Spaces:
Runtime error
Runtime error
sander-wood
commited on
Commit
•
b3bc81c
1
Parent(s):
4c5fbb0
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,106 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import torch
|
3 |
+
import random
|
4 |
+
from unidecode import unidecode
|
5 |
+
from samplings import top_p_sampling, temperature_sampling
|
6 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
7 |
+
|
8 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
9 |
+
|
10 |
+
description = """
|
11 |
+
<div>
|
12 |
+
|
13 |
+
<a style="display:inline-block" href='https://github.com/sander-wood/text-to-music'><img src='https://img.shields.io/github/stars/sander-wood/text-to-music?style=social' /></a>
|
14 |
+
<a style="display:inline-block" href="https://arxiv.org/pdf/2211.11216.pdf"><img src="https://img.shields.io/badge/arXiv-2211.11216-b31b1b.svg"></a>
|
15 |
+
<a class="duplicate-button" style="display:inline-block" target="_blank" href="https://huggingface.co/spaces/sander-wood/text-to-music?duplicate=true"><img style="margin-top:0;margin-bottom:0" src="https://huggingface.co/datasets/huggingface/badges/raw/main/duplicate-this-space-md-dark.svg" alt="Duplicate Space"></a>
|
16 |
+
</div>
|
17 |
+
|
18 |
+
## ℹ️ How to use this demo?
|
19 |
+
1. Enter a query in the text box.
|
20 |
+
2. You can set the parameters (i.e., number of tunes, maximum length, top-p, temperature, and random seed) for the generation. (optional)
|
21 |
+
3. Click "Submit" and wait for the result.
|
22 |
+
4. The generated ABC notation can be converted to MIDI or PDF using [EasyABC](https://sourceforge.net/projects/easyabc/), you can also use this [online renderer](https://ldzhangyx.github.io/abc/) to render the ABC notation.
|
23 |
+
|
24 |
+
## ❕Notice
|
25 |
+
- The text box is case-sensitive.
|
26 |
+
- The demo is based on BART-base and fine-tuned on the Textune dataset (282,870 text-music pairs).
|
27 |
+
- The demo only supports English text as the input.
|
28 |
+
- The demo is still in the early stage, and the generated music is not perfect. If you have any suggestions, please feel free to contact me via [email](mailto:[email protected]).
|
29 |
+
"""
|
30 |
+
|
31 |
+
|
32 |
+
examples = [
|
33 |
+
["This is a traditional Irish dance music.\nNote Length-1/8\nMeter-6/8\nKey-D", 3, 1024, 0.9, 1.0, 0],
|
34 |
+
["This is a jazz-swing lead sheet with chord and vocal.", 3, 1024, 0.9, 1.0, 0]
|
35 |
+
]
|
36 |
+
|
37 |
+
|
38 |
+
def generate_abc(text, num_tunes, max_length, top_p, temperature, seed):
|
39 |
+
|
40 |
+
try:
|
41 |
+
seed = int(seed)
|
42 |
+
except:
|
43 |
+
seed = None
|
44 |
+
|
45 |
+
print("Input Text: " + text)
|
46 |
+
text = unidecode(text)
|
47 |
+
tokenizer = AutoTokenizer.from_pretrained('sander-wood/text-to-music')
|
48 |
+
model = AutoModelForSeq2SeqLM.from_pretrained('sander-wood/text-to-music')
|
49 |
+
model = model.to(device)
|
50 |
+
|
51 |
+
input_ids = tokenizer(text,
|
52 |
+
return_tensors='pt',
|
53 |
+
truncation=True,
|
54 |
+
max_length=max_length)['input_ids'].to(device)
|
55 |
+
decoder_start_token_id = model.config.decoder_start_token_id
|
56 |
+
eos_token_id = model.config.eos_token_id
|
57 |
+
random.seed(seed)
|
58 |
+
tunes = ""
|
59 |
+
|
60 |
+
for n_idx in range(num_tunes):
|
61 |
+
print("\nX:"+str(n_idx+1)+"\n", end="")
|
62 |
+
tunes += "X:"+str(n_idx+1)+"\n"
|
63 |
+
decoder_input_ids = torch.tensor([[decoder_start_token_id]])
|
64 |
+
|
65 |
+
for t_idx in range(max_length):
|
66 |
+
|
67 |
+
if seed!=None:
|
68 |
+
n_seed = random.randint(0, 1000000)
|
69 |
+
random.seed(n_seed)
|
70 |
+
else:
|
71 |
+
n_seed = None
|
72 |
+
outputs = model(input_ids=input_ids,
|
73 |
+
decoder_input_ids=decoder_input_ids.to(device))
|
74 |
+
probs = outputs.logits[0][-1]
|
75 |
+
probs = torch.nn.Softmax(dim=-1)(probs).cpu().detach().numpy()
|
76 |
+
sampled_id = temperature_sampling(probs=top_p_sampling(probs,
|
77 |
+
top_p=top_p,
|
78 |
+
seed=n_seed,
|
79 |
+
return_probs=True),
|
80 |
+
seed=n_seed,
|
81 |
+
temperature=temperature)
|
82 |
+
decoder_input_ids = torch.cat((decoder_input_ids, torch.tensor([[sampled_id]])), 1)
|
83 |
+
if sampled_id!=eos_token_id:
|
84 |
+
sampled_token = tokenizer.decode([sampled_id])
|
85 |
+
print(sampled_token, end="")
|
86 |
+
tunes += sampled_token
|
87 |
+
else:
|
88 |
+
tunes += '\n'
|
89 |
+
break
|
90 |
+
|
91 |
+
return tunes
|
92 |
+
|
93 |
+
input_text = gr.inputs.Textbox(lines=5, label="Input Text", placeholder="Describe the music you want to generate ...")
|
94 |
+
input_num_tunes = gr.inputs.Slider(minimum=1, maximum=10, step=1, default=1, label="Number of Tunes")
|
95 |
+
input_max_length = gr.inputs.Slider(minimum=10, maximum=1000, step=10, default=500, label="Max Length")
|
96 |
+
input_top_p = gr.inputs.Slider(minimum=0.0, maximum=1.0, step=0.05, default=0.9, label="Top P")
|
97 |
+
input_temperature = gr.inputs.Slider(minimum=0.0, maximum=2.0, step=0.1, default=1.0, label="Temperature")
|
98 |
+
input_seed = gr.inputs.Textbox(lines=1, label="Seed (int)", default="None")
|
99 |
+
output_abc = gr.outputs.Textbox(label="Generated Tunes")
|
100 |
+
|
101 |
+
gr.Interface(fn=generate_abc,
|
102 |
+
inputs=[input_text, input_num_tunes, input_max_length, input_top_p, input_temperature, input_seed],
|
103 |
+
outputs=output_abc,
|
104 |
+
title="Textune: Generating Tune from Text",
|
105 |
+
description=description,
|
106 |
+
examples=examples).launch(debug=True)
|