Shift in loss? about codebert HOT 5 CLOSED

Because the loss of generation is P(y_n|y_0,...y_n-1), you need to use vector of the last token to predict the next token. Therefore, you need to shift the label.

from codebert.

Thanks, @guody5 . Could you also explain why there is min(len(dfg),0) instead of a max?

from codebert.

Thanks. The code should be code_tokens=code_tokens[:args.max_source_length-3]. Please ignore this.

from codebert.

thanks for the quick reply. to clarify, if the code is too long, we only keep the code and ignore the dataflow, right?

from codebert.

Yes.

from codebert.