wei's picture
Update README.md
31ed1af
metadata
tags:
  - summarization
widget:
  - text: >-
      protected String renderUri ( URI uri ) { return uri . toASCIIString ( ) ;
      }

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 multi-task training on 13 supervised tasks in the software development domain and 7 unsupervised datasets. It is then fine-tuned on the code comment 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_comment_generation_java_multitask_finetune"),
    tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_large_code_comment_generation_java_multitask_finetune", skip_special_tokens=True),
    device=0
)

tokenized_code = "protected String renderUri ( URI uri ) { return uri . toASCIIString ( ) ; }"
pipeline([tokenized_code])

Run this example in colab notebook.

Training data

The supervised training tasks datasets can be downloaded on Link

Training procedure

Multi-task Pretraining

The model was trained on a single TPU Pod V3-8 for 260,000 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 V3-8 for 25,000 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 :

Language / Model Java
CodeTrans-ST-Small 37.98
CodeTrans-ST-Base 38.07
CodeTrans-TF-Small 38.56
CodeTrans-TF-Base 39.06
CodeTrans-TF-Large 39.50
CodeTrans-MT-Small 20.15
CodeTrans-MT-Base 27.44
CodeTrans-MT-Large 34.69
CodeTrans-MT-TF-Small 38.37
CodeTrans-MT-TF-Base 38.90
CodeTrans-MT-TF-Large 39.25
State of the art 38.17

Created by Ahmed Elnaggar | LinkedIn and Wei Ding | LinkedIn