Gotcha! Tensor Shape

While working on the second part of my style-transfer project, I needed to obtain the shape of a tensor. I decided to try using the tf.shape function.

Reading the example provided on the documentation, it seemed like it would do what I needed, nevermind the odd:

This operation returns a 1-D integer tensor representing the shape of input.

When I tried this in a notebook:

dummy_data = np.random.rand(1,4,2)
dummy_tensor = tf.constant(dummy_data)
print("TF shape function outside a session: {}".format(tf.shape(dummy_tensor)))

The behaviour was unexpected:

TF shape function outside a session: Tensor(“Shape:0”, shape=(3,), dtype=int32)

I had a hunch about what I was doing wrong so I tried the following:

with tf.Session() as sess:
    dummy_tensor_shape_ = sess.run(dummy_tensor_shape)
    print("TF shape function inside a session: {}".format(dummy_tensor_shape_))

Which produced the following:

TF shape function inside a session: [1 4 2]

As expected!

Once again, I was reminded that tensor operations will only provide what you want within a session! Before that, the tensor is simply an operation waiting to be executed. The 3 noted in the shape outside the session seems to be the length of the list containing the actual shape.

Thankfully, TF offers another function: get_shape() Note that get_shape is just an alias for shape. This method infers the static shape but can fail in some cases, for example where the input data shape isn’t known until runtime. This explanation provides a clear example of a failure case.

The TF documentation notes that this function can provide debugging information and early warnings. My use case was slightly different in that I actually used the shape information in a loss calcuation. So perhaps I should have figured out how to use tf.shape properly within the graph. However, get_shape worked fine and it was easier to understand at the time so it is what I ended up using.

If you’d like to see for yourself the difference ways of getting a tensor’s shape, here’s a notebook for you to tinker with!

• 2020 3
• 2019 3
• 2018 11