Indexing a 1D tensor in tensorflow

Question

Is it possible to get the element from a tensor at a given index in order to obtain a scalar? For example given an image I can retrieve its shape with shape = tf.shape(image) , but how can I retrieve its height, width and depth?

The only way I found is the following:

height = tf.reshape(tf.slice(shape, [0], [1]), [])
width = tf.reshape(tf.slice(shape, [1], [1]), [])
depth = tf.reshape(tf.slice(shape, [2], [1]), [])

Is there any other way?

Answer 1

The slice syntax (ie using the [] operator) is based on NumPy slicing, and gives a slightly more concise way of getting the height, width and depth from a shape tensor:

shape = tf.shape(image)
height = shape[0]  # returns a scalar
width = shape[1]   # returns a scalar
depth = shape[2]   # returns a scalar

Nessuno's answer will also work well if the tensor has a statically determined shape. However, variable-sized images (eg the result of tf.image.decode_jpeg() ) will typically give None for the height and width dimensions when you use get_shape() , because these may vary from one image to the next.

Answer 2

Use tf.Variable.get_shape(variable) .

image = tf.Variable([ [ [1,2,3],[3,4,4],[1,1,1] ], [[1,2,3],[3,4,4],[1,1,1] ]])
shape = image.get_shape()
print( [shape[i].value for i in range(len(shape))] )

Outputs:

[2, 3, 3]