Method prediction

[crimsonfan]

First of all, I would like to thanks all of your works. I'm a college student and I'm studying, doing research about this project. I would like to ask that: when code2vec predicts a snippet of code, it also gives us the vector of that snipped, after the method prediction (99% [solve] for example), so if I want to calculate the distance between the snippet's vector and the method prediction's vectors ([solve] vector), can I find those prediction's vectors from this project and where can I find these hardcodes location?

Many thanks.

[crimsonfan]

Dear author, I read the solved issues and found they are located in models/java14_model/target.txt, am I right? P/s: If in a class has many function, can the code2vec shows the predict method name of each function along with its line number (in case functions have the same name)? Thanks.

[urialon]

Hi @crimsonfan , Thank you for your interest in code2vec!

Yes, you can take the code vector and compute its distance with the target embedding. Did you find the target.txt file? See also https://github.com/tech-srl/code2vec#exporting-the-trained-token-vectors-and-target-vectors

Alternatively, you can perform this computation (code-target distance) in runtime, and run a --test run on the test set that will output the distances for the entire test set.

[crimsonfan]

Dear @urialon,

Thanks, I've almost done that part. As i read, the code2vec analyzes functions in a class one by one and transfer them to vectors (4 functions to 4 vectors). Is there any way for me to find the starting line of each function for the case that all of them have the same names?

Thanks, Tin.

[urialon]

Currently, this is not supported. Maybe the easiest way would be to append the line number to each method name, during preprocessing (during the run of the JavaExtractor). For example, toString can become toString_123.

Best, Uri

[crimsonfan]

Dear @urialon , Thanks for all. Tin.