Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,137 +1,211 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
import note_seq
|
3 |
import numpy as np
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
else:
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
if
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
)
|
136 |
|
137 |
-
|
|
|
|
|
|
|
1 |
import numpy as np
|
2 |
+
import torch
|
3 |
+
import gradio as gr
|
4 |
+
import spaces
|
5 |
+
from queue import Queue
|
6 |
+
from threading import Thread
|
7 |
+
from typing import Optional
|
8 |
+
from transformers import MusicgenForConditionalGeneration, MusicgenProcessor, set_seed
|
9 |
+
from transformers.generation.streamers import BaseStreamer
|
10 |
+
|
11 |
+
model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small")
|
12 |
+
processor = MusicgenProcessor.from_pretrained("facebook/musicgen-small")
|
13 |
+
|
14 |
+
title = "9🌍MusicHub - Text to Music Stream Generator"
|
15 |
+
description = """ Facebook MusicGen-Small Model - Generate and stream music with model https://huggingface.co/facebook/musicgen-small """
|
16 |
+
article = """
|
17 |
+
## How It Works:
|
18 |
+
MusicGen is an auto-regressive transformer-based model, meaning generates audio codes (tokens) in a causal fashion.
|
19 |
+
At each decoding step, the model generates a new set of audio codes, conditional on the text input and all previous audio codes. From the
|
20 |
+
frame rate of the [EnCodec model](https://huggingface.co/facebook/encodec_32khz) used to decode the generated codes to audio waveform.
|
21 |
+
"""
|
22 |
+
|
23 |
+
|
24 |
+
class MusicgenStreamer(BaseStreamer):
|
25 |
+
def __init__(
|
26 |
+
self,
|
27 |
+
model: MusicgenForConditionalGeneration,
|
28 |
+
device: Optional[str] = None,
|
29 |
+
play_steps: Optional[int] = 10,
|
30 |
+
stride: Optional[int] = None,
|
31 |
+
timeout: Optional[float] = None,
|
32 |
+
):
|
33 |
+
"""
|
34 |
+
Streamer that stores playback-ready audio in a queue, to be used by a downstream application as an iterator. This is
|
35 |
+
useful for applications that benefit from acessing the generated audio in a non-blocking way (e.g. in an interactive
|
36 |
+
Gradio demo).
|
37 |
+
Parameters:
|
38 |
+
model (`MusicgenForConditionalGeneration`):
|
39 |
+
The MusicGen model used to generate the audio waveform.
|
40 |
+
device (`str`, *optional*):
|
41 |
+
The torch device on which to run the computation. If `None`, will default to the device of the model.
|
42 |
+
play_steps (`int`, *optional*, defaults to 10):
|
43 |
+
The number of generation steps with which to return the generated audio array. Using fewer steps will
|
44 |
+
mean the first chunk is ready faster, but will require more codec decoding steps overall. This value
|
45 |
+
should be tuned to your device and latency requirements.
|
46 |
+
stride (`int`, *optional*):
|
47 |
+
The window (stride) between adjacent audio samples. Using a stride between adjacent audio samples reduces
|
48 |
+
the hard boundary between them, giving smoother playback. If `None`, will default to a value equivalent to
|
49 |
+
play_steps // 6 in the audio space.
|
50 |
+
timeout (`int`, *optional*):
|
51 |
+
The timeout for the audio queue. If `None`, the queue will block indefinitely. Useful to handle exceptions
|
52 |
+
in `.generate()`, when it is called in a separate thread.
|
53 |
+
"""
|
54 |
+
self.decoder = model.decoder
|
55 |
+
self.audio_encoder = model.audio_encoder
|
56 |
+
self.generation_config = model.generation_config
|
57 |
+
self.device = device if device is not None else model.device
|
58 |
+
|
59 |
+
# variables used in the streaming process
|
60 |
+
self.play_steps = play_steps
|
61 |
+
if stride is not None:
|
62 |
+
self.stride = stride
|
63 |
+
else:
|
64 |
+
hop_length = np.prod(self.audio_encoder.config.upsampling_ratios)
|
65 |
+
self.stride = hop_length * (play_steps - self.decoder.num_codebooks) // 6
|
66 |
+
self.token_cache = None
|
67 |
+
self.to_yield = 0
|
68 |
+
|
69 |
+
# varibles used in the thread process
|
70 |
+
self.audio_queue = Queue()
|
71 |
+
self.stop_signal = None
|
72 |
+
self.timeout = timeout
|
73 |
+
|
74 |
+
def apply_delay_pattern_mask(self, input_ids):
|
75 |
+
# build the delay pattern mask for offsetting each codebook prediction by 1 (this behaviour is specific to MusicGen)
|
76 |
+
_, decoder_delay_pattern_mask = self.decoder.build_delay_pattern_mask(
|
77 |
+
input_ids[:, :1],
|
78 |
+
pad_token_id=self.generation_config.decoder_start_token_id,
|
79 |
+
max_length=input_ids.shape[-1],
|
80 |
+
)
|
81 |
+
# apply the pattern mask to the input ids
|
82 |
+
input_ids = self.decoder.apply_delay_pattern_mask(input_ids, decoder_delay_pattern_mask)
|
83 |
+
|
84 |
+
# revert the pattern delay mask by filtering the pad token id
|
85 |
+
input_ids = input_ids[input_ids != self.generation_config.pad_token_id].reshape(
|
86 |
+
1, self.decoder.num_codebooks, -1
|
87 |
+
)
|
88 |
+
|
89 |
+
# append the frame dimension back to the audio codes
|
90 |
+
input_ids = input_ids[None, ...]
|
91 |
+
|
92 |
+
# send the input_ids to the correct device
|
93 |
+
input_ids = input_ids.to(self.audio_encoder.device)
|
94 |
+
|
95 |
+
output_values = self.audio_encoder.decode(
|
96 |
+
input_ids,
|
97 |
+
audio_scales=[None],
|
98 |
+
)
|
99 |
+
audio_values = output_values.audio_values[0, 0]
|
100 |
+
return audio_values.cpu().float().numpy()
|
101 |
+
|
102 |
+
def put(self, value):
|
103 |
+
batch_size = value.shape[0] // self.decoder.num_codebooks
|
104 |
+
if batch_size > 1:
|
105 |
+
raise ValueError("MusicgenStreamer only supports batch size 1")
|
106 |
+
|
107 |
+
if self.token_cache is None:
|
108 |
+
self.token_cache = value
|
109 |
+
else:
|
110 |
+
self.token_cache = torch.concatenate([self.token_cache, value[:, None]], dim=-1)
|
111 |
+
|
112 |
+
if self.token_cache.shape[-1] % self.play_steps == 0:
|
113 |
+
audio_values = self.apply_delay_pattern_mask(self.token_cache)
|
114 |
+
self.on_finalized_audio(audio_values[self.to_yield : -self.stride])
|
115 |
+
self.to_yield += len(audio_values) - self.to_yield - self.stride
|
116 |
+
|
117 |
+
def end(self):
|
118 |
+
"""Flushes any remaining cache and appends the stop symbol."""
|
119 |
+
if self.token_cache is not None:
|
120 |
+
audio_values = self.apply_delay_pattern_mask(self.token_cache)
|
121 |
else:
|
122 |
+
audio_values = np.zeros(self.to_yield)
|
123 |
+
|
124 |
+
self.on_finalized_audio(audio_values[self.to_yield :], stream_end=True)
|
125 |
+
|
126 |
+
def on_finalized_audio(self, audio: np.ndarray, stream_end: bool = False):
|
127 |
+
"""Put the new audio in the queue. If the stream is ending, also put a stop signal in the queue."""
|
128 |
+
self.audio_queue.put(audio, timeout=self.timeout)
|
129 |
+
if stream_end:
|
130 |
+
self.audio_queue.put(self.stop_signal, timeout=self.timeout)
|
131 |
+
|
132 |
+
def __iter__(self):
|
133 |
+
return self
|
134 |
+
|
135 |
+
def __next__(self):
|
136 |
+
value = self.audio_queue.get(timeout=self.timeout)
|
137 |
+
if not isinstance(value, np.ndarray) and value == self.stop_signal:
|
138 |
+
raise StopIteration()
|
139 |
+
else:
|
140 |
+
return value
|
141 |
+
|
142 |
+
|
143 |
+
sampling_rate = model.audio_encoder.config.sampling_rate
|
144 |
+
frame_rate = model.audio_encoder.config.frame_rate
|
145 |
+
|
146 |
+
target_dtype = np.int16
|
147 |
+
max_range = np.iinfo(target_dtype).max
|
148 |
+
|
149 |
+
|
150 |
+
@spaces.GPU
|
151 |
+
def generate_audio(text_prompt, audio_length_in_s=10.0, play_steps_in_s=2.0, seed=0):
|
152 |
+
max_new_tokens = int(frame_rate * audio_length_in_s)
|
153 |
+
play_steps = int(frame_rate * play_steps_in_s)
|
154 |
+
|
155 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
156 |
+
if device != model.device:
|
157 |
+
model.to(device)
|
158 |
+
if device == "cuda:0":
|
159 |
+
model.half()
|
160 |
+
|
161 |
+
inputs = processor(
|
162 |
+
text=text_prompt,
|
163 |
+
padding=True,
|
164 |
+
return_tensors="pt",
|
165 |
+
)
|
166 |
+
|
167 |
+
streamer = MusicgenStreamer(model, device=device, play_steps=play_steps)
|
168 |
+
|
169 |
+
generation_kwargs = dict(
|
170 |
+
**inputs.to(device),
|
171 |
+
streamer=streamer,
|
172 |
+
max_new_tokens=max_new_tokens,
|
173 |
+
)
|
174 |
+
thread = Thread(target=model.generate, kwargs=generation_kwargs)
|
175 |
+
thread.start()
|
176 |
+
|
177 |
+
set_seed(seed)
|
178 |
+
for new_audio in streamer:
|
179 |
+
print(f"Sample of length: {round(new_audio.shape[0] / sampling_rate, 2)} seconds")
|
180 |
+
new_audio = (new_audio * max_range).astype(np.int16)
|
181 |
+
yield (sampling_rate, new_audio)
|
182 |
+
|
183 |
+
|
184 |
+
demo = gr.Interface(
|
185 |
+
fn=generate_audio,
|
186 |
+
inputs=[
|
187 |
+
gr.Text(label="Prompt", value="80s pop track with synth and instrumentals"),
|
188 |
+
gr.Slider(10, 30, value=15, step=5, label="Audio length in seconds"),
|
189 |
+
gr.Slider(0.5, 2.5, value=0.5, step=0.5, label="Streaming interval in seconds", info="Lower = shorter chunks, lower latency, more codec steps"),
|
190 |
+
gr.Slider(0, 10, value=5, step=1, label="Seed for random generations"),
|
191 |
+
],
|
192 |
+
outputs=[
|
193 |
+
gr.Audio(label="Generated Music", streaming=True, autoplay=True)
|
194 |
+
],
|
195 |
+
examples = [
|
196 |
+
["Country acoustic guitar fast line dance singer like Kenny Chesney and Garth brooks and Luke Combs and Chris Stapleton. bpm: 100", 30, 0.5, 5],
|
197 |
+
["Electronic Dance track with pulsating bass and high energy synths. bpm: 126", 30, 0.5, 5],
|
198 |
+
["Rap Beats with deep bass and snappy snares. bpm: 80", 30, 0.5, 5],
|
199 |
+
["Lo-Fi track with smooth beats and chill vibes. bpm: 100", 30, 0.5, 5],
|
200 |
+
["Global Groove track with international instruments and dance rhythms. bpm: 128", 30, 0.5, 5],
|
201 |
+
["Relaxing Meditation music with ambient pads and soothing melodies. bpm: 80", 30, 0.5, 5],
|
202 |
+
["Rave Dance track with hard-hitting beats and euphoric synths. bpm: 128", 30, 0.5, 5]
|
203 |
+
],
|
204 |
+
|
205 |
+
title=title,
|
206 |
+
description=description,
|
207 |
+
article=article,
|
208 |
+
cache_examples=False,
|
209 |
)
|
210 |
|
211 |
+
demo.queue().launch()
|