Passing in new java source code snippets to get predictions

[jetbrains-research]

Passing in new java source code snippets to get predictions

[SpirinEgor]

It seems that there are correct steps. Could you provide more information:

• The mined paths from the code snippet
• The input tensors shapes (they should be something like [5; n_paths])

Also, the model was trained with masked method names, so you should replace them too in your example. As far as I remember, mask token in <MN>. But you can check it. It should be in the top of the label_to_id counter. The correct code snippet is

def <MN>(n):
    if n == 0:
        return 1
    else:
        return n * <MN>(n-1)

Output shape is [7; 1; vocab size], so correct decoding is:

# [7]
predictions = output.squeeze(1).argmax(-1)
labels = [id_to_label(i.item()) for i in predictions]

[SpirinEgor]

Hi!

First of all, you need to extract paths from the code snippet. This should be done by the tool you used for model training.

After that, you should write a little wrapper to pass your datapoint into the model. Look into an example from readme for correct model loading. Pass your example into the model according to forward method signature.

[Thirunayan22]

Hi @SpirinEgor, thanks for responding !
To train the model I downloaded the dataset and ran the train script provided in the readme. I am a bit unclear, what do you mean by the "tool" used for model training?

[SpirinEgor]

What dataset did you download?

The tool I was speaking about is a tool for extracting paths from AST of code. There are some different ways to do it:

1. JavaExtractor from original code2seq repo by Alon
2. astminer
3. PSIMiner

All these tools use different parsers therefore they build different ASTs and extract different paths.

[hehehwang]

hello, i'm doing same thing with my own python ast miner
can you confirm whether i'm doing correctly?
https://gist.github.com/hehehwang/d058c6fca986a5b479afe10245f63a3e
(i think something went wrong since it's keep printing out and token..)

thank you!

[hehehwang]

seems that i've made a silly mistake on decoding.. thank you for your help!
now it works just fine and clear (without mask token though)

thanks again!

[SpirinEgor]

Awesome!
If you have other questions, feel free to open an issue!