CODEWITHSUNDEEP
tensorflow
le4m
le4m

Reputation: 578

tensorflow function that projects max value to 1 and others -1 without using zeros

How can I make x from y where

x = tf.constant([[1,5,3], [100,20,3]])
y = ([[-1,1,-1], [1,-1,-1]])

So it basically projects the max value to 1 and makes other elements to -1. One important constraint is we should not use zero. (Because if I use zero, the gradient does not flow on that node.) Using tf.argmax we can get the max indices but don't really know how to make y from it.

Could you please help?

For pedagogical purpose I set x as constant, but in the actual problem I'am solving, x is actually a placeholder that gets inputs of a network.

Upvotes: 1

Views: 2058

Answers (2)

Yogendra Miraje
Yogendra Miraje

Reputation: 101

Following is much cleaner solution that works for me:

y = tf.one_hot(tf.argmax(x, axis=1), x.shape[1], on_value=1, off_value=-1)

Basically, tf.argmax gives you the index of the max value along the dimension 1 and the one_hot method creates the desired tesnor.

Upvotes: 0

akuiper
akuiper

Reputation: 214927

You can use tf.reduce_max to calculate the max per row, compare with original tensor, and use tf.where to set values conditionally:

x = tf.constant([[1,5,3], [100,20,3]])

sess = tf.InteractiveSession()
sess.run(
    tf.where(
        tf.equal(tf.reduce_max(x, axis=1, keep_dims=True), x), 
        tf.constant(1, shape=x.shape), 
        tf.constant(-1, shape=x.shape)
    )
)

array([[-1,  1, -1],
       [ 1, -1, -1]], dtype=int32)

Upvotes: 3

Related Questions