How to update a subset of 2D tensor in Tensorflow?

Question

I want to update an index in a 2D tensor with value 0. So data is a 2D tensor whose 2nd row 2nd column index value is to be replaced by 0. However, I am getting a type error. Can anyone help me with it?

TypeError: Input 'ref' of 'ScatterUpdate' Op requires l-value input

data = tf.Variable([[1,2,3,4,5], [6,7,8,9,0], [1,2,3,4,5]])
data2 = tf.reshape(data, [-1])
sparse_update = tf.scatter_update(data2, tf.constant([7]), tf.constant([0]))
#data = tf.reshape(data, [N,S])
init_op = tf.initialize_all_variables()

sess = tf.Session()
sess.run([init_op])
print "Values before:", sess.run([data])
#sess.run([updated_data_subset])
print "Values after:", sess.run([sparse_update])

In general, depending on how complex the updating is and on whether you plan to do this process as part of a big graph, you may want to check out `tf.py_func`, a wrapper that allows you to add numpy operations within tf graphs. — dopexxx, Jul 09 '18 at 07:31

score 11 · Answer 1 · edited Sep 14 '17 at 14:21

11

Scatter update only works on variables. Instead try this pattern.

Tensorflow version < 1.0: a = tf.concat(0, [a[:i], [updated_value], a[i+1:]])

Tensorflow version >= 1.0: a = tf.concat(axis=0, values=[a[:i], [updated_value], a[i+1:]])

edited Sep 14 '17 at 14:21

Sogartar

2,035
2
19
35

answered Oct 04 '16 at 19:39

Yaroslav Bulatov

57,332
22
139
197

Extremely helpful, answering the question that should have been asked – jon_simon Nov 30 '19 at 22:25
1

Why doesn't tensorflow 2 have a function for this? – user2585501 Jan 06 '21 at 06:10

Da Tong · Accepted Answer · 2017-04-07T03:26:19.947

3

tf.scatter_update could only be applied to Variable type. data in your code IS a Variable, while data2 IS NOT, because the return type of tf.reshape is Tensor.

Solution:

for tensorflow after v1.0

data = tf.Variable([[1,2,3,4,5], [6,7,8,9,0], [1,2,3,4,5]])
row = tf.gather(data, 2)
new_row = tf.concat([row[:2], tf.constant([0]), row[3:]], axis=0)
sparse_update = tf.scatter_update(data, tf.constant(2), new_row)

for tensorflow before v1.0

data = tf.Variable([[1,2,3,4,5], [6,7,8,9,0], [1,2,3,4,5]])
row = tf.gather(data, 2)
new_row = tf.concat(0, [row[:2], tf.constant([0]), row[3:]])
sparse_update = tf.scatter_update(data, tf.constant(2), new_row)

edited Apr 07 '17 at 03:26

answered Oct 10 '16 at 13:51

Da Tong

2,018
18
25

1

It does not work: ValueError: Dimension 0 in both shapes must be equal, but are 1 and 2 From merging shape 1 with other shapes. for 'concat/concat_dim' (op: 'Pack') with input shapes: [2], [1], [2]. – Antonio Sesto Apr 05 '17 at 11:35
1

@AntonioSesto TensorFlow v1.0 changed some interface to keep consistent with numpy. `concat` was one of the changes. I have edited my answer to work with v1.0. – Da Tong Apr 07 '17 at 03:27

score 0 · Answer 3 · answered Oct 08 '21 at 02:35

This is the function I use to modify a subset (row/column) of a 2D tensor in Tensorflow 2:

#note if updatedValue isVector, updatedValue should be provided in 2D format
def modifyTensorRowColumn(a, isRow, index, updatedValue, isVector):
    
    if(not isRow):
        a = tf.transpose(a)
        if(isVector):
            updatedValue = tf.transpose(updatedValue)
    
    if(index == 0):
        if(isVector):
            values = [updatedValue, a[index+1:]]
        else:
            values = [[updatedValue], a[index+1:]]
    elif(index == a.shape[0]-1):
        if(isVector):
            values = [a[:index], updatedValue]
        else:
            values = [a[:index], [updatedValue]]
    else:
        if(isVector):
            values = [a[:index], updatedValue, a[index+1:]]
        else:
            values = [a[:index], [updatedValue], a[index+1:]]
            
    a = tf.concat(axis=0, values=values)
            
    if(not isRow):
        a = tf.transpose(a)
        
    return a

How to update a subset of 2D tensor in Tensorflow?

3 Answers3

Linked