Deep learning和tensorflow学习记录（二十八）：static and dynamic shapes

tensorflow中tensor的static shape是在graph构建的时候就已经确定了。

import tensorflow as tf

a = tf.placeholder(tf.float32, [None, 128])
static_shape = a.shape.as_list()
dynamic_shape = tf.shape(a)

print(static_shape)
print(dynamic_shape)

None表示不确定，可以为任何尺寸，可以在Session.run()的时候动态确定。tf.shape()获取dynamic shape，返回的是一个tensor。

输出：

[None, 128]
Tensor("Shape:0", shape=(2,), dtype=int32)

可通过set_shape()设置tensor的shape，

a.set_shape([32, 128])
static_shape = a.shape.as_list()
dynamic_shape = tf.shape(a)
print(static_shape)
print(dynamic_shape)

输出：

[32, 128]
Tensor("Shape:0", shape=(2,), dtype=int32)

也可以通过tf.reshape()重设shape，

a = tf.placeholder(tf.float32, [None, 128])
a =  tf.reshape(a, [32, 128])
static_shape = a.shape.as_list()
dynamic_shape = tf.shape(a)
print(static_shape)
print(dynamic_shape)

输出：

[32, 128]
Tensor("Shape:0", shape=(2,), dtype=int32)

可以定义一个函数，当static shape为None时返回dynamic shape，否则返回static shape。

def get_shape(tensor):
  static_shape = tensor.shape.as_list()
  dynamic_shape = tf.unstack(tf.shape(tensor))
  dims = [s[1] if s[0] is None else s[0]
          for s in zip(static_shape, dynamic_shape)]
  return dims

b = tf.placeholder(tf.float32, [None, 10, 32])
shape = get_shape(b)
print(shape)
b = tf.reshape(b, [shape[0], shape[1] * shape[2]])
shape = get_shape(b)
print(shape)

输出：

[<tf.Tensor 'unstack:0' shape=() dtype=int32>, 10, 32]
[<tf.Tensor 'unstack_1:0' shape=() dtype=int32>, 320]