Keras fitting ignoring nan values

问题

I am training a neural network to do regression, (1 input and 1 output). Let's x and y be the usual input and output dataset, respectively.

My problem is that the y dataset (not the x) have some values set to nan, so the fitting goes to nan. I wonder if there is an option to ignore the nan values in the fitting, in a similar way to the numpy functions np.nanmean to calculate the mean ignoring nans and so on.

If that option does not exist I suppose I would have to find the nan values and erase them manually, and at the same time erase the values in x corresponding to the nan position in y.

x       y
2       4
3       2
4      np.nan
5       7
6      np.nan
7      np.nan

In this simple example the nan values in the y column should be removed and at the same time the corresponding values in the x column (4, 6, 7).

Thank you.

EDIT: Ok, I have a problem filtering the nans, I do:

for index, x in np.ndenumerate(a):
    if x == np.nan:
        print index, x

and it doesn't print anything and I am sure there are nan values...

EDIT (SELF ANSWER): Ok, I have found a way to localize the nans:

for index, x in np.ndenumerate(a):
      if x != x:
           print index, x

回答1:

As said in the comments, simply remove the nan as a preprocessing step:

import numpy as np

x = range(2,8)
y = [4,2,np.nan,7,np.nan,np.nan]

for a,b in zip(x,y):
    if str(b) == 'nan':
        x.remove(a)
        y.remove(b)

print x,y

produces [2, 3, 5] [4, 2, 7].

If you're using some tool to preprocess the data which gives you the np.nan, check whether the API allows you to disable this behavior and take a minute to think whether this is really the behavior you want (or if you e.g. want to map this to constants because you find your input to be valuable even though they have no labels).

来源：https://stackoverflow.com/questions/37747886/keras-fitting-ignoring-nan-values