tf.keras.layers.PReLU outputs NaN on positive input about tensorflow HOT 3 OPEN

Hi @maybeLee, I tried to reproduce the issue, from the given constraint/dummy code, and i am able to get the error output as

and after spending some time over your issue, got to know, this issue may be arised due to mismatch in the package version.

from tensorflow.

HI @AkashKumarSingh11032001
My previous code is working for tensorflow 2.15.0 and it seems that this API does not support the weights parameter in 2.16.0.
Here is the adjusted code for 2.16.0:

import numpy as np
import tensorflow as tf
input = tf.constant(1.1234, dtype='float32')
out = tf.keras.layers.PReLU(alpha_initializer=tf.keras.initializers.Constant(np.inf))(input)
print(out)

from tensorflow.

Hi @maybeLee ,

I tried to run your code on colab using tf 2.15, 2.16 and tf-nightly versions faced same issue. Please find the gist here for reference. We need to work on it, In few days i will provide a solution for it.

Thank you!

from tensorflow.