wei commited on
Commit
6fdfdb5
1 Parent(s): f4c39f5

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +75 -0
README.md CHANGED
@@ -5,3 +5,78 @@ widget:
5
  - text: "you are given an array of numbers a and a number b , compute the difference of elements in a and b"
6
 
7
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  - text: "you are given an array of numbers a and a number b , compute the difference of elements in a and b"
6
 
7
  ---
8
+
9
+
10
+ # CodeTrans model for program synthesis
11
+ Pretrained model on programming language lisp inspired DSL using the t5 base model architecture. It was first released in
12
+ [this repository](https://github.com/agemagician/CodeTrans).
13
+
14
+
15
+ ## Model description
16
+
17
+ 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. It is then fine-tuned on the program synthesis task for the lisp inspired DSL code.
18
+
19
+ ## Intended uses & limitations
20
+
21
+ The model could be used to generate lisp inspired DSL code given the human language description tasks.
22
+
23
+ ### How to use
24
+
25
+ Here is how to use this model to generate lisp inspired DSL code using Transformers SummarizationPipeline:
26
+
27
+ ```python
28
+ from transformers import AutoTokenizer, AutoModelWithLMHead, SummarizationPipeline
29
+
30
+ pipeline = SummarizationPipeline(
31
+ model=AutoModelWithLMHead.from_pretrained("SEBIS/code_trans_t5_base_program_synthese_multitask_finetune"),
32
+ tokenizer=AutoTokenizer.from_pretrained("SEBIS/code_trans_t5_base_program_synthese_multitask_finetune", skip_special_tokens=True),
33
+ device=0
34
+ )
35
+
36
+ tokenized_code = "you are given an array of numbers a and a number b , compute the difference of elements in a and b"
37
+ pipeline([tokenized_code])
38
+ ```
39
+ Run this example in [colab notebook](https://github.com/agemagician/CodeTrans/blob/main/prediction/multitask/fine-tuning/program%20synthesis/base_model.ipynb).
40
+ ## Training data
41
+
42
+ 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)
43
+
44
+
45
+ ## Training procedure
46
+
47
+ ### Multi-task Pretraining
48
+
49
+ The model was trained on a single TPU Pod V3-8 for 500,000 steps in total, using sequence length 512 (batch size 4096).
50
+ It has a total of approximately 220M parameters and was trained using the encoder-decoder architecture.
51
+ The optimizer used is AdaFactor with inverse square root learning rate schedule for pre-training.
52
+
53
+ ### Fine-tuning
54
+
55
+ This model was then fine-tuned on a single TPU Pod V2-8 for 30,000 steps in total, using sequence length 512 (batch size 256), using only the dataset only containing lisp inspired DSL data.
56
+
57
+
58
+ ## Evaluation results
59
+
60
+ For the code documentation tasks, different models achieves the following results on different programming languages (in BLEU score):
61
+
62
+ Test results :
63
+
64
+ | Language / Model | LISP |
65
+ | -------------------- | :------------: |
66
+ | CodeTrans-ST-Small | 89.43 |
67
+ | CodeTrans-ST-Base | 89.65 |
68
+ | CodeTrans-TF-Small | 90.30 |
69
+ | CodeTrans-TF-Base | 90.24 |
70
+ | CodeTrans-TF-Large | 90.21 |
71
+ | CodeTrans-MT-Small | 82.88 |
72
+ | CodeTrans-MT-Base | 86.99 |
73
+ | CodeTrans-MT-Large | 90.27 |
74
+ | CodeTrans-MT-TF-Small | **90.31** |
75
+ | CodeTrans-MT-TF-Base | 90.30 |
76
+ | CodeTrans-MT-TF-Large | 90.17 |
77
+ | State of the art | 85.80 |
78
+
79
+
80
+
81
+ > 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/)
82
+