File size: 16,488 Bytes
5238467
 
 
 
 
 
 
 
1897b6f
8e10a53
5238467
 
925b7f8
efabdc6
 
9138f15
1897b6f
b76a81b
1a6de5e
14af4d8
aef7fad
1897b6f
5238467
595ae94
1a6de5e
595ae94
 
 
5238467
1a6de5e
 
 
efabdc6
 
 
 
 
 
 
 
 
 
 
 
 
5238467
 
de8ae12
5238467
de8ae12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5238467
 
aef7fad
1a6de5e
ee1911a
1a6de5e
 
 
 
 
 
 
 
 
 
5238467
 
de8ae12
 
 
4f37585
 
 
 
 
5238467
e3f64dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1a6de5e
e3f64dd
 
 
 
 
 
 
 
1dda6b6
 
e3f64dd
5238467
e3f64dd
 
 
c81e4b7
e3f64dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14af4d8
e3f64dd
 
 
 
 
 
50d48cc
e3f64dd
 
5238467
1a6de5e
 
 
 
 
 
14af4d8
 
e3f64dd
 
 
 
 
 
 
d7ef5a5
 
b76a81b
 
d7ef5a5
 
b76a81b
 
d7ef5a5
b76a81b
d7ef5a5
 
 
 
 
 
e3f64dd
14af4d8
e83dc6d
14af4d8
 
 
1a6de5e
1897b6f
5d66b58
1a6de5e
5d66b58
23fe483
 
b76a81b
1a6de5e
de8ae12
 
 
 
 
 
aef7fad
5238467
de8ae12
 
 
 
 
 
 
0ffc43b
de8ae12
 
 
bedb522
0ffc43b
bedb522
 
de8ae12
 
bedb522
de8ae12
 
5238467
de8ae12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1dda6b6
50d48cc
de8ae12
 
 
1a6de5e
b76a81b
de8ae12
 
1a6de5e
 
1dda6b6
 
de8ae12
 
 
 
 
 
 
 
 
 
 
 
 
 
9766876
de8ae12
 
 
 
9766876
de8ae12
 
 
 
9766876
de8ae12
 
 
 
9766876
de8ae12
 
 
 
9766876
de8ae12
 
 
595ae94
de8ae12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
595ae94
de8ae12
 
 
 
 
 
 
 
 
 
8e10a53
de8ae12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
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
85
86
87
88
89
90
91
92
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
"""
Copyright (c) Meta Platforms, Inc. and affiliates.
All rights reserved.

This source code is licensed under the license found in the
LICENSE file in the root directory of this source tree.
"""

from tempfile import NamedTemporaryFile
import argparse
import torch
import gradio as gr
import os
import time
import warnings
from audiocraft.models import MusicGen
from audiocraft.data.audio import audio_write
from audiocraft.data.audio_utils import apply_fade, apply_tafade
from audiocraft.utils.extend import generate_music_segments, add_settings_to_image, INTERRUPTING
import numpy as np
import random

MODEL = None
MODELS = None
IS_SHARED_SPACE = "Surn/UnlimitedMusicGen" in os.environ.get('SPACE_ID', '')
INTERRUPTED = False
UNLOAD_MODEL = False
MOVE_TO_CPU = False

def interrupt_callback():
    return INTERRUPTED

def interrupt():
    global INTERRUPTING
    INTERRUPTING = True


def make_waveform(*args, **kwargs):
    # Further remove some warnings.
    be = time.time()
    with warnings.catch_warnings():
        warnings.simplefilter('ignore')
        out = gr.make_waveform(*args, **kwargs)
        print("Make a video took", time.time() - be)
        return out

def load_model(version):
    global MODEL, MODELS, UNLOAD_MODEL
    print("Loading model", version)
    if MODELS is None:
        return MusicGen.get_pretrained(version)
    else:
        t1 = time.monotonic()
        if MODEL is not None:
            MODEL.to('cpu') # move to cache
            print("Previous model moved to CPU in %.2fs" % (time.monotonic() - t1))
            t1 = time.monotonic()
        if MODELS.get(version) is None:
            print("Loading model %s from disk" % version)
            result = MusicGen.get_pretrained(version)
            MODELS[version] = result
            print("Model loaded in %.2fs" % (time.monotonic() - t1))
            return result
        result = MODELS[version].to('cuda')
        print("Cached model loaded in %.2fs" % (time.monotonic() - t1))
        return result


def predict(model, text, melody, duration, dimension, topk, topp, temperature, cfg_coef, background, title, include_settings, settings_font, settings_font_color, seed, overlap=1):
    global MODEL, INTERRUPTED, INTERRUPTING
    output_segments = None

    INTERRUPTED = False
    INTERRUPTING = False
    if temperature < 0:
        raise gr.Error("Temperature must be >= 0.")
    if topk < 0:
        raise gr.Error("Topk must be non-negative.")
    if topp < 0:
        raise gr.Error("Topp must be non-negative.")

    if MODEL is None or MODEL.name != model:
        MODEL = load_model(model)
    else:
        if MOVE_TO_CPU:
            MODEL.to('cuda')
    
    # prevent hacking
    duration = min(duration, 720)
    overlap =  min(overlap, 15)
    # 

    output = None
    segment_duration = duration
    initial_duration = duration
    output_segments = []
    while duration > 0:
        if not output_segments: # first pass of long or short song
            if segment_duration > MODEL.lm.cfg.dataset.segment_duration:
                segment_duration = MODEL.lm.cfg.dataset.segment_duration
            else:
                segment_duration = duration
        else: # next pass of long song
            if duration + overlap < MODEL.lm.cfg.dataset.segment_duration:
                segment_duration = duration + overlap
            else:
                segment_duration = MODEL.lm.cfg.dataset.segment_duration
        # implement seed
        if seed < 0:
            seed = random.randint(0, 0xffff_ffff_ffff)
        torch.manual_seed(seed)


        print(f'Segment duration: {segment_duration}, duration: {duration}, overlap: {overlap}')
        MODEL.set_generation_params(
            use_sampling=True,
            top_k=topk,
            top_p=topp,
            temperature=temperature,
            cfg_coef=cfg_coef,
            duration=segment_duration,
            two_step_cfg=False,
            rep_penalty=0.5
        )

        if melody:
            # todo return excess duration, load next model and continue in loop structure building up output_segments
            if duration > MODEL.lm.cfg.dataset.segment_duration:
                output_segments, duration = generate_music_segments(text, melody, seed, MODEL, duration, overlap, MODEL.lm.cfg.dataset.segment_duration)
            else:
                # pure original code
                sr, melody = melody[0], torch.from_numpy(melody[1]).to(MODEL.device).float().t().unsqueeze(0)
                print(melody.shape)
                if melody.dim() == 2:
                    melody = melody[None]
                melody = melody[..., :int(sr * MODEL.lm.cfg.dataset.segment_duration)]
                output = MODEL.generate_with_chroma(
                    descriptions=[text],
                    melody_wavs=melody,
                    melody_sample_rate=sr,
                    progress=True
                )
            # All output_segments are populated, so we can break the loop or set duration to 0
            break
        else:
            #output = MODEL.generate(descriptions=[text], progress=False)
            if not output_segments:
                next_segment = MODEL.generate(descriptions=[text], progress=True)
                duration -= segment_duration
            else:
                last_chunk = output_segments[-1][:, :, -overlap*MODEL.sample_rate:]
                next_segment = MODEL.generate_continuation(last_chunk, MODEL.sample_rate, descriptions=[text], progress=False)
                duration -= segment_duration - overlap
            output_segments.append(next_segment)

        if INTERRUPTING:
            INTERRUPTED = True
            INTERRUPTING = False
            print("Function execution interrupted!")
            raise gr.Error("Interrupted.")

    if output_segments:
        try:
            # Combine the output segments into one long audio file or stack tracks
            #output_segments = [segment.detach().cpu().float()[0] for segment in output_segments]
            #output = torch.cat(output_segments, dim=dimension)
            
            output = output_segments[0]
            for i in range(1, len(output_segments)):
                overlap_samples = overlap * MODEL.sample_rate
                #stack tracks and fade out/in
                overlapping_output_fadeout = output[:, :, -overlap_samples:]
                overlapping_output_fadeout = apply_fade(overlapping_output_fadeout,sample_rate=MODEL.sample_rate,duration=overlap,out=True,start=True, curve_end=0.0, current_device=MODEL.device)
                #overlapping_output_fadeout = apply_tafade(overlapping_output_fadeout,sample_rate=MODEL.sample_rate,duration=overlap,out=True,start=True,shape="exponential")

                overlapping_output_fadein = output_segments[i][:, :, :overlap_samples]
                overlapping_output_fadein = apply_fade(overlapping_output_fadein,sample_rate=MODEL.sample_rate,duration=overlap,out=False,start=False, curve_start=0.0, current_device=MODEL.device)
                #overlapping_output_fadein = apply_tafade(overlapping_output_fadein,sample_rate=MODEL.sample_rate,duration=overlap,out=False,start=False, shape="linear")

                overlapping_output = torch.cat([overlapping_output_fadeout[:, :, :-(overlap_samples // 2)], overlapping_output_fadein],dim=2)
                print(f" overlap size Fade:{overlapping_output.size()}\n output: {output.size()}\n segment: {output_segments[i].size()}")
                ##overlapping_output = torch.cat([output[:, :, -overlap_samples:], output_segments[i][:, :, :overlap_samples]], dim=1) #stack tracks
                ##print(f" overlap size stack:{overlapping_output.size()}\n output: {output.size()}\n segment: {output_segments[i].size()}")
                #overlapping_output = torch.cat([output[:, :, -overlap_samples:], output_segments[i][:, :, :overlap_samples]], dim=2) #stack tracks
                #print(f" overlap size cat:{overlapping_output.size()}\n output: {output.size()}\n segment: {output_segments[i].size()}")               
                output = torch.cat([output[:, :, :-overlap_samples], overlapping_output, output_segments[i][:, :, overlap_samples:]], dim=dimension)
            output = output.detach().cpu().float()[0]
        except Exception as e:
            print(f"Error combining segments: {e}. Using the first segment only.")
            output = output_segments[0].detach().cpu().float()[0]
    else:
        output = output.detach().cpu().float()[0]

    with NamedTemporaryFile("wb", suffix=".wav", delete=False) as file:
        if include_settings:
            video_description = f"{text}\n Duration: {str(initial_duration)} Dimension: {dimension}\n Top-k:{topk} Top-p:{topp}\n Randomness:{temperature}\n cfg:{cfg_coef} overlap: {overlap}\n Seed: {seed}\n Model: {model}\n Melody File:#todo"
            background = add_settings_to_image(title, video_description, background_path=background, font=settings_font, font_color=settings_font_color)
        audio_write(
            file.name, output, MODEL.sample_rate, strategy="loudness",
            loudness_headroom_db=18, loudness_compressor=True, add_suffix=False, channels=2)
        waveform_video = make_waveform(file.name,bg_image=background, bar_count=45)
    if MOVE_TO_CPU:
        MODEL.to('cpu')
    if UNLOAD_MODEL:
        MODEL = None
    torch.cuda.empty_cache()
    torch.cuda.ipc_collect()
    return waveform_video, seed

def ui(**kwargs):
    css="""
    #col-container {max-width: 910px; margin-left: auto; margin-right: auto;}
    a {text-decoration-line: underline; font-weight: 600;}
    """
    with gr.Blocks(title="UnlimitedMusicGen", css=css) as demo:
        gr.Markdown(
            """            
            # UnlimitedMusicGen
            This is your private demo for [UnlimitedMusicGen](https://github.com/Oncorporation/audiocraft), a simple and controllable model for music generation
            presented at: ["Simple and Controllable Music Generation"](https://huggingface.co/papers/2306.05284)
            
            Disclaimer: This won't run on CPU only. Clone this App and run on GPU instance!
                        
            Todo: Working on improved Melody Conditioned Music Generation transitions and consistency.
            """
        )
        if IS_SHARED_SPACE and not torch.cuda.is_available():
            gr.Markdown("""
                ⚠ This Space doesn't work in this shared UI ⚠

                <a href="https://huggingface.co/spaces/musicgen/MusicGen?duplicate=true" style="display: inline-block;margin-top: .5em;margin-right: .25em;" target="_blank">
                <img style="margin-bottom: 0em;display: inline;margin-top: -.25em;" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
                to use it privately, or use the <a href="https://huggingface.co/spaces/facebook/MusicGen">public demo</a>
                """)
        with gr.Row():
            with gr.Column():
                with gr.Row():
                    text = gr.Text(label="Input Text", interactive=True, value="4/4 100bpm 320kbps 48khz, Industrial/Electronic Soundtrack, Dark, Intense, Sci-Fi")
                    melody = gr.Audio(source="upload", type="numpy", label="Melody Condition (optional)", interactive=True)
                with gr.Row():
                    submit = gr.Button("Submit")
                    # Adapted from https://github.com/rkfg/audiocraft/blob/long/app.py, MIT license.
                    _ = gr.Button("Interrupt").click(fn=interrupt, queue=False)
                with gr.Row():
                    background= gr.Image(value="./assets/background.png", source="upload", label="Background", shape=(768,512), type="filepath", interactive=True)
                    include_settings = gr.Checkbox(label="Add Settings to background", value=True, interactive=True)
                with gr.Row():
                    title = gr.Textbox(label="Title", value="UnlimitedMusicGen", interactive=True)
                    settings_font = gr.Text(label="Settings Font", value="./assets/arial.ttf", interactive=True)
                    settings_font_color = gr.ColorPicker(label="Settings Font Color", value="#c87f05", interactive=True)
                with gr.Row():
                    model = gr.Radio(["melody", "medium", "small", "large"], label="Model", value="melody", interactive=True)
                with gr.Row():
                    duration = gr.Slider(minimum=1, maximum=720, value=10, label="Duration", interactive=True)
                    overlap = gr.Slider(minimum=1, maximum=15, value=5, step=1, label="Overlap", interactive=True)
                    dimension = gr.Slider(minimum=-2, maximum=2, value=2, step=1, label="Dimension", info="determines which direction to add new segements of audio. (1 = stack tracks, 2 = lengthen, -2..0 = ?)", interactive=True)
                with gr.Row():
                    topk = gr.Number(label="Top-k", value=250, precision=0, interactive=True)
                    topp = gr.Number(label="Top-p", value=0, precision=0, interactive=True)
                    temperature = gr.Number(label="Randomness Temperature", value=0.75, precision=None, interactive=True)
                    cfg_coef = gr.Number(label="Classifier Free Guidance", value=5.5, precision=None, interactive=True)
                with gr.Row():
                    seed = gr.Number(label="Seed", value=-1, precision=0, interactive=True)
                    gr.Button('\U0001f3b2\ufe0f').style(full_width=False).click(fn=lambda: -1, outputs=[seed], queue=False)
                    reuse_seed = gr.Button('\u267b\ufe0f').style(full_width=False)
            with gr.Column() as c:
                output = gr.Video(label="Generated Music")
                seed_used = gr.Number(label='Seed used', value=-1, interactive=False)

        reuse_seed.click(fn=lambda x: x, inputs=[seed_used], outputs=[seed], queue=False)
        submit.click(predict, inputs=[model, text, melody, duration, dimension, topk, topp, temperature, cfg_coef, background, title, include_settings, settings_font, settings_font_color, seed, overlap], outputs=[output, seed_used])
        gr.Examples(
            fn=predict,
            examples=[
                [
                    "4/4 120bpm 320kbps 48khz, An 80s driving pop song with heavy drums and synth pads in the background",
                    "./assets/bach.mp3",
                    "melody"
                ],
                [
                    "4/4 120bpm 320kbps 48khz, A cheerful country song with acoustic guitars",
                    "./assets/bolero_ravel.mp3",
                    "melody"
                ],
                [
                    "4/4 120bpm 320kbps 48khz, 90s rock song with electric guitar and heavy drums",
                    None,
                    "medium"
                ],
                [
                    "4/4 120bpm 320kbps 48khz, a light and cheerly EDM track, with syncopated drums, aery pads, and strong emotions",
                    "./assets/bach.mp3",
                    "melody"
                ],
                [
                    "4/4 320kbps 48khz, lofi slow bpm electro chill with organic samples",
                    None,
                    "medium",
                ],
            ],
            inputs=[text, melody, model],
            outputs=[output]
        )

        # Show the interface
        launch_kwargs = {}
        share = kwargs.get('share', False)
        if share:
            launch_kwargs['share'] = share



        demo.queue(max_size=15).launch(**launch_kwargs )

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    
    parser.add_argument(
        '--share', action='store_true', help='Share the gradio UI'
    )
    parser.add_argument(
        '--unload_model', action='store_true', help='Unload the model after every generation to save GPU memory'
    )

    parser.add_argument(
        '--unload_to_cpu', action='store_true', help='Move the model to main RAM after every generation to save GPU memory but reload faster than after full unload (see above)'
    )

    parser.add_argument(
        '--cache', action='store_true', help='Cache models in RAM to quickly switch between them'
    )

    args = parser.parse_args()
    UNLOAD_MODEL = args.unload_model
    MOVE_TO_CPU = args.unload_to_cpu
    if args.cache:
        MODELS = {}

    ui(
        unload_to_cpu = MOVE_TO_CPU,
        share=args.share
        
    )