Why Tensor.get_shape() can work immediately before initializing the session?

Question

As indicating in many TF docs, the graph begin to build only when we create the sess and execute the sess.run() method.

Just now, I tested that:

X = tf.placeholder(dtype=tf.float32, shape=[100,28,28,1])
W = tf.get_variable(...)
Y = tf.matmul(X,W)
print(Y.get_shape()[0].value)

As see in this code, it outputs the particular the shape value of Y even I did not t create the session instance.

How to understand? Thank you!

You don't need to calculate the matrix multiplication to determine the shape of the result probably would be one of the reasons. For instance a 3X2 tensor times a 2X5 matrix will always end up with a 3X5 matrix, you can predetermine this without the actual calculation. — Psidom, Aug 24 '17 at 01:03

score 5 · Accepted Answer · answered Aug 24 '17 at 01:34

5

There's "static shape" and "dynamic shape", get_shape() is static and only needs graph, tf.shape is dynamic and needs session -- How to understand static shape and dynamic shape in TensorFlow?

answered Aug 24 '17 at 01:34

Yaroslav Bulatov

Thank you very much and Now I fully understood the shape method. – Kevin Sun Aug 24 '17 at 03:21

1 Answers1