Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,119 @@
|
|
1 |
---
|
2 |
license: cc-by-4.0
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: cc-by-4.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
tags:
|
6 |
+
- music
|
7 |
+
- art
|
8 |
---
|
9 |
+
|
10 |
+
# MusiLingo-short-v1
|
11 |
+
This repo contains the code for the following paper.
|
12 |
+
__[MusiLingo: Bridging Music and Text with Pre-trained Language Models for Music Captioning and Query Response](https://arxiv.org/abs/2309.08730)__
|
13 |
+
|
14 |
+
You can refer to more information at the [GitHub repo](https://github.com/zihaod/MusiLingo)
|
15 |
+
|
16 |
+
You can use the [MusicInstruct (MI)](https://huggingface.co/datasets/m-a-p/Music-Instruct) dataset for the following demo
|
17 |
+
|
18 |
+
This checkpoint is developped on the MI-short.
|
19 |
+
|
20 |
+
# Inference Code
|
21 |
+
```
|
22 |
+
from tqdm.auto import tqdm
|
23 |
+
|
24 |
+
import torch
|
25 |
+
from torch.utils.data import DataLoader
|
26 |
+
from transformers import Wav2Vec2FeatureExtractor
|
27 |
+
from transformers import StoppingCriteria, StoppingCriteriaList
|
28 |
+
|
29 |
+
|
30 |
+
|
31 |
+
class StoppingCriteriaSub(StoppingCriteria):
|
32 |
+
def __init__(self, stops=[], encounters=1):
|
33 |
+
super().__init__()
|
34 |
+
self.stops = stops
|
35 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
|
36 |
+
for stop in self.stops:
|
37 |
+
if torch.all((stop == input_ids[0][-len(stop):])).item():
|
38 |
+
return True
|
39 |
+
return False
|
40 |
+
|
41 |
+
def answer(self, samples, stopping, max_new_tokens=300, num_beams=1, min_length=1, top_p=0.5,
|
42 |
+
repetition_penalty=1.0, length_penalty=1, temperature=0.1, max_length=2000):
|
43 |
+
audio = samples["audio"].cuda()
|
44 |
+
audio_embeds, atts_audio = self.encode_audio(audio)
|
45 |
+
if 'instruction_input' in samples: # instruction dataset
|
46 |
+
#print('Instruction Batch')
|
47 |
+
instruction_prompt = []
|
48 |
+
for instruction in samples['instruction_input']:
|
49 |
+
prompt = '<Audio><AudioHere></Audio> ' + instruction
|
50 |
+
instruction_prompt.append(self.prompt_template.format(prompt))
|
51 |
+
audio_embeds, atts_audio = self.instruction_prompt_wrap(audio_embeds, atts_audio, instruction_prompt)
|
52 |
+
self.llama_tokenizer.padding_side = "right"
|
53 |
+
batch_size = audio_embeds.shape[0]
|
54 |
+
bos = torch.ones([batch_size, 1],
|
55 |
+
dtype=torch.long,
|
56 |
+
device=torch.device('cuda')) * self.llama_tokenizer.bos_token_id
|
57 |
+
bos_embeds = self.llama_model.model.embed_tokens(bos)
|
58 |
+
atts_bos = atts_audio[:, :1]
|
59 |
+
inputs_embeds = torch.cat([bos_embeds, audio_embeds], dim=1)
|
60 |
+
attention_mask = torch.cat([atts_bos, atts_audio], dim=1)
|
61 |
+
outputs = self.llama_model.generate(
|
62 |
+
inputs_embeds=inputs_embeds,
|
63 |
+
max_new_tokens=max_new_tokens,
|
64 |
+
stopping_criteria=stopping,
|
65 |
+
num_beams=num_beams,
|
66 |
+
do_sample=True,
|
67 |
+
min_length=min_length,
|
68 |
+
top_p=top_p,
|
69 |
+
repetition_penalty=repetition_penalty,
|
70 |
+
length_penalty=length_penalty,
|
71 |
+
temperature=temperature,
|
72 |
+
)
|
73 |
+
output_token = outputs[0]
|
74 |
+
if output_token[0] == 0: # the model might output a unknow token <unk> at the beginning. remove it
|
75 |
+
output_token = output_token[1:]
|
76 |
+
if output_token[0] == 1: # if there is a start token <s> at the beginning. remove it
|
77 |
+
output_token = output_token[1:]
|
78 |
+
output_text = self.llama_tokenizer.decode(output_token, add_special_tokens=False)
|
79 |
+
output_text = output_text.split('###')[0] # remove the stop sign '###'
|
80 |
+
output_text = output_text.split('Assistant:')[-1].strip()
|
81 |
+
return output_text
|
82 |
+
|
83 |
+
processor = Wav2Vec2FeatureExtractor.from_pretrained("m-a-p/MERT-v1-330M",trust_remote_code=True)
|
84 |
+
ds = CMIDataset(processor, 'path/to/MI_dataset', 'test', question_type='short')
|
85 |
+
dl = DataLoader(
|
86 |
+
ds,
|
87 |
+
batch_size=1,
|
88 |
+
num_workers=0,
|
89 |
+
pin_memory=True,
|
90 |
+
shuffle=False,
|
91 |
+
drop_last=True,
|
92 |
+
collate_fn=ds.collater
|
93 |
+
)
|
94 |
+
|
95 |
+
stopping = StoppingCriteriaList([StoppingCriteriaSub([torch.tensor([835]).cuda(),
|
96 |
+
torch.tensor([2277, 29937]).cuda()])])
|
97 |
+
|
98 |
+
from transformers import AutoModel
|
99 |
+
model_short = AutoModel.from_pretrained("m-a-p/MusiLingo-short-v1")
|
100 |
+
|
101 |
+
for idx, sample in tqdm(enumerate(dl)):
|
102 |
+
ans = answer(Musilingo_short.model, sample, stopping, length_penalty=100, temperature=0.1)
|
103 |
+
txt = sample['text_input'][0]
|
104 |
+
print(txt)
|
105 |
+
print(and)
|
106 |
+
```
|
107 |
+
|
108 |
+
# Citing This Work
|
109 |
+
|
110 |
+
If you find the work useful for your research, please consider citing it using the following BibTeX entry:
|
111 |
+
```
|
112 |
+
@inproceedings{deng2024musilingo,
|
113 |
+
title={MusiLingo: Bridging Music and Text with Pre-trained Language Models for Music Captioning and Query Response},
|
114 |
+
author={Deng, Zihao and Ma, Yinghao and Liu, Yudong and Guo, Rongchen and Zhang, Ge and Chen, Wenhu and Huang, Wenhao and Benetos, Emmanouil},
|
115 |
+
booktitle={Proceedings of the 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2024)},
|
116 |
+
year={2024},
|
117 |
+
organization={Association for Computational Linguistics}
|
118 |
+
}
|
119 |
+
```
|