This repository contains code and checkpoints for CPT.
CPT: A Pre-Trained Unbalanced Transformer for Both Chinese Language Understanding and Generation
Yunfan Shao, Zhichao Geng, Yitao Liu, Junqi Dai, Fei Yang, Li Zhe, Hujun Bao, Xipeng Qiu
Aiming to unify both NLU and NLG tasks, We propose a novel Chinese Pre-trained Un-balanced Transformer (CPT), which is an unbalanced Transformer encoder-decoder pre-trained with MLM and DAE jointly.
The architecture of CPT is a variant of the full Transformer and consists of three parts:
- Shared Encoder (S-Enc): a Transformer encoder with fully-connected self-attention, which is designed to capture the common semantic representation for both language understanding and generation.
- Understanding Decoder (U-Dec): a shallow Transformer encoder with fully-connected self-attention, which is designed for NLU tasks. The input of U-Dec is the output of S-Enc.
- Generation Decoder (G-Dec): a Transformer decoder with masked self-attention, which is designed for generation tasks with auto-regressive fashion. G-Dec utilizes the output of S-Enc with cross-attention.
We provide the pre-trained weights of CPT and Chinese BART with source code, which can be directly used in Huggingface-Transformers.
Chinese BART-base: 6 layers Encoder, 6 layers Decoder, 12 Heads and 768 Model dim.Chinese BART-large: 12 layers Encoder, 12 layers Decoder, 16 Heads and 1024 Model dim.CPT-base: 10 layers S-Enc, 2 layers U-Dec/G-Dec, 12 Heads and 768 Model dim.CPT-large: 20 layers S-Enc, 4 layers U-Dec/G-Dec, 16 Heads and 1024 Model dim.
The pre-trained weights can be downloaded here.
| Model | MODEL_NAME |
|---|---|
Chinese BART-base | fnlp/bart-base-chinese |
Chinese BART-large | fnlp/bart-large-chinese |
CPT-base | fnlp/cpt-base |
CPT-large | fnlp/cpt-large |
To use CPT, please import the file finetune/modeling_cpt.py that define the architecture of CPT into your project. Then, use the PTMs as the following example, where MODEL_NAME is the corresponding string that refers to the model.
For CPT:
from modeling_cpt import BertTokenizer, CPTForConditionalGeneration tokenizer = BertTokenizer.from_pretrained("MODEL_NAME") model = CPTForConditionalGeneration.from_pretrained("MODEL_NAME") print(model)For Chinese BART:
from transformers import BertTokenizer, BartForConditionalGeneration tokenizer = BertTokenizer.from_pretrained("MODEL_NAME") model = BartForConditionalGeneration.from_pretrained("MODEL_NAME") print(model)Pre-training code and examples can be find Here.
Fine-tuning code and examples can be find Here.
@article{shao2021cpt, title={CPT: A Pre-Trained Unbalanced Transformer for Both Chinese Language Understanding and Generation}, author={Yunfan Shao and Zhichao Geng and Yitao Liu and Junqi Dai and Fei Yang and Li Zhe and Hujun Bao and Xipeng Qiu}, journal={arXiv preprint arXiv:2109.05729}, year={2021} }