Tensorflow slice assignment.

slice_assignment.py

import tensorflow as tf

# tf.enable_eager_execution()  # uncomment if tf version < 2.0

def assign(x, slices, values):
    """Assign slicing-tensor to the given values.

    Args
    :x: A `Tensor` as input.
    :slices: A `Tensor`, `list` or `numpy array` indicates the region
    of the tensor that would be assigned.
    :values: A `Tensor` represents the assigned values.
    
    Returns
      A tensor where the specified region was assigned.
    """
    input_shape = tf.shape(x)
    shape = [(end - beg) for beg, end in slices]
    ones = tf.ones(shape)
    padding = [(beg, dim - end) for dim, (beg, end) in zip(input_shape, slices)]
    mask = tf.cast(tf.pad(ones, padding), x.dtype)
    padded_values = tf.cast(tf.pad(values, padding), x.dtype)
    return mask * padded_values + (1. - mask) * x


x = tf.random.uniform([3, 3], 0, 1)
slices = [(0, 2), (0, 2)]
values = tf.constant([[0, 0], [0, 0]])
x = assign(x, slices, values)

print(x)

As you know, Tensorflow hasn't provided a convenient way for sliced tensor assignment yet. I don't know why. It's a crucial feature. To get over that suck, I had to write this script. It's not perfect but anyway it's still helpful for me. It's heavily inspired by an idea I have found in a Tensorflow issue.