Clip to maximum / minimum finite int/float/double?

Question

I'd like to clip some Tensorflow variables with tf.clip_by_value to the most positive / negative finite value possible. I'd like to reach a similar behavior like numpy.nan_to_num()

How do I get these values, given some data type? Does this clipping affect the optimizer?

Answer 1

All TensorFlow datatypes have the min and max properties, that return the maximum and minimum values the type can hold. E.g.,

import tensorflow as tf

print( tf.float32.max )

Outputs:

3.4028235e+38