metadata

language: Chinese
widget:
  - text: 最美的不是下雨天，是曾与你躲过雨的屋檐

Chinese GPT2 Model

Model description

The model is used to generate Chinese lyrics. You can download the model from the link gpt2-chinese-lyric

How to use

You can use the model directly with a pipeline for text generation:

>>> from transformers import BertTokenizer, GPT2LMHeadModel, TextGenerationPipeline
>>> tokenizer = BertTokenizer.from_pretrained("uer/gpt2-chinese-lyric")
>>> model = GPT2LMHeadModel.from_pretrained("uer/gpt2-chinese-lyric")
>>> text_generator = TextGenerationPipeline(model, tokenizer)   
>>> text_generator("最美的不是下雨天，是曾与你躲过雨的屋檐", max_length=100, do_sample=True)
    [{'generated_text': '最美的不是下雨天，是曾与你躲过雨的屋檐 ， 下 课 铃 声 响 起 的 瞬 间 ， 我 们 的 笑 脸 ， 有 太 多 回 忆 在 浮 现 ， 是 你 总 在 我 身 边 ， 不 知 道 会 不 会 再 见 ， 从 现 在 开 始 到 永 远 ， 想 说 的 语 言 凝 结 成 一 句 ， 不 管 我 们 是 否 能 够 兑 现 ， 想 说 的 语 言 凝 结'}]

Training data

Training data contains 150,000 Chinese lyrics which are collected by Chinese-Lyric-Corpus and MusicLyricChatbot projects

Training procedure

The model is pre-trained by UER-py on Tencent Cloud TI-ONE. We pre-train 100,000 steps with a sequence length of 512 on the basis of the model gpt2-base-chinese-cluecorpussmall

python3 preprocess.py --corpus_path corpora/lyric.txt \
                      --vocab_path models/google_zh_vocab.txt \
                      --dataset_path lyric_lm_seq512_dataset.pt \
                      --seq_length 512 --processes_num 32 --target lm

python3 pretrain.py --dataset_path lyric_lm_seq512_dataset.pt \
                    --pretrained_model_path gpt2-base-chinese-cluecorpussmall/pytorch_model.bin\
                    --vocab_path models/google_zh_vocab.txt \
                    --output_model_path models/lyric_gpt2_seq512_model.bin \
                    --config_path models/bert_base_config.json --learning_rate 5e-5 \
                    --world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 --tie_weight \
                    --embedding word_pos --remove_embedding_layernorm \
                    --encoder transformer --mask causal --layernorm_positioning pre \
                    --target lm --batch_size 64 --total_steps 100000 \
                    --save_checkpoint_steps 10000 --report_steps 5000

Finally, we convert the pre-trained model into Huggingface's format:

python3 scripts/convert_gpt2_from_uer_to_huggingface.py --input_model_path lyric_gpt2_seq512_model.bin-100000 \
                                                        --output_model_path pytorch_model.bin \
                                                        --layers_num 12

BibTeX entry and citation info

@article{zhao2019uer,
  title={UER: An Open-Source Toolkit for Pre-training Models},
  author={Zhao, Zhe and Chen, Hui and Zhang, Jinbin and Zhao, Xin and Liu, Tao and Lu, Wei and Chen, Xi and Deng, Haotang and Ju, Qi and Du, Xiaoyong},
  journal={EMNLP-IJCNLP 2019},
  pages={241},
  year={2019}
}