Wsrtjtyk

I have to tensors, for example

A = tf.Tensor(

       [[1.0986123 0.6931472 0.        0.6931472 0.       ]

        [0.        0.        0.        0.        0.       ]

        [3.7376697 3.7612002 3.7841897 3.8066626 3.8286414]], shape=(3, 5), dtype=float32)



B = tf.Tensor(

   [[2 1]

    [2 2]], shape=(2, 2), dtype=int64)

Tensor B holds indices in tensor A. I want to update every value in tensor A to zero that is listed in the index list B.

So, the expected result would be

tf.Tensor(

       [[1.0986123 0.6931472 0.        0.6931472 0.       ]

        [0.        0.        0.        0.        0.       ]

        [3.7376697 0 0 3.8066626 3.8286414]], shape=(3, 5), dtype=float32)

So the entries at index [2,1] and [2, 2] are set to 0.

I looked at tf.assign but they can only be used for tf.Variable's. tf.boolean_mask would be a nice way to do it, but i do not know and could not find out how i can create a boolean mask with a list of indices.

I looked at the tensor flow functions i could find and related S/O answers but couldn't find a satisfying solution.

You can use tf.scatter_nd_update for this. For example:

A = tf.Variable(

    [[1.0986123, 0.6931472, 0.       , 0.6931472, 0.       ],

     [0.       , 0.       , 0.       , 0.       , 0.       ],

     [3.7376697, 3.7612002, 3.7841897, 3.8066626, 3.8286414]], dtype=tf.float32)



B = tf.Variable(

    [[2, 1],

     [2, 2]], dtype=tf.int64)



C = tf.scatter_nd_update(A, B, tf.zeros(shape=tf.shape(B)[0]))



with tf.Session() as sess:

    sess.run(tf.global_variables_initializer())

    print(sess.run(C))

or

A = tf.constant(

    [[1.0986123, 0.6931472, 0.       , 0.6931472, 0.       ],

     [0.       , 0.       , 0.       , 0.       , 0.       ],

     [3.7376697, 3.7612002, 3.7841897, 3.8066626, 3.8286414]], dtype=tf.float32)



B = tf.constant(

    [[2, 1],

     [2, 2]], dtype=tf.int64)



AV = tf.Variable(A)



C = tf.scatter_nd_update(AV, B, tf.zeros(shape=tf.shape(B)[0]))



with tf.Session() as sess:

    sess.run(tf.global_variables_initializer())

    print(sess.run(C))