CodeTrans model for code documentation generation java

Pretrained model on programming language java using the t5 large model architecture. It was first released in this repository. This model is trained on tokenized java code functions: it works best with tokenized java functions.

Model description

This CodeTrans model is based on the t5-large model. It has its own SentencePiece vocabulary model. It used transfer-learning pre-training on 7 unsupervised datasets in the software development domain. It is then fine-tuned on the code documentation generation task for the java function/method.

Intended uses & limitations

The model could be used to generate the description for the java function or be fine-tuned on other java code tasks. It can be used on unparsed and untokenized java code. However, if the java code is tokenized, the performance should be better.

How to use

Here is how to use this model to generate java function documentation using Transformers SummarizationPipeline:

from transformers import AutoTokenizer, AutoModelWithLMHead, SummarizationPipeline

pipeline = SummarizationPipeline(
    model=AutoModelWithLMHead.from_pretrained("SEBIS/code_trans_t5_large_code_documentation_generation_java_transfer_learning_finetune"),
    tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_large_code_documentation_generation_java_transfer_learning_finetune", skip_special_tokens=True),
    device=0
)

tokenized_code = "public static < T , U > Function < T , U > castFunction  ( Class < U > target ) { return new CastToClass < T , U > ( target ) ; }"
pipeline([tokenized_code])

Run this example in colab notebook.

Training data

The supervised training tasks datasets can be downloaded on Link

Training procedure

Transfer-learning Pretraining

The model was trained on a single TPU Pod V3-8 for 240,000 steps in total, using sequence length 512 (batch size 4096). It has a total of approximately 220M parameters and was trained using the encoder-decoder architecture. The optimizer used is AdaFactor with inverse square root learning rate schedule for pre-training.

Fine-tuning

This model was then fine-tuned on a single TPU Pod V2-8 for 500 steps in total, using sequence length 512 (batch size 256), using only the dataset only containing java code.

Evaluation results

For the code documentation tasks, different models achieves the following results on different programming languages (in BLEU score):

Test results :

Created by Ahmed Elnaggar | LinkedIn and Wei Ding | LinkedIn