104

Suppose I have a Tensorflow tensor. How do I get the dimensions (shape) of the tensor as integer values? I know there are two methods, tensor.get_shape() and tf.shape(tensor), but I can't get the shape values as integer int32 values.

For example, below I've created a 2-D tensor, and I need to get the number of rows and columns as int32 so that I can call reshape() to create a tensor of shape (num_rows * num_cols, 1). However, the method tensor.get_shape() returns values as Dimension type, not int32.

import tensorflow as tf
import numpy as np

sess = tf.Session()    
tensor = tf.convert_to_tensor(np.array([[1001,1002,1003],[3,4,5]]), dtype=tf.float32)

sess.run(tensor)    
# array([[ 1001.,  1002.,  1003.],
#        [    3.,     4.,     5.]], dtype=float32)

tensor_shape = tensor.get_shape()    
tensor_shape
# TensorShape([Dimension(2), Dimension(3)])    
print tensor_shape    
# (2, 3)

num_rows = tensor_shape[0] # ???
num_cols = tensor_shape[1] # ???

tensor2 = tf.reshape(tensor, (num_rows*num_cols, 1))    
# Traceback (most recent call last):
#   File "<stdin>", line 1, in <module>
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.py", line 1750, in reshape
#     name=name)
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 454, in apply_op
#     as_ref=input_arg.is_ref)
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 621, in convert_to_tensor
#     ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/constant_op.py", line 180, in _constant_tensor_conversion_function
#     return constant(v, dtype=dtype, name=name)
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/constant_op.py", line 163, in constant
#     tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape))
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/tensor_util.py", line 353, in make_tensor_proto
#     _AssertCompatible(values, dtype)
#   File "/usr/local/lib/python2.7/site-packages/tensorflow/python/framework/tensor_util.py", line 290, in _AssertCompatible
#     (dtype.name, repr(mismatch), type(mismatch).__name__))
# TypeError: Expected int32, got Dimension(6) of type 'Dimension' instead.
Joel Carneiro
  • 3,287
  • 3
  • 16
  • 20
stackoverflowuser2010
  • 38,621
  • 48
  • 169
  • 217

6 Answers6

146

To get the shape as a list of ints, do tensor.get_shape().as_list().

To complete your tf.shape() call, try tensor2 = tf.reshape(tensor, tf.TensorShape([num_rows*num_cols, 1])). Or you can directly do tensor2 = tf.reshape(tensor, tf.TensorShape([-1, 1])) where its first dimension can be inferred.

martianwars
  • 6,380
  • 5
  • 35
  • 44
yuefengz
  • 3,338
  • 1
  • 17
  • 24
  • Thanks, that lets me call and complete `tf.reshape()`, but I would really like to get `num_rows` and `num_cols` as integers for other operations. – stackoverflowuser2010 Nov 17 '16 at 22:45
  • 10
    Try `tensor.get_shape().as_list()` – yuefengz Nov 17 '16 at 22:47
  • 3
    For completeness, this code works: `num_rows, num_cols = x.get_shape().as_list()` – stackoverflowuser2010 Nov 17 '16 at 23:35
  • 1
    Nice! I was using python int() to cast the results of x.get_shape(). ie num_rows=int(x.get_shape()[1]), num_cols=int(x.get_shape()[2]), etc.Yep, kinda a hacky to get around that pesky error, but it worked. Thanks for enlightening me to a better way :-) – SherylHohman Mar 17 '17 at 18:54
  • In TF2.0, it seems not work: as_list() is not defined on an unknown TensorShape – EyesBear Jul 25 '19 at 08:38
  • Can anyone say whether this also applies within custom loss functions? I get "AttributeError: 'Tensor' object has no attribute 'as_list'" when I try to use within a loss function. – Jason K. May 16 '21 at 15:19
33

Another way to solve this is like this:

tensor_shape[0].value

This will return the int value of the Dimension object.

maazza
  • 7,016
  • 15
  • 63
  • 96
tijmen Verhulsdonck
  • 356
  • 3
  • 4
13

2.0 Compatible Answer: In Tensorflow 2.x (2.1), you can get the dimensions (shape) of the tensor as integer values, as shown in the Code below:

Method 1 (using tf.shape):

import tensorflow as tf
c = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
Shape = c.shape.as_list()
print(Shape)   # [2,3]

Method 2 (using tf.get_shape()):

import tensorflow as tf
c = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
Shape = c.get_shape().as_list()
print(Shape)   # [2,3]
8

for a 2-D tensor, you can get the number of rows and columns as int32 using the following code:

rows, columns = map(lambda i: i.value, tensor.get_shape())
Anna
  • 91
  • 1
  • 2
2

Another simple solution is to use map() as follows:

tensor_shape = map(int, my_tensor.shape)

This converts all the Dimension objects to int

Avandale
  • 242
  • 1
  • 7
0

In later versions (tested with TensorFlow 1.14) there's a more numpy-like way to get the shape of a tensor. You can use tensor.shape to get the shape of the tensor.

tensor_shape = tensor.shape
print(tensor_shape)
thushv89
  • 10,865
  • 1
  • 26
  • 39