This question is with respect to accessing individual elements in a tensor, say [[1,2,3]]. I need to access the inner element [1,2,3] (This can be performed using .eval() or sess.run()) but it takes longer when the size of the tensor is huge)
Is there any method to do the same faster?
Thanks in Advance.
There are two main ways to access subsets of the elements in a tensor, either of which should work for your example.

Use the indexing operator (based on
tf.slice()
) to extract a contiguous slice from the tensor.input = tf.constant([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) output = input[0, :] print sess.run(output) # ==> [1 2 3]
The indexing operator supports many of the same slice specifications as NumPy does.

Use the
tf.gather()
op to select a noncontiguous slice from the tensor.input = tf.constant([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) output = tf.gather(input, 0) print sess.run(output) # ==> [1 2 3] output = tf.gather(input, [0, 2]) print sess.run(output) # ==> [[1 2 3] [7 8 9]]
Note that
tf.gather()
only allows you to select whole slices in the 0th dimension (whole rows in the example of a matrix), so you may need totf.reshape()
ortf.transpose()
your input to obtain the appropriate elements.
You simply can’t get value of 0th element of [[1,2,3]] without run()ning or eval()ing an operation which would be getting it. Because before you ‘run’ or ‘eval’, you have only a description how to get this inner element(because TF uses symbolic graphs/calculations). So even if you would use tf.gather/tf.slice, you still would have to get values of these operations via eval/run. See @mrry’s answer.
I suspect it’s the rest of the computation that takes time, rather than accessing one element.
Also the result might require a copy from whatever memory is stored in, so if it’s on the graphics card it will need to be copied back to RAM first and then you get access to your element. If this is the case you might skip it by adding an tensorflow operation to take the first element, and only return that.
Tags: python, tensorflow