CodeTrans model for api recommendation generation

Pretrained model for api recommendation generation using the t5 large model architecture. It was first released in this repository.

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 api recommendation generation task for the java apis.

Intended uses & limitations

The model could be used to generate api usage for the java programming tasks.

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_api_generation_transfer_learning_finetune"),
    tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_large_api_generation_transfer_learning_finetune", skip_special_tokens=True),
    device=0
)

tokenized_code = "parse the uses licence node of this package , if any , and returns the license definition if theres"
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 V3-8 for 180,000 steps in total, using sequence length 512 (batch size 256), using only the dataset only containing api recommendation generation data.

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	68.71
CodeTrans-ST-Base	70.45
CodeTrans-TF-Small	68.90
CodeTrans-TF-Base	72.11
CodeTrans-TF-Large	73.26
CodeTrans-MT-Small	58.43
CodeTrans-MT-Base	67.97
CodeTrans-MT-Large	72.29
CodeTrans-MT-TF-Small	69.29
CodeTrans-MT-TF-Base	72.89
CodeTrans-MT-TF-Large	73.39
State of the art	54.42

Created by Ahmed Elnaggar | LinkedIn and Wei Ding | LinkedIn