| --- |
| license: apache-2.0 |
| pipeline_tag: text-generation |
| --- |
| |
|
|
|
|
|
|
| <p align="center"> |
| <b><font size="6">[ACL 2025] SongComposer</font></b> |
| <p> |
| |
| <div align="center"> |
|
|
| [💻Github Repo](https://github.com/pjlab-songcomposer/songcomposer) |
|
|
| [📖Paper](https://arxiv.org/abs/2402.17645) |
|
|
| </div> |
|
|
| **SongComposer** is a language large model (LLM) based on [InternLM2](https://github.com/InternLM/InternLM) for lyric and melody composition in song generation. |
|
|
| We release SongComposer series in two versions: |
|
|
| - SongComposer_pretrain: The pretrained SongComposer with InternLM2 as the initialization of the LLM, gains basic knowledge of lyric and melody. |
| - SongComposer_sft: The finetuned SongComposer for *instruction-following song generation* including lyric to melody, melody to lyric, song continuation, text to song. |
|
|
| ### Import from Transformers |
| To load the SongComposer_pretrain model using Transformers, use the following code: |
| ```python |
| from transformers import AutoTokenizer, AutoModel |
| ckpt_path = "Mar2Ding/songcomposer_pretrain" |
| tokenizer = AutoTokenizer.from_pretrained(ckpt_path, trust_remote_code=True) |
| model = AutoModel.from_pretrained(ckpt_path, trust_remote_code=True).cuda().half() |
| prompt = '<bop> Total 7 lines. The first line:可,<D4>,<137>,<79>|惜,<D#4>,<137>,<79>|这,<F4>,<137>,<88>|是,<F4>,<121>,<79>|属,<F4>,<121>,<79>|于,<D#4>,<214>,<88>|你,<D#4>,<141>,<79>|的,<D4>,<130>,<79>|风,<C4>,<151>,<79>|景,<A#3> <F3>,<181><137>,<79>\n' |
| model.inference_pretrain(prompt, tokenizer) |
| ``` |
| |
| ### 通过 Transformers 加载 |
| 通过以下的代码加载 SongComposer_pretrain 模型 |
| ```python |
| from transformers import AutoTokenizer, AutoModel |
| ckpt_path = "Mar2Ding/songcomposer_pretrain" |
| tokenizer = AutoTokenizer.from_pretrained(ckpt_path, trust_remote_code=True) |
| model = AutoModel.from_pretrained(ckpt_path, trust_remote_code=True).cuda().half() |
| prompt = '<bop> Total 7 lines. The first line:可,<D4>,<137>,<79>|惜,<D#4>,<137>,<79>|这,<F4>,<137>,<88>|是,<F4>,<121>,<79>|属,<F4>,<121>,<79>|于,<D#4>,<214>,<88>|你,<D#4>,<141>,<79>|的,<D4>,<130>,<79>|风,<C4>,<151>,<79>|景,<A#3> <F3>,<181><137>,<79>\n' |
| model.inference_pretrain(prompt, tokenizer) |
| ``` |
| |
| ### Open Source License |
| The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow free commercial usage. |
