training stops due to NaN loss value

The training process stops due NaN loss... How to avoid this to complete the training ..and what is the possible issue that causses..

[Chunru]

Likely you have some nan data in your training samples. You can try

rmmissing(A, dim)

[Shoaib Ali]

the data do not have any "Nan" value

[Chunru]

Then what is the loss function?

[Shoaib Ali]

Weighted cross entropy loss

[Chunru]

try "dbstop error" and then run the program. Check if the network output is 0. There might be a problem if network output is 0 since entropy loss has term of T*log(Y) where T is target and Y is the network output.

Answer 1

[Shoaib Ali]

Can you explain, where to use this inside the loss funtion or before the training command??

[Chunru]

"dbstop error" can be used in command line before you train the network. Then it should stop when error occurs and then you check out what is wrong at which part of program.

[Shoaib Ali]

ok Thanks