How does matconvnet deal with 'NaN'?

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB

Other

1.4k stars 752 forks source link

How does matconvnet deal with 'NaN'? #1190

Open leeeeeeo opened 6 years ago

leeeeeeo commented 6 years ago

My training data contains 'Nan', how does matconvnet deal with 'NaN'? For example, I assume, In the conv layer, 'Nan' is still 'NaN' after forward propogation. What's the gradient when backward propogation?