Can i build function classifier using code-transformer

[ZJUFLS]

Hi, I would like to build a method/function classifier using code-transformer, the labels are going to be 0,1,2,..9,

Could i get some suggestions on how to build it?

Any guidelines would be hugely appreciated! :D

[vthost]

I am currently trying the same and add the classes as label instead of the function name, setting separate_label_vocabulary to True and num_subtokens_output to 1. But maybe the authors have a better solution, or additional requirements :)

[danielzuegner]

Yes, that should be a fairly direct way to achieve what you want. Closing this, feel free to reopen in case you have follow-up questions.

[tobias-kirschstein]

Cool idea to use the Code Transformer for method classification. A few remarks:

• Preprocessing: When using the separate label vocabulary approach you will have to adapt how the output vocabulary is populated (by default it is filled with all word tokens that appear in method names). This is handled in the preprocess-1.py (Counting of tokens) and in preprocess-2.py (Collecting the most common tokens into a vocabulary). If you want to avoid running the preprocessing pipeline to generate your own dataset, it should also be possible to add your classification output vocabulary to the already existing vocabularies (the vocabularies.p file is just a pickled 3- or 4-tuple of Vocabulary objects).
• Model Input: Obviously, you also need to change how the preprocessed methods are fed into the model. The standard dataset view we used for the Code Summarization task masks the method name, which you probably don't want.
• Model: Furthermore, by default, the training script uses an Code Transformer encoder together with a regular Transformer Decoder. For Method Classification having a simple linear layer in the end instead of the Decoder might be better.

[vthost]

Thank you! I considered the first two, but it's good to have a confirmation.