alecsharpie
commited on
Commit
•
c464789
1
Parent(s):
30a0aef
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- code
|
4 |
+
license: bsd-3-clause
|
5 |
+
tags:
|
6 |
+
- code
|
7 |
+
- generative
|
8 |
+
datasets:
|
9 |
+
- bigcode/the-stack
|
10 |
+
---
|
11 |
+
|
12 |
+
# CodeGen (CodeGen-CSS 350M)
|
13 |
+
|
14 |
+
## Model description
|
15 |
+
|
16 |
+
CodeGen is a family of autoregressive language models for **program synthesis** from the paper: [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong. The models are originally released in [this repository](https://github.com/salesforce/CodeGen), under 3 pre-training data variants (`NL`, `Multi`, `Mono`) and 4 model size variants (`350M`, `2B`, `6B`, `16B`).
|
17 |
+
|
18 |
+
The checkpoint included in this repository is finetuned on top of the **CodeGen-Multi 350M**, where "Multi" means the model is initialized with *CodeGen-NL 350M* and further pre-trained on a dataset of multiple programming languages, and "350M" refers to the number of trainable parameters.
|
19 |
+
|
20 |
+
It has been finetuned on CSS code contained in bigcode/the-stack dataset on huggingface
|
21 |
+
|
22 |
+
## Training data
|
23 |
+
|
24 |
+
This checkpoint (CodeGen-Multi 350M) was firstly initialized with *CodeGen-NL 350M*, and then pre-trained on [BigQuery](https://console.cloud.google.com/marketplace/details/github/github-repos), a large-scale dataset of multiple programming languages from GitHub repositories. The data consists of 119.2B tokens and includes C, C++, Go, Java, JavaScript, and Python.
|
25 |
+
|
26 |
+
Lastly it has been finetuned on CSS code contained in [bigcode/the-stack](https://huggingface.co/datasets/bigcode/the-stack) dataset on huggingface
|
27 |
+
|
28 |
+
## Training procedure
|
29 |
+
|
30 |
+
Initially:
|
31 |
+
CodeGen was trained using cross-entropy loss to maximize the likelihood of sequential inputs.
|
32 |
+
The family of models are trained using multiple TPU-v4-512 by Google, leveraging data and model parallelism.
|
33 |
+
See Section 2.3 of the [paper](https://arxiv.org/abs/2203.13474) for more details.
|
34 |
+
|
35 |
+
Finetune:
|
36 |
+
I fine tuned the 350M model on a single A100 with 40Gb of RAM, with batch size 10 and an input length of 512 tokens
|
37 |
+
Used 80-90% of the RAM
|
38 |
+
|
39 |
+
## Intended Use and Limitations
|
40 |
+
|
41 |
+
As an autoregressive language model, CodeGen is capable of extracting features from given natural language and programming language texts, and calculating the likelihood of them.
|
42 |
+
However, the model is intended for and best at **program synthesis**, that is, generating executable code given English prompts, where the prompts should be in the form of a comment string. The model can complete partially-generated code as well.
|
43 |
+
|
44 |
+
## How to use
|
45 |
+
|
46 |
+
This model can be easily loaded using the `AutoModelForCausalLM` functionality:
|
47 |
+
|
48 |
+
```python
|
49 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
50 |
+
|
51 |
+
tokenizer = AutoTokenizer.from_pretrained("Salesforce/codegen-350M-multi")
|
52 |
+
model = AutoModelForCausalLM.from_pretrained("alecsharpie/codegen_350m_css")
|
53 |
+
|
54 |
+
text = ".header-container {"
|
55 |
+
|
56 |
+
input_ids = tokenizer(text, return_tensors="pt").input_ids
|
57 |
+
generated_ids = model.generate(input_ids, max_length=128)
|
58 |
+
print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
|
59 |
+
```
|
60 |
+
|
61 |
+
## BibTeX entry and citation info
|
62 |
+
|
63 |
+
```bibtex
|
64 |
+
@article{Nijkamp2022ACP,
|
65 |
+
title={A Conversational Paradigm for Program Synthesis},
|
66 |
+
author={Nijkamp, Erik and Pang, Bo and Hayashi, Hiroaki and Tu, Lifu and Wang, Huan and Zhou, Yingbo and Savarese, Silvio and Xiong, Caiming},
|
67 |
+
journal={arXiv preprint},
|
68 |
+
year={2022}
|
69 |
+
}
|
70 |
+
```
|