How can I implement a layer which can keep several decimal?

Question

I want to achieve, in the output layer to retain two decimal functions. Because I want to use it between two convolution layers, so I want to use it to achieve this.
But because the two decimals it keeps often overflow, I don't know how to solve it?

import tensorflow as tf

input = tf.Variable([3.5115155, 3.365, 3.38115155, 3.81151536, 3.38115159, 3.38115158, 3.398115155], dtype=tf.float32)

@tf.custom_gradient
def round_test(x):
   def grad(dy):
     return 1.0*dy
   return tf.math.round(x * 100)/100, grad

output_clip = round_test(input)

grad_clip = tf.gradients(output_clip, input)

with tf.Session() as sess:
   sess.run(tf.global_variables_initializer())
   print("input:", sess.run(input))
   print("output_clipping:", sess.run(output_clip))
   print("with clipping:", sess.run(grad_clip)[0])

This is a mistake.
input: [3.5115156 3.365 3.3811514] output_clipping: [3.51 3.36 3.3799999]

I expect the output of roud_test(3.3811514) to be 3.38 , but the actual output is 3.3799999
I just want to keep two decimal places.

Answer 1

Try tf.py_func :

import numpy as np #add

return tf.py_func(lambda a:np.round(a,2),[x],tf.float32),grad

The results:

input: [3.5115156 3.365     3.3811514]
output_clipping: [3.51 3.36 3.38]
with clipping: [1. 1. 1.]