File size: 13,194 Bytes
3888ab7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import glob
import os.path

import gradio as gr

import tqdm
import json

import MIDI
from midi_synthesizer import synthesis

in_space = os.getenv("SYSTEM") == "spaces"


def generate(model, prompt=None, max_len=512, temp=1.0, top_p=0.98, top_k=20,
             disable_patch_change=False, disable_control_change=False, disable_channels=None):
    if disable_channels is not None:
        disable_channels = [tokenizer.parameter_ids["channel"][c] for c in disable_channels]
    else:
        disable_channels = []
    max_token_seq = tokenizer.max_token_seq
    if prompt is None:
        input_tensor = np.full((1, max_token_seq), tokenizer.pad_id, dtype=np.int64)
        input_tensor[0, 0] = tokenizer.bos_id  # bos
    else:
        prompt = prompt[:, :max_token_seq]
        if prompt.shape[-1] < max_token_seq:
            prompt = np.pad(prompt, ((0, 0), (0, max_token_seq - prompt.shape[-1])),
                            mode="constant", constant_values=tokenizer.pad_id)
        input_tensor = prompt
    input_tensor = input_tensor[None, :, :]
    cur_len = input_tensor.shape[1]
    bar = tqdm.tqdm(desc="generating", total=max_len - cur_len, disable=in_space)
    with bar:
        while cur_len < max_len:
            end = False
            hidden = model[0].run(None, {'x': input_tensor})[0][:, -1]
            next_token_seq = np.empty((1, 0), dtype=np.int64)
            event_name = ""
            for i in range(max_token_seq):
                mask = np.zeros(tokenizer.vocab_size, dtype=np.int64)
                if i == 0:
                    mask_ids = list(tokenizer.event_ids.values()) + [tokenizer.eos_id]
                    if disable_patch_change:
                        mask_ids.remove(tokenizer.event_ids["patch_change"])
                    if disable_control_change:
                        mask_ids.remove(tokenizer.event_ids["control_change"])
                    mask[mask_ids] = 1
                else:
                    param_name = tokenizer.events[event_name][i - 1]
                    mask_ids = tokenizer.parameter_ids[param_name]
                    if param_name == "channel":
                        mask_ids = [i for i in mask_ids if i not in disable_channels]
                    mask[mask_ids] = 1
                logits = model[1].run(None, {'x': next_token_seq, "hidden": hidden})[0][:, -1:]
                scores = softmax(logits / temp, -1) * mask
                sample = sample_top_p_k(scores, top_p, top_k)
                if i == 0:
                    next_token_seq = sample
                    eid = sample.item()
                    if eid == tokenizer.eos_id:
                        end = True
                        break
                    event_name = tokenizer.id_events[eid]
                else:
                    next_token_seq = np.concatenate([next_token_seq, sample], axis=1)
                    if len(tokenizer.events[event_name]) == i:
                        break
            if next_token_seq.shape[1] < max_token_seq:
                next_token_seq = np.pad(next_token_seq, ((0, 0), (0, max_token_seq - next_token_seq.shape[-1])),
                                        mode="constant", constant_values=tokenizer.pad_id)
            next_token_seq = next_token_seq[None, :, :]
            input_tensor = np.concatenate([input_tensor, next_token_seq], axis=1)
            cur_len += 1
            bar.update(1)
            yield next_token_seq.reshape(-1)
            if end:
                break


def create_msg(name, data):
    return {"name": name, "data": data}


def run(model_name, tab, instruments, drum_kit, mid, midi_events, gen_events, temp, top_p, top_k, allow_cc):
    mid_seq = []
    gen_events = int(gen_events)
    max_len = gen_events

    disable_patch_change = False
    disable_channels = None
    if tab == 0:
        i = 0
        mid = [[tokenizer.bos_id] + [tokenizer.pad_id] * (tokenizer.max_token_seq - 1)]
        patches = {}
        for instr in instruments:
            patches[i] = patch2number[instr]
            i = (i + 1) if i != 8 else 10
        if drum_kit != "None":
            patches[9] = drum_kits2number[drum_kit]
        for i, (c, p) in enumerate(patches.items()):
            mid.append(tokenizer.event2tokens(["patch_change", 0, 0, i, c, p]))
        mid_seq = mid
        mid = np.asarray(mid, dtype=np.int64)
        if len(instruments) > 0:
            disable_patch_change = True
            disable_channels = [i for i in range(16) if i not in patches]
    elif mid is not None:
        mid = tokenizer.tokenize(MIDI.midi2score(mid))
        mid = np.asarray(mid, dtype=np.int64)
        mid = mid[:int(midi_events)]
        max_len += len(mid)
        for token_seq in mid:
            mid_seq.append(token_seq.tolist())
    init_msgs = [create_msg("visualizer_clear", None)]
    for tokens in mid_seq:
        init_msgs.append(create_msg("visualizer_append", tokenizer.tokens2event(tokens)))
    yield mid_seq, None, None, init_msgs
    model = models[model_name]
    generator = generate(model, mid, max_len=max_len, temp=temp, top_p=top_p, top_k=top_k,
                         disable_patch_change=disable_patch_change, disable_control_change=not allow_cc,
                         disable_channels=disable_channels)
    for i, token_seq in enumerate(generator):
        token_seq = token_seq.tolist()
        mid_seq.append(token_seq)
        event = tokenizer.tokens2event(token_seq)
        yield mid_seq, None, None, [create_msg("visualizer_append", event), create_msg("progress", [i + 1, gen_events])]
    mid = tokenizer.detokenize(mid_seq)
    with open(f"output.mid", 'wb') as f:
        f.write(MIDI.score2midi(mid))
    audio = synthesis(MIDI.score2opus(mid), soundfont_path)
    yield mid_seq, "output.mid", (44100, audio), [create_msg("visualizer_end", None)]


def cancel_run(mid_seq):
    if mid_seq is None:
        return None, None
    mid = tokenizer.detokenize(mid_seq)
    with open(f"output.mid", 'wb') as f:
        f.write(MIDI.score2midi(mid))
    audio = synthesis(MIDI.score2opus(mid), soundfont_path)
    return "output.mid", (44100, audio), [create_msg("visualizer_end", None)]


def load_javascript(dir="javascript"):
    scripts_list = glob.glob(f"{dir}/*.js")
    javascript = ""
    for path in scripts_list:
        with open(path, "r", encoding="utf8") as jsfile:
            javascript += f"\n<!-- {path} --><script>{jsfile.read()}</script>"
    template_response_ori = gr.routes.templates.TemplateResponse

    def template_response(*args, **kwargs):
        res = template_response_ori(*args, **kwargs)
        res.body = res.body.replace(
            b'</head>', f'{javascript}</head>'.encode("utf8"))
        res.init_headers()
        return res

    gr.routes.templates.TemplateResponse = template_response


class JSMsgReceiver(gr.HTML):

    def __init__(self, **kwargs):
        super().__init__(elem_id="msg_receiver", visible=False, **kwargs)

    def postprocess(self, y):
        if y:
            y = f"<p>{json.dumps(y)}</p>"
        return super().postprocess(y)

    def get_block_name(self) -> str:
        return "html"


number2drum_kits = {-1: "None", 0: "Standard", 8: "Room", 16: "Power", 24: "Electric", 25: "TR-808", 32: "Jazz",
                    40: "Blush", 48: "Orchestra"}
patch2number = {v: k for k, v in MIDI.Number2patch.items()}
drum_kits2number = {v: k for k, v in number2drum_kits.items()}

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
    parser.add_argument("--port", type=int, default=7860, help="gradio server port")
    parser.add_argument("--max-gen", type=int, default=1024, help="max")
    opt = parser.parse_args()
    soundfont_path = hf_hub_download(repo_id="skytnt/midi-model", filename="soundfont.sf2")
    models_info = {"generic pretrain model": ["skytnt/midi-model", ""],
                   "j-pop finetune model": ["skytnt/midi-model-ft", "jpop/"],
                   "touhou finetune model": ["skytnt/midi-model-ft", "touhou/"]}
    models = {}
    tokenizer = MIDITokenizer()
    providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
    for name, (repo_id, path) in models_info.items():
        model_base_path = hf_hub_download(repo_id=repo_id, filename=f"{path}onnx/model_base.onnx")
        model_token_path = hf_hub_download(repo_id=repo_id, filename=f"{path}onnx/model_token.onnx")
        model_base = rt.InferenceSession(model_base_path, providers=providers)
        model_token = rt.InferenceSession(model_token_path, providers=providers)
        models[name] = [model_base, model_token]

    load_javascript()
    app = gr.Blocks()
    with app:
        gr.Markdown("<h1 style='text-align: center; margin-bottom: 1rem'>Midi Composer</h1>")
        gr.Markdown("![Visitors](https://api.visitorbadge.io/api/visitors?path=skytnt.midi-composer&style=flat)\n\n"
                    "Midi event transformer for music generation\n\n"
                    "Demo for [SkyTNT/midi-model](https://github.com/SkyTNT/midi-model)\n\n"
                    "[Open In Colab]"
                    "(https://colab.research.google.com/github/SkyTNT/midi-model/blob/main/demo.ipynb)"
                    " for faster running and longer generation"
                    )
        js_msg = JSMsgReceiver()
        input_model = gr.Dropdown(label="select model", choices=list(models.keys()),
                                  type="value", value=list(models.keys())[0])
        tab_select = gr.Variable(value=0)
        with gr.Tabs():
            with gr.TabItem("instrument prompt") as tab1:
                input_instruments = gr.Dropdown(label="instruments (auto if empty)", choices=list(patch2number.keys()),
                                                multiselect=True, max_choices=15, type="value")
                input_drum_kit = gr.Dropdown(label="drum kit", choices=list(drum_kits2number.keys()), type="value",
                                             value="None")
                example1 = gr.Examples([
                    [[], "None"],
                    [["Acoustic Grand"], "None"],
                    [["Acoustic Grand", "Violin", "Viola", "Cello", "Contrabass"], "Orchestra"],
                    [["Flute", "Cello", "Bassoon", "Tuba"], "None"],
                    [["Violin", "Viola", "Cello", "Contrabass", "Trumpet", "French Horn", "Brass Section",
                      "Flute", "Piccolo", "Tuba", "Trombone", "Timpani"], "Orchestra"],
                    [["Acoustic Guitar(nylon)", "Acoustic Guitar(steel)", "Electric Guitar(jazz)",
                      "Electric Guitar(clean)", "Electric Guitar(muted)", "Overdriven Guitar", "Distortion Guitar",
                      "Electric Bass(finger)"], "Standard"]
                ], [input_instruments, input_drum_kit])
            with gr.TabItem("midi prompt") as tab2:
                input_midi = gr.File(label="input midi", file_types=[".midi", ".mid"], type="binary")
                input_midi_events = gr.Slider(label="use first n midi events as prompt", minimum=1, maximum=512,
                                              step=1,
                                              value=128)
                example2 = gr.Examples([[file, 128] for file in glob.glob("example/*.mid")],
                                       [input_midi, input_midi_events])

        tab1.select(lambda: 0, None, tab_select, queue=False)
        tab2.select(lambda: 1, None, tab_select, queue=False)
        input_gen_events = gr.Slider(label="generate n midi events", minimum=1, maximum=opt.max_gen,
                                     step=1, value=opt.max_gen // 2)
        with gr.Accordion("options", open=False):
            input_temp = gr.Slider(label="temperature", minimum=0.1, maximum=1.2, step=0.01, value=1)
            input_top_p = gr.Slider(label="top p", minimum=0.1, maximum=1, step=0.01, value=0.98)
            input_top_k = gr.Slider(label="top k", minimum=1, maximum=20, step=1, value=12)
            input_allow_cc = gr.Checkbox(label="allow midi cc event", value=True)
            example3 = gr.Examples([[1, 0.98, 12], [1.2, 0.95, 8]], [input_temp, input_top_p, input_top_k])
        run_btn = gr.Button("generate", variant="primary")
        stop_btn = gr.Button("stop and output")
        output_midi_seq = gr.Variable()
        output_midi_visualizer = gr.HTML(elem_id="midi_visualizer_container")
        output_audio = gr.Audio(label="output audio", format="mp3", elem_id="midi_audio")
        output_midi = gr.File(label="output midi", file_types=[".mid"])
        run_event = run_btn.click(run, [input_model, tab_select, input_instruments, input_drum_kit, input_midi,
                                        input_midi_events, input_gen_events, input_temp, input_top_p, input_top_k,
                                        input_allow_cc],
                                  [output_midi_seq, output_midi, output_audio, js_msg])
        stop_btn.click(cancel_run, output_midi_seq, [output_midi, output_audio, js_msg], cancels=run_event, queue=False)
    app.queue(2).launch(server_port=opt.port, share=opt.share, inbrowser=True)