Replace nan values in tensorflow tensor

Question

I\'m working on a convolutional neural network in tensorflow and I have a problem. The problem is the input image I read through tfrecords contains a certain number of nan v

Answer 1

Clip by value made NaN infinity and where was overkill for one variable. I used this to convert a single value to 0 if it's NaN:

value_not_nan = tf.dtypes.cast(tf.math.logical_not(tf.math.is_nan(value)), dtype=tf.float32)
tf.math.multiply_no_nan(value, value_not_nan)

Answer 2

If someone is looking for the solution in Tensorflow 2.0, the adapted code of Allen Lavoie is :

import tensorflow as tf
with tf.compat.v1.Session():
    has_nans = tf.constant([float('NaN'), 1.])
    print(tf.where(tf.math.is_nan(has_nans), tf.zeros_like(has_nans), has_nans).eval())

Answer 3

A combination of tf.where and tf.is_nan should work:

import tensorflow as tf
with tf.Session():
    has_nans = tf.constant([float('NaN'), 1.])
    print(tf.where(tf.is_nan(has_nans), tf.zeros_like(has_nans), has_nans).eval())

Prints (using TensorFlow 0.12.1):

[ 0.  1.]

Answer 4

A much easier approach, compatible with TF2.0, is to just use tf.clip_by_value, which mirrors np.clip and removes NaNs (see here):

no_nans = tf.clip_by_value(has_nans, -1e12, 1e12)

Some caveats: 1) this also removes infs 2) Depending on your application you may need to set the clip value to a high value to avoid losing info.