         Num: 1          Shape [1]

Total number of params: 1814801