Uri Alon

The main idea is to assign a random vector to every kind of symbol, whether that symbol is a path or a token. Then, during training, these vectors are modified...

This is where the random vectors are initialized: https://github.com/tech-srl/code2vec/blob/master/tensorflow_model.py#L206-L220 The reader converts string inputs into integer indices, and these integer indices allow looking up the specific random vector: https://github.com/tech-srl/code2vec/blob/master/path_context_reader.py#L205-L207 The...

> So the network doesn't really have granularity at the AST level. Paths are treated just as whole strings Thus is correct, and was addressed in a follow-up paper [Code2seq](https://code2seq.org/)...

Hi @wenhaozheng-nju , Thank you for your interest in our work! We haven't tried PPI just because GNNs' results in some papers reach close to 100%, so there might not...

@Russell91 - can you please take a look?

No, when you use model=0, it doesn't matter because the "word" vectors and the "context" vectors are concatenated, and therefore they lose their "role" (it doesn't matter which of them...

and if so - would you rather me fixing it in the saving part of glove.c, or in the creation of the CREC struct in cooccur.c?

Well, if the users rely on that, they are relying on something that contradicts the documentation! Anyway - we can fix it and at the same time keep the old...

Do you have something in mind about "evidence that I can assemble in terms of functionality"? I mean - both the current version and my proposed fix *will work*, the...

@Russell91 what do you think? can I send a fix?