CodeTrans model for source code summarization csharp
Pretrained model on programming language csharp using the t5 base model architecture. It was first released in this repository. This model is trained on tokenized csharp code functions: it works best with tokenized csharp functions.
Model description
This CodeTrans model is based on the t5-base
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 csharp function or be fine-tuned on other csharp code tasks. It can be used on unparsed and untokenized csharp code. However, if the csharp code is tokenized, the performance should be better.
How to use
Here is how to use this model to generate csharp function documentation using Transformers SummarizationPipeline:
from transformers import AutoTokenizer, AutoModelWithLMHead, SummarizationPipeline
pipeline = SummarizationPipeline(
model=AutoModelWithLMHead.from_pretrained("SEBIS/code_trans_t5_base_source_code_summarization_csharp_multitask"),
tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_base_source_code_summarization_csharp_multitask", skip_special_tokens=True),
device=0
)
tokenized_code = "public static DateTime ParseUnixDateTime ( double unixTime ) { var dt = new DateTime ( CODE_INTEGER , CODE_INTEGER , CODE_INTEGER , CODE_INTEGER , CODE_INTEGER , CODE_INTEGER , CODE_INTEGER , System . DateTimeKind . Utc ) ; dt = dt . AddSeconds ( unixTimeStamp ) . ToLocalTime ( ) ; return dt ; }"
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 160,000 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 source code summarization tasks, different models achieves the following results on different programming languages (in BLEU score):
Test results :
Language / Model | Python | SQL | C# |
---|---|---|---|
CodeTrans-ST-Small | 8.45 | 17.55 | 19.74 |
CodeTrans-ST-Base | 9.12 | 15.00 | 18.65 |
CodeTrans-TF-Small | 10.06 | 17.71 | 20.40 |
CodeTrans-TF-Base | 10.94 | 17.66 | 21.12 |
CodeTrans-TF-Large | 12.41 | 18.40 | 21.43 |
CodeTrans-MT-Small | 13.11 | 19.15 | 22.39 |
CodeTrans-MT-Base | 13.37 | 19.24 | 23.20 |
CodeTrans-MT-Large | 13.24 | 19.40 | 23.57 |
CodeTrans-MT-TF-Small | 12.10 | 18.25 | 22.03 |
CodeTrans-MT-TF-Base | 10.64 | 16.91 | 21.40 |
CodeTrans-MT-TF-Large | 12.14 | 19.98 | 21.10 |
CODE-NN | -- | 18.40 | 20.50 |
Created by Ahmed Elnaggar | LinkedIn and Wei Ding | LinkedIn
- Downloads last month
- 3