README.md

---
language: ja
tags:
- ja
- japanese
- gpt2
- text-generation
- lm
- nlp
license: mit
widget:
- text: "未来に揺れる花 過去にもあった花"
---

# Japanese GPT2 Lyric Model

## Model description

The model is used to generate Japanese lyrics.

## How to use

```python
import torch
from transformers import T5Tokenizer, GPT2LMHeadModel, TextGenerationPipeline

tokenizer = T5Tokenizer.from_pretrained("skytnt/gpt2-japanese-lyric-small")
model = GPT2LMHeadModel.from_pretrained("skytnt/gpt2-japanese-lyric-small")


def gen_lyric(prompt_text: str):
    prompt_text = prompt_text.replace("\n", "[SEP]")
    prompt_tokens = tokenizer.tokenize(prompt_text)
    prompt_token_ids = tokenizer.convert_tokens_to_ids(prompt_tokens)
    prompt_tensor = torch.LongTensor(prompt_token_ids)
    prompt_tensor = prompt_tensor.view(1, -1)
    # model forward
    output_sequences = model.generate(
        input_ids=prompt_tensor,
        max_length=512,
        top_p=0.95,
        top_k=40,
        temperature=1.0,
        do_sample=True,
        early_stopping=True,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
        num_return_sequences=1
    )

    # convert model outputs to readable sentence
    generated_sequence = output_sequences.tolist()[0]
    generated_tokens = tokenizer.convert_ids_to_tokens(generated_sequence)
    generated_text = tokenizer.convert_tokens_to_string(generated_tokens)
    generated_text = "\n".join([s.strip() for s in generated_text.split('[SEP]')]).replace(' ', '\u3000').replace(
        '</s>', '\n\n---end---')
    return generated_text


print(gen_lyric("未来に揺れる花 過去にもあった花"))

```

## Training data

Training data ([click to download](https://data.anyweb.xyz/dataset/lyric.zip)) contains 48,394 Japanese lyrics which are collected from [NetEasyMusic](https://music.163.com/) by [lyric_download](https://github.com/SkyTNT/lyric_downlowd)