Tf.strided_slice(), what does new_axis_mask do?

TensorFlow General Discussion
,
1

Hello, I am quite confused about the parameter new_axis_mask of tf.strided_slice() function. The official documentation says:

If the ith bit of new_axis_mask is set, then begin, end, and stride are ignored and a new length 1 dimension is added at this point in the output tensor.
For example, foo[:4, tf.newaxis, :2] would produce a shape (4, 1, 2) tensor.

I did some tests…

import tensorflow as tf
import numpy as np

arr = np.random.rand(4,6,7)
out1 = tf.strided_slice(arr, begin=[1,1,1], end=[4,5,6], strides=[1,1,1])
print(out1.shape) // returns TensorShape([3, 4, 5])

So according to my understanding the result of(when using new_axis_mask)

out2 = tf.strided_slice(arr, begin=[1,1,1], end=[4,5,6], strides=[1,1,1], new_axis_mask=1)
print(out2.shape)

should be TensorShape([1, 3, 4, 5]) cause we add an extra dimension to the output at the first axis. However the result is TensorShape([1, 3, 5, 7]).

Can anyone explain me how the parameter new_axis_mask of tf.strided_slice() function works?
Code is here: Google Colab

3

Hi @kanye, when we pass the value to the new_axis_mask a new dimension will be added at the position indicated by the mask, and the slicing operation is then applied to the resulting array. so that might be the reason for you will get the [1, 3, 5, 7]. Thank You.