Models: TypeError: Failed to convert object of type <class 'list'> to Tensor. Contents: [1, 1, Dimension(None)]. Consider casting elements to a supported type.

[nd7141]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux
• TensorFlow version (use command below): 1.12

I'm trying to reshape some of my data with Dataset API. In particular, I have created a dataset like this:

dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(parse).filter(filter_empty).batch(batch_size)

After this if I print my shape within map function I get the following:

def print_shape(x):
    print(x.shape)

dataset.map(print_shape)

>>> (?, ?, ?)

So the shape of the batch is not clear and hence if I want to reshape some of my variables by accessing them via x.shape, it fails with an error:
TypeError: Failed to convert object of type <class 'list'> to Tensor. Contents: [1, 1, Dimension(None)]. Consider casting elements to a supported type.

How can I access shape within dataset creation so that I cam reshape some of my variables?

[nd7141]

Another detail: I use eager mode.

[nd7141]

Solved via using tf.shape instead.

[Shruthi-Sampathkumar]

Solved via using tf.shape instead.

How did you use tf.shape to solve this issue?

[sachin5nowal]

I have been facing similar issues ,try using tf.expand_dims.
x=tf.expand_dims(x, -1)
This helped me sorting the issue.

[Liusandian]

Solved via using tf.shape instead.

How did you use tf.shape to solve this issue? could you please give more details about the solution?

[wookayin]

@Liusandian @Shruthi-Sampathkumar

Let's assume x is a Tensor of shape [10, ?, ?]. It can be a tensor with a undetermined size when it is TF 1.x, running non-eagerly on TF 2.x (e.g. using tf.function + autograph), or running inside tf.Dataset (as in @nd7141's case).

x.shape will give you [10, None, None]. You cannot reshape it because you do not know the dimension statically. For example, some tensor manipulations tf.reshape(x, [-1, x.shape[1], x.shape[0]]) will fail.

However, h = tf.shape(x)[1]; w = tf.shape(x)[2] will let h, w be symbolic or graph-mode tensors (integer) that will contain a dimension of x. The value will be determined runtime. In such a case, tf.reshape(x, [-1, w, h]) will produce a (symbolic) tensor of shape [?, ?, ?] (still unknown) whose tensor shape will be known on runtime.

TL;DR) tf.shape(x) gives a tf.Tensor of integer dtype whose elements are not None (it is just undetermined), where as x.shape gives a static value as a list/tuple. But if any shape dimension is unknown, it will be None.