CodeTrans model for code documentation generation javascript

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

Intended uses & limitations

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

How to use

Here is how to use this model to generate javascript 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_javascript_multitask"),
    tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_large_code_documentation_generation_javascript_multitask", skip_special_tokens=True),
    device=0
)

tokenized_code = "function isStandardBrowserEnv ( ) { if ( typeof navigator !== 'undefined' && ( navigator . product === 'ReactNative' || navigator . product === 'NativeScript' || navigator . product === 'NS' ) ) { return false ; } return ( typeof window !== 'undefined' && typeof document !== 'undefined' ) ; }"
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 120,000 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.

Evaluation results

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

Test results :

Language / Model	Python	Java	Go	Php	Ruby	JavaScript
CodeTrans-ST-Small	17.31	16.65	16.89	23.05	9.19	13.7
CodeTrans-ST-Base	16.86	17.17	17.16	22.98	8.23	13.17
CodeTrans-TF-Small	19.93	19.48	18.88	25.35	13.15	17.23
CodeTrans-TF-Base	20.26	20.19	19.50	25.84	14.07	18.25
CodeTrans-TF-Large	20.35	20.06	19.54	26.18	14.94	18.98
CodeTrans-MT-Small	19.64	19.00	19.15	24.68	14.91	15.26
CodeTrans-MT-Base	20.39	21.22	19.43	26.23	15.26	16.11
CodeTrans-MT-Large	20.18	21.87	19.38	26.08	15.00	16.23
CodeTrans-MT-TF-Small	19.77	20.04	19.36	25.55	13.70	17.24
CodeTrans-MT-TF-Base	19.77	21.12	18.86	25.79	14.24	18.62
CodeTrans-MT-TF-Large	18.94	21.42	18.77	26.20	14.19	18.83
State of the art	19.06	17.65	18.07	25.16	12.16	14.90

Created by Ahmed Elnaggar | LinkedIn and Wei Ding | LinkedIn