Tensorflow NN: Custom loss and accuracy work, MSE returns NaN

Question

I've come into a bit of a confusing scenario. I'm training a Tensorflow/Keras neural network with a custom loss function and two metrics, MSE and accuracy. When evaluated on the training data, the loss and accuracy return numbers but the MSE is NaN. I was wondering if there was a possible explanation for this. Thanks.

my suggestions: 1- change your optimizer 2- check for nans in your input 3- normalize your input 4-in your target check if labels are correct (the range and etc.) — SoheilStar, May 09 '21 at 05:26

Dammio · Answer 1 · 2021-05-09T06:00:08.330

0

Check your input again. The input may have some "Not a Number" values. Refer this: How can I check for NaN values?

Or a quick test:

for value in values:
    print(f"{repr(value):<8} : {is_nan(value)}")

edited May 09 '21 at 06:00

answered May 09 '21 at 05:09

Dammio

911
1
7
15

Interesting - I'll try doing that. Thanks! – Andre May 09 '21 at 05:11
@Dammio How it is possible? MSE means the sum of the error^2 divided by the number of samples. the number of samples is always >0. – SoheilStar May 09 '21 at 05:20
I modified my answer a bit, there may have some Not a Number exists. – Dammio May 09 '21 at 05:52

Tensorflow NN: Custom loss and accuracy work, MSE returns NaN

1 Answers1