CODEWITHSUNDEEP
pythontensorflowtensorflow2.0
Damian Grzanka
Damian Grzanka

Reputation: 305

Partial update tensor based on boolean mask in TensorFlow

I want to update part of the tensor based on some conditions.

I know that TensorFlow tensors are immutable so creating a new tensor would be ok for me. I tried tensor_scatter_nd_update method but I couldn't make it work

This is code that I want to replicate in TensorFlow written in NumPy.

import numpy as np

a = np.random.random((1, 3))
b = np.array([[0, 1, 0]])

c = np.zeros_like(a)
mask = b == 1
c[mask] = np.log(a[mask])

Upvotes: 2

Views: 925

Answers (1)

Alexey Tochin
Alexey Tochin

Reputation: 683

In TensorFlow, we do not update tensors that are in fact immutable objects. Instead we create new tensors from other tensors like in functional languages.

import tensorflow as tf

a = tf.random.uniform(shape=(1, 3))
b = tf.constant([[0, 1, 0]], dtype=tf.int32)

c = tf.zeros_like(a)
mask = b == 1
c_updated = tf.where(mask, tf.math.log(a), c)
# [[ 0.      , -4.175911,  0.      ]]```

Upvotes: 4

Related Questions