yuewang-sf commited on
Commit
94f88f9
1 Parent(s): ad0691e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +110 -0
README.md CHANGED
@@ -1,3 +1,113 @@
1
  ---
2
  license: bsd-3-clause
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: bsd-3-clause
3
  ---
4
+
5
+ # CodeT5+ 110M Embedding Models
6
+
7
+ ## Model description
8
+
9
+ [CodeT5+](https://github.com/salesforce/CodeT5/tree/main/CodeT5+) is a new family of open code large language models
10
+ with an encoder-decoder architecture that can flexibly operate in different modes (i.e. _encoder-only_, _decoder-only_,
11
+ and _encoder-decoder_) to support a wide range of code understanding and generation tasks.
12
+ It is introduced in the paper:
13
+
14
+ [CodeT5+: Open Code Large Language Models for Code Understanding and Generation](https://arxiv.org/pdf/2305.07922.pdf)
15
+ by [Yue Wang](https://yuewang-cuhk.github.io/)\*, [Hung Le](https://sites.google.com/view/henryle2018/home?pli=1)\*, [Akhilesh Deepak Gotmare](https://akhileshgotmare.github.io/), [Nghi D.Q. Bui](https://bdqnghi.github.io/), [Junnan Li](https://sites.google.com/site/junnanlics), [Steven C.H. Hoi](https://sites.google.com/view/stevenhoi/home) (*
16
+ indicates equal contribution).
17
+
18
+ Compared to the original CodeT5 family (base: `220M`, large: `770M`), CodeT5+ is pretrained with a diverse set of
19
+ pretraining tasks including _span denoising_, _causal language modeling_, _contrastive learning_, and _text-code
20
+ matching_ to learn rich representations from both unimodal code data and bimodal code-text data.
21
+ Additionally, it employs a simple yet effective _compute-efficient pretraining_ method to initialize the model
22
+ components with frozen off-the-shelf LLMs such as [CodeGen](https://github.com/salesforce/CodeGen) to efficiently scale
23
+ up the model (i.e. `2B`, `6B`, `16B`), and adopts a "shallow encoder and deep decoder" architecture.
24
+ Furthermore, it is instruction-tuned to align with natural language instructions (see our InstructCodeT5+ 16B)
25
+ following [Code Alpaca](https://github.com/sahil280114/codealpaca).
26
+
27
+ ## How to use
28
+
29
+ This checkpoint consists of an encoder of CodeT5+ 220M model (pretrained from 2 stages on both unimodal and bimodal) and a projection layer, which can be used to extract code
30
+ embeddings of 256 dimension. It can be easily loaded using the `AutoModel` functionality and employs the
31
+ same [CodeT5](https://github.com/salesforce/CodeT5) tokenizer.
32
+
33
+ ```python
34
+ from transformers import AutoModel, AutoTokenizer
35
+
36
+ checkpoint = "Salesforce/codet5p-110m-embedding"
37
+ device = "cuda" # for GPU usage or "cpu" for CPU usage
38
+
39
+ tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True)
40
+ model = AutoModel.from_pretrained(checkpoint, trust_remote_code=True).to(device)
41
+
42
+ inputs = tokenizer.encode("def print_hello_world():\tprint('Hello World!')", return_tensors="pt").to(device)
43
+ embedding = model(inputs)[0]
44
+ print(f'Dimension of the embedding: {embedding.size()[0]}, with norm={embedding.norm().item()}')
45
+ # Dimension of the embedding: 256, with norm=1.0
46
+ print(embedding)
47
+ # tensor([ 0.0185, 0.0229, -0.0315, -0.0307, -0.1421, -0.0575, -0.0275, 0.0501,
48
+ # 0.0203, 0.0337, -0.0067, -0.0075, -0.0222, -0.0107, -0.0250, -0.0657,
49
+ # 0.1571, -0.0994, -0.0370, 0.0164, -0.0948, 0.0490, -0.0352, 0.0907,
50
+ # -0.0198, 0.0130, -0.0921, 0.0209, 0.0651, 0.0319, 0.0299, -0.0173,
51
+ # -0.0693, -0.0798, -0.0066, -0.0417, 0.1076, 0.0597, -0.0316, 0.0940,
52
+ # -0.0313, 0.0993, 0.0931, -0.0427, 0.0256, 0.0297, -0.0561, -0.0155,
53
+ # -0.0496, -0.0697, -0.1011, 0.1178, 0.0283, -0.0571, -0.0635, -0.0222,
54
+ # 0.0710, -0.0617, 0.0423, -0.0057, 0.0620, -0.0262, 0.0441, 0.0425,
55
+ # -0.0413, -0.0245, 0.0043, 0.0185, 0.0060, -0.1727, -0.1152, 0.0655,
56
+ # -0.0235, -0.1465, -0.1359, 0.0022, 0.0177, -0.0176, -0.0361, -0.0750,
57
+ # -0.0464, -0.0846, -0.0088, 0.0136, -0.0221, 0.0591, 0.0876, -0.0903,
58
+ # 0.0271, -0.1165, -0.0169, -0.0566, 0.1173, -0.0801, 0.0430, 0.0236,
59
+ # 0.0060, -0.0778, -0.0570, 0.0102, -0.0172, -0.0051, -0.0891, -0.0620,
60
+ # -0.0536, 0.0190, -0.0039, -0.0189, -0.0267, -0.0389, -0.0208, 0.0076,
61
+ # -0.0676, 0.0630, -0.0962, 0.0418, -0.0172, -0.0229, -0.0452, 0.0401,
62
+ # 0.0270, 0.0677, -0.0111, -0.0089, 0.0175, 0.0703, 0.0714, -0.0068,
63
+ # 0.1214, -0.0004, 0.0020, 0.0255, 0.0424, -0.0030, 0.0318, 0.1227,
64
+ # 0.0676, -0.0723, 0.0970, 0.0637, -0.0140, -0.0283, -0.0120, 0.0343,
65
+ # -0.0890, 0.0680, 0.0514, 0.0513, 0.0627, -0.0284, -0.0479, 0.0068,
66
+ # -0.0794, 0.0202, 0.0208, -0.0113, -0.0747, 0.0045, -0.0854, -0.0609,
67
+ # -0.0078, 0.1168, 0.0618, -0.0223, -0.0755, 0.0182, -0.0128, 0.1116,
68
+ # 0.0240, 0.0342, 0.0119, -0.0235, -0.0150, -0.0228, -0.0568, -0.1528,
69
+ # 0.0164, -0.0268, 0.0727, -0.0569, 0.1306, 0.0643, -0.0158, -0.1070,
70
+ # -0.0107, -0.0139, -0.0363, 0.0366, -0.0986, -0.0628, -0.0277, 0.0316,
71
+ # 0.0363, 0.0038, -0.1092, -0.0679, -0.1398, -0.0648, 0.1711, -0.0666,
72
+ # 0.0563, 0.0581, 0.0226, 0.0347, -0.0672, -0.0229, -0.0565, 0.0623,
73
+ # 0.1089, -0.0687, -0.0901, -0.0073, 0.0426, 0.0870, -0.0390, -0.0144,
74
+ # -0.0166, 0.0262, -0.0310, 0.0467, -0.0164, -0.0700, -0.0602, -0.0720,
75
+ # -0.0386, 0.0067, -0.0337, -0.0053, 0.0829, 0.1004, 0.0427, 0.0026,
76
+ # -0.0537, 0.0951, 0.0584, -0.0583, -0.0208, 0.0124, 0.0067, 0.0403,
77
+ # 0.0091, -0.0044, -0.0036, 0.0524, 0.1103, -0.1511, -0.0479, 0.1709,
78
+ # 0.0772, 0.0721, -0.0332, 0.0866, 0.0799, -0.0581, 0.0713, 0.0218],
79
+ # device='cuda:0', grad_fn=<SelectBackward0>)
80
+ ```
81
+
82
+ ## Pretraining data
83
+
84
+ This checkpoint is trained on the stricter permissive subset of the deduplicated version of
85
+ the [github-code dataset](https://huggingface.co/datasets/codeparrot/github-code).
86
+ The data is preprocessed by reserving only permissively licensed code ("mit" “apache-2”, “bsd-3-clause”, “bsd-2-clause”,
87
+ “cc0-1.0”, “unlicense”, “isc”).
88
+ Supported languages (9 in total) are as follows:
89
+ `c`, `c++`, `c-sharp`, `go`, `java`, `javascript`, `php`, `python`, `ruby.`
90
+
91
+ ## Training procedure
92
+
93
+ This checkpoint is first trained on the unimodal code data at the first-stage pretraining and then on bimodal text-code
94
+ pair data using the proposed mixture of pretraining tasks.
95
+ Please refer to the paper for more details.
96
+
97
+ ## Evaluation results
98
+
99
+ We show the zero-shot results of this checkpoint on 6 downstream code retrieval tasks from CodeXGLUE in the following table.
100
+ | Ruby | JavaScript | Go | Python | Java | PHP | Overall |
101
+ | ----- | ---------- | ----- | ------ | ----- | ----- | ------- |
102
+ | 74.51 | 69.07 | 90.69 | 71.55 | 71.82 | 67.72 | 74.23 |
103
+
104
+ ## BibTeX entry and citation info
105
+
106
+ ```bibtex
107
+ @article{wang2023codet5plus,
108
+ title={CodeT5+: Open Code Large Language Models for Code Understanding and Generation},
109
+ author={Wang, Yue and Le, Hung and Gotmare, Akhilesh Deepak and Bui, Nghi D.Q. and Li, Junnan and Hoi, Steven C. H.},
110
+ journal={arXiv preprint},
111
+ year={2023}
112
+ }
113
+ ```