code-code CodeCompletion input about codexglue

[celbree]

I would answer your second question first.

another question, is the CodeGPT pretrained and finetuned with the same code corpus, but different format? (raw code for pretrain, token format for finetune?)

The pre-trained dataset and fine-tuned dataset are different. CodeGPT is pre-trained on Python and Java corpus from CodeSearchNet in raw code format, and fine-tuned on PY150 and Github Java Corpus in token format. We'll update more details about CodeGPT to our repo.

For the first question.

But the performance of finetuned model != pretrain model, right?
If I want to use finetuned model in the production environment, I have to preprocess the input code before sending them into model, and postprocess the output before presenting in IDE...

Yes. The fine-tuned model is more likely to predict code sequences in token level format. And if you would like to use it in real code completion scenario, you could use our released pre-trained CodeGPT model which is suitable for predicting raw codes. Alternately, you can try to fine-tuning CodeGPT on more data corpus (e.g. you could download PY150 and Github Java Corpus to obtain the raw code of each code file without running our preprocessing script). According to the time-cost analysis, it won't take too much time or computing resources.

[celbree]

Yes, indeed. We do the preprocessing to ignore all the indentation and line breaks. CodeGPT is a pre-train model for CodeXGLUE participants to play with. It's train on a not large corpus and it can be viewed as a baseline for CodeXGLUE benchmark.

[celbree]

Thanks for your suggestion.
As in most of related works about code completion, many researchers focus on token-level code completion and they tokenize raw codes into token sequences. We keep this setting in CodeXGLUE. To ensure consistency, we also use this input format for line-level code completion.
If a model works on the initial raw code, it is expected to perform well on token format after fine-tuned. In fact, our CodeGPT is pre-trained on the initial format of raw code and fine-tuned on token format.

[carter54]

I see...thx for the reply. I can understand such token level format is easier for the evaluations.
But the performance of finetuned model != pretrain model, right?
If I want to use finetuned model in the production environment, I have to preprocess the input code before sending them into model, and postprocess the output before presenting in IDE...

[carter54]

another question, is the CodeGPT pretrained and finetuned with the same code corpus, but different format? (raw code for pretrain, token format for finetune?)
Thx in advance.

[carter54]

@celbree Thx for these details~
I have tried the pre-trained CodeGPT model and found that all indentation and line break are missing. Do I made any mistake when deploying the model?
Or did you preprocess the raw code by removing indentation and line break before pratraining the model? If so, any reason to do this preprocess?

[carter54]

@celbree Thanks~
to my point of view, it is better to keep indentation and line breaks in the input, as indentation is import to the semantics of code.
a simple example:

a = 1
b = 2
if a > b:
    print ('I don't agree')
    print('Neither do I')

a = 1
b = 2
if a > b:
    print ('I don't agree')
print('Neither do I')

are different

[celbree]

Totally agree. Thanks for your suggestion.