CODEWITHSUNDEEP
pythontensorflow
blacknight12321
blacknight12321

Reputation: 91

getting the higher value out of two tensors elementwise in tensorflow

I have 2 2d tensors of the same size (let's say shape =[80,90]). I want to get the value of the variable whose absolute value is bigger elementwise.

In python I would do something like this:

mask = np.abs(a)>np.abs(b)

c = a*mask + b*~mask

How do I do this in tensorflow while still being able to calculate gradients?

I know I can do this:

mask = tf.abs(a) > tf.abs(b)

but then:

c= tf.cast(mask, tf.float32) * a + tf.cast(~mask, tf.float32) * b

doesn't pass gradients as the cast operation doesn't pass gradients.

Upvotes: 0

Views: 575

Answers (2)

cs95
cs95

Reputation: 402852

Use np.maximum:

np.maximum(a, b, where=abs(a) > abs(b))

Upvotes: 2

Abhishek Mishra
Abhishek Mishra

Reputation: 2004

@coldspped answers work fine. But if you want a general purpose method that can extract element based on a random mask, you can use tf.where api. For your problem, the answer is as following:

mask = tf.abs(a) - tf.abs(b)
c = tf.where(mask, a, b)

Upvotes: 1

Related Questions