CODEWITHSUNDEEP
tensorflow
armundle
armundle

Reputation: 1179

tf.reshape loses shape of the tensor

When I reshape a tensor, tf.reshape as well as <Tensor>.shape are unable to infer the shape. If I understood this Issue correctly: https://github.com/tensorflow/tensorflow/issues/3311 , it should have been fixed.

Can anyone help me where I might be missing something?

import tensorflow as tf

sess = tf.InteractiveSession()

m = 100
n = 300

x = 123
y = 456

a = tf.get_variable(dtype=tf.int32, shape=[m, n, x], name="a")
b = tf.get_variable(dtype=tf.int32, shape=[m, n, y], name="b")

print(a.shape) # => (100, 300, 123)
print(b.shape) # => (100, 300, 456)
print(tf.shape(a)) # => Tensor("Shape_4:0", shape=(3,), dtype=int32)
print(tf.shape(b)) # => Tensor("Shape_5:0", shape=(3,), dtype=int32

c = tf.concat([a, b], axis=-1)

print(c.shape) # => (100, 300, 579)
print(tf.shape(c)) # = >Tensor("Shape:0", shape=(3,), dtype=int32)

s = tf.shape(c)
cc = tf.reshape(c, [s[0]*s[1], -1])

print(cc.shape)  # => (?, ?)
print(tf.shape(cc)) # => Tensor("Shape_3:0", shape=(2,), dtype=int32)

Upvotes: 1

Views: 596

Answers (1)

JoshVarty
JoshVarty

Reputation: 9426

I think you want to use:

s = c.get_shape().as_list()

or

s = c.shape.as_list()

I've never really used tf.shape() myself, but when I use the above I receive the proper shape (30000, 579)

Upvotes: 1

Related Questions