1 ---
2 tags:
3 - summarization
4 widget:
5 - text: "select time ( col0 ) from tab0"
6
7 ---
8
9
10
11 # CodeTrans model for source code summarization sql
12 Pretrained model on programming language sql using the t5 large model architecture. It was first released in
13 [this repository](https://github.com/agemagician/CodeTrans). This model is trained on tokenized sql code functions: it works best with tokenized sql functions.
14
15
16 ## Model description
17
18 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.
19
20
21 ## Intended uses & limitations
22
23 The model could be used to generate the description for the sql function or be fine-tuned on other sql code tasks. It can be used on unparsed and untokenized sql code. However, if the sql code is tokenized, the performance should be better.
24
25 ### How to use
26
27 Here is how to use this model to generate sql function documentation using Transformers SummarizationPipeline:
28
29 ```python
30 from transformers import AutoTokenizer, AutoModelWithLMHead, SummarizationPipeline
31
32 pipeline = SummarizationPipeline(
33 model=AutoModelWithLMHead.from_pretrained("SEBIS/code_trans_t5_large_source_code_summarization_sql_multitask"),
34 tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_large_source_code_summarization_sql_multitask", skip_special_tokens=True),
35 device=0
36 )
37
38 tokenized_code = "select time ( col0 ) from tab0"
39 pipeline([tokenized_code])
40 ```
41 Run this example in [colab notebook](https://github.com/agemagician/CodeTrans/blob/main/prediction/multitask/pre-training/source%20code%20summarization/sql/large_model.ipynb).
42 ## Training data
43
44 The supervised training tasks datasets can be downloaded on [Link](https://www.dropbox.com/sh/488bq2of10r4wvw/AACs5CGIQuwtsD7j_Ls_JAORa/finetuning_dataset?dl=0&subfolder_nav_tracking=1)
45
46
47 ## Training procedure
48
49 ### Multi-task Pretraining
50
51 The model was trained on a single TPU Pod V3-8 for 120,000 steps in total, using sequence length 512 (batch size 4096).
52 It has a total of approximately 220M parameters and was trained using the encoder-decoder architecture.
53 The optimizer used is AdaFactor with inverse square root learning rate schedule for pre-training.
54
55
56 ## Evaluation results
57
58 For the source code summarization tasks, different models achieves the following results on different programming languages (in BLEU score):
59
60 Test results :
61
62 | Language / Model | Python | SQL | C# |
63 | -------------------- | :------------: | :------------: | :------------: |
64 | CodeTrans-ST-Small | 8.45 | 17.55 | 19.74 |
65 | CodeTrans-ST-Base | 9.12 | 15.00 | 18.65 |
66 | CodeTrans-TF-Small | 10.06 | 17.71 | 20.40 |
67 | CodeTrans-TF-Base | 10.94 | 17.66 | 21.12 |
68 | CodeTrans-TF-Large | 12.41 | 18.40 | 21.43 |
69 | CodeTrans-MT-Small | 13.11 | 19.15 | 22.39 |
70 | CodeTrans-MT-Base | **13.37** | 19.24 | 23.20 |
71 | CodeTrans-MT-Large | 13.24 | 19.40 | **23.57** |
72 | CodeTrans-MT-TF-Small | 12.10 | 18.25 | 22.03 |
73 | CodeTrans-MT-TF-Base | 10.64 | 16.91 | 21.40 |
74 | CodeTrans-MT-TF-Large | 12.14 | **19.98** | 21.10 |
75 | CODE-NN | -- | 18.40 | 20.50 |
76
77
78 > Created by [Ahmed Elnaggar](https://twitter.com/Elnaggar_AI) | [LinkedIn](https://www.linkedin.com/in/prof-ahmed-elnaggar/) and Wei Ding | [LinkedIn](https://www.linkedin.com/in/wei-ding-92561270/)
79
80
81