How to manage NaNs in responses training a convolutional neural network?

Hello,

I am training a UNET for regression. I am facing the issue of managing the NaNs in the responses (reference) data. My input data is a 4-D matrix (48x48x9xN), while the reference is always a 4-D matrix (48x48xx1xN). A number of the reference images (i.e. some of the N 48x48 images) are partially filled, it means that some values are NaN.

When I start the trining process I get the following error message:

"Invalid training data. Responses must not contain NaNs."

Is there a way to manage NaNs? It is important to highlight that the input pixels corresponding to reference NaN pixel, are not NaN but have reliable values.

Thanks.

Leo Pio

Answer 1

[KSSV]

You can fill NaN's using either fillmissing, interp2. Also have a look on the fileexchange: https://in.mathworks.com/matlabcentral/fileexchange/15590-fillnans

[Leo Pio D'Adderio]

I cannot replace NaN's with anything. Consider a 48x48 image, supposing that the upper triangle of reference data is all NaN, while the lower triangle has reliable values. I would that the network trained only on the lower triangle, but the image size must be always 48x48.

[KSSV]

Replace NaN's with zeros then.

A(isnan(A))=0 ;