How can I change the shape of a variable in TensorFlow?

Question

TensorFlow tutorial says that at creation time we need to specify the shape of tensors. That shape automatically becomes the shape of the tensor. It also says that TensorF

Answer 1

The tf.Variable class is the recommended way to create variables, but it restricts your ability to change the shape of the variable once it has been created.

If you need to change the shape of a variable, you can do the following (e.g. for a 32-bit floating point tensor):

var = tf.Variable(tf.placeholder(tf.float32))
# ...
new_value = ...  # Tensor or numpy array.
change_shape_op = tf.assign(var, new_value, validate_shape=False)
# ...
sess.run(change_shape_op)  # Changes the shape of `var` to new_value's shape.

Note that this feature is not in the documented public API, so it is subject to change. If you do find yourself needing to use this feature, let us know, and we can investigate a way to support it moving forward.

Answer 2

As said by Mayou36, you can now change the variable shape after it was first declared. Here is a working example:

v = tf.Variable([1], shape=tf.TensorShape(None), dtype=tf.int32) 
tf.print(v)
v.assign([1, 1, 1])
tf.print(v)

And this outputs:

[1]
[1 1 1]

Answer 3

tf.Variable: Use the shape argument with None

A feature was added in 1.14 that allows to specify unknown shapes.

If shape is None, the initial shape value is used.

If shape is specified, this is used as the shape and allows to have None.

Example:

var = tf.Variable(array, shape=(None, 10))

This allows to later on assign values with shapes matching the shape above (e.g. arbitrary shapes in axis 0)

var.assign(new_value)

Answer 4

Documentation shows methods for reshaping. They are:

• reshape
• squeeze (removes dimensions of size 1 from the shape of a tensor)
• expand_dims (adds dimensions of size 1)

as well as bunch of methods to get shape, size, rank of your tensor. Probably the most used is reshape and here is a code example with a couple of edge cases (-1):

import tensorflow as tf

v1 = tf.Variable([
    [1, 2, 3, 4],
    [5, 6, 7, 8],
    [9, 10, 11, 12]
])
v2 = tf.reshape(v1, [2, 6])
v3 = tf.reshape(v1, [2, 2, -1])
v4 = tf.reshape(v1, [-1])
# v5 = tf.reshape(v1, [2, 4, -1]) will fail, because you can not find such an integer for -1
v6 = tf.reshape(v1, [1, 4, 1, 3, 1])
v6_shape = tf.shape(v6)
v6_squeezed = tf.squeeze(v6)
v6_squeezed_shape = tf.shape(v6_squeezed)

init = tf.initialize_all_variables()

sess = tf.Session()
sess.run(init)
a, b, c, d, e, f, g = sess.run([v2, v3, v4, v6, v6_shape, v6_squeezed, v6_squeezed_shape])
# print all variables to see what is there
print e # shape of v6
print g # shape of v6_squeezed

Answer 5

tf.Variable(tf.placeholder(tf.float32))

is not valid at tensorflow 1.2.1

in python shell:

import tensorflow as tf
tf.Variable(tf.placeholder(tf.float32))

You will get:

ValueError: initial_value must have a shape specified: Tensor("Placeholder:0", dtype=float32)

Update: if you add validate_shape=False, there will be no error.

tf.Variable(tf.placeholder(tf.float32), validate_shape=False)

if tf.py_func matches your requirement:

def init():
    return numpy.random.rand(2,3)
a = tf.pyfun(init, [], tf.float32)

You can create variable that has any shape by passing your own init function.

Another way:

var = tf.get_varible('my-name', initializer=init, shape=(1,1))

You can pass tf.constant or any init function that returns numpy array. The shape provided will not be validated. The output shape is your real data shape.

Answer 6

Take a look at shapes-and-shaping from TensorFlow documentation. It describes different shape transformations available.

The most common function is probably tf.reshape, which is similar to its numpy equivalent. It allows you to specify any shape that you want as long as the number of elements stays the same. There are some examples available in the documentation.