Update README.md
Browse files
README.md
CHANGED
@@ -20,7 +20,66 @@ musicgen-songstarter-v0.1 is a [`musicgen-melody`](https://huggingface.co/facebo
|
|
20 |
|
21 |
This is a proof of concept. Hopefully, we will be able to collect more data and train a better models in the future.
|
22 |
|
23 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
## Prompt Format
|
26 |
|
@@ -30,8 +89,14 @@ Follow the following prompt format:
|
|
30 |
{tag_1}, {tag_1}, ..., {tag_n}, {key}, {bpm} bpm
|
31 |
```
|
32 |
|
33 |
-
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
```
|
36 |
hip hop
|
37 |
trap
|
@@ -194,72 +259,3 @@ music box
|
|
194 |
glitch
|
195 |
clarinet
|
196 |
```
|
197 |
-
</details>
|
198 |
-
|
199 |
-
|
200 |
-
For example:
|
201 |
-
|
202 |
-
```
|
203 |
-
hip hop, soul, piano, chords, jazz, neo jazz, G# minor, 140 bpm
|
204 |
-
```
|
205 |
-
|
206 |
-
## Usage
|
207 |
-
|
208 |
-
Install [audiocraft](https://github.com/facebookresearch/audiocraft):
|
209 |
-
|
210 |
-
```
|
211 |
-
pip install -U git+https://github.com/facebookresearch/audiocraft#egg=audiocraft
|
212 |
-
```
|
213 |
-
|
214 |
-
Then, you should be able to load this model just like any other musicgen checkpoint here on the Hub:
|
215 |
-
|
216 |
-
```python
|
217 |
-
from audiocraft.models import musicgen
|
218 |
-
|
219 |
-
model = musicgen.MusicGen.get_pretrained('nateraw/musicgen-songstarter-v0.1', device='cuda')
|
220 |
-
```
|
221 |
-
|
222 |
-
To generate and save audio samples, you can do:
|
223 |
-
|
224 |
-
```python
|
225 |
-
from datetime import datetime
|
226 |
-
from pathlib import Path
|
227 |
-
|
228 |
-
from audiocraft.models import musicgen
|
229 |
-
from audiocraft.data.audio import audio_write
|
230 |
-
from audiocraft.utils.notebook import display_audio
|
231 |
-
|
232 |
-
model = musicgen.MusicGen.get_pretrained('nateraw/musicgen-songstarter-v0.1', device='cuda')
|
233 |
-
|
234 |
-
# path to save our samples.
|
235 |
-
out_dir = Path("./samples")
|
236 |
-
out_dir.mkdir(exist_ok=True, parents=True)
|
237 |
-
|
238 |
-
model.set_generation_params(
|
239 |
-
duration=15,
|
240 |
-
use_sampling=True,
|
241 |
-
temperature=1.0,
|
242 |
-
top_k=250,
|
243 |
-
cfg_coef=3.0,
|
244 |
-
)
|
245 |
-
|
246 |
-
text = "hip hop, soul, piano, chords, jazz, neo jazz, G# minor, 140 bpm"
|
247 |
-
N = 4
|
248 |
-
out = model.generate(
|
249 |
-
[text] * N,
|
250 |
-
progress=True,
|
251 |
-
)
|
252 |
-
|
253 |
-
# Write to files
|
254 |
-
dt_str = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
|
255 |
-
for i in range(N):
|
256 |
-
audio_write(
|
257 |
-
out_dir / f"{dt_str}_{i:02d}",
|
258 |
-
out[i].cpu(),
|
259 |
-
model.sample_rate,
|
260 |
-
strategy="loudness",
|
261 |
-
)
|
262 |
-
|
263 |
-
# Or, if in a notebook, display audio widgets
|
264 |
-
# display_audio(out, model.sample_rate)
|
265 |
-
```
|
|
|
20 |
|
21 |
This is a proof of concept. Hopefully, we will be able to collect more data and train a better models in the future.
|
22 |
|
23 |
+
## Usage
|
24 |
+
|
25 |
+
Install [audiocraft](https://github.com/facebookresearch/audiocraft):
|
26 |
+
|
27 |
+
```
|
28 |
+
pip install -U git+https://github.com/facebookresearch/audiocraft#egg=audiocraft
|
29 |
+
```
|
30 |
+
|
31 |
+
Then, you should be able to load this model just like any other musicgen checkpoint here on the Hub:
|
32 |
+
|
33 |
+
```python
|
34 |
+
from audiocraft.models import musicgen
|
35 |
+
|
36 |
+
model = musicgen.MusicGen.get_pretrained('nateraw/musicgen-songstarter-v0.1', device='cuda')
|
37 |
+
```
|
38 |
+
|
39 |
+
To generate and save audio samples, you can do:
|
40 |
+
|
41 |
+
```python
|
42 |
+
from datetime import datetime
|
43 |
+
from pathlib import Path
|
44 |
+
|
45 |
+
from audiocraft.models import musicgen
|
46 |
+
from audiocraft.data.audio import audio_write
|
47 |
+
from audiocraft.utils.notebook import display_audio
|
48 |
+
|
49 |
+
model = musicgen.MusicGen.get_pretrained('nateraw/musicgen-songstarter-v0.1', device='cuda')
|
50 |
+
|
51 |
+
# path to save our samples.
|
52 |
+
out_dir = Path("./samples")
|
53 |
+
out_dir.mkdir(exist_ok=True, parents=True)
|
54 |
+
|
55 |
+
model.set_generation_params(
|
56 |
+
duration=15,
|
57 |
+
use_sampling=True,
|
58 |
+
temperature=1.0,
|
59 |
+
top_k=250,
|
60 |
+
cfg_coef=3.0,
|
61 |
+
)
|
62 |
+
|
63 |
+
text = "hip hop, soul, piano, chords, jazz, neo jazz, G# minor, 140 bpm"
|
64 |
+
N = 4
|
65 |
+
out = model.generate(
|
66 |
+
[text] * N,
|
67 |
+
progress=True,
|
68 |
+
)
|
69 |
+
|
70 |
+
# Write to files
|
71 |
+
dt_str = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
|
72 |
+
for i in range(N):
|
73 |
+
audio_write(
|
74 |
+
out_dir / f"{dt_str}_{i:02d}",
|
75 |
+
out[i].cpu(),
|
76 |
+
model.sample_rate,
|
77 |
+
strategy="loudness",
|
78 |
+
)
|
79 |
+
|
80 |
+
# Or, if in a notebook, display audio widgets
|
81 |
+
# display_audio(out, model.sample_rate)
|
82 |
+
```
|
83 |
|
84 |
## Prompt Format
|
85 |
|
|
|
89 |
{tag_1}, {tag_1}, ..., {tag_n}, {key}, {bpm} bpm
|
90 |
```
|
91 |
|
92 |
+
For example:
|
93 |
+
|
94 |
+
```
|
95 |
+
hip hop, soul, piano, chords, jazz, neo jazz, G# minor, 140 bpm
|
96 |
+
```
|
97 |
+
|
98 |
+
The training dataset had the following tags in it:
|
99 |
+
|
100 |
```
|
101 |
hip hop
|
102 |
trap
|
|
|
259 |
glitch
|
260 |
clarinet
|
261 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|