File size: 10,926 Bytes
1ea42dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ff5739c
1ea42dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import glob

import PIL
import gradio as gr
import numpy as np
import torch

import torch.nn.functional as F
import tqdm

import MIDI
from midi_model import MIDIModel
from midi_tokenizer import MIDITokenizer
from midi_synthesizer import synthesis
from huggingface_hub import hf_hub_download

@torch.inference_mode()
def generate(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, amp=True):
    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 = torch.full((1, max_token_seq), tokenizer.pad_id, dtype=torch.long, device=model.device)
        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 = torch.from_numpy(prompt).to(dtype=torch.long, device=model.device)
    input_tensor = input_tensor.unsqueeze(0)
    cur_len = input_tensor.shape[1]
    bar = tqdm.tqdm(desc="generating", total=max_len - cur_len)
    with bar, torch.cuda.amp.autocast(enabled=amp):
        while cur_len < max_len:
            end = False
            hidden = model.forward(input_tensor)[0, -1].unsqueeze(0)
            next_token_seq = None
            event_name = ""
            for i in range(max_token_seq):
                mask = torch.zeros(tokenizer.vocab_size, dtype=torch.int64, device=model.device)
                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.forward_token(hidden, next_token_seq)[:, -1:]
                scores = torch.softmax(logits / temp, dim=-1) * mask
                sample = model.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 = torch.cat([next_token_seq, sample], dim=1)
                    if len(tokenizer.events[event_name]) == i:
                        break
            if next_token_seq.shape[1] < max_token_seq:
                next_token_seq = F.pad(next_token_seq, (0, max_token_seq - next_token_seq.shape[1]),
                                       "constant", value=tokenizer.pad_id)
            next_token_seq = next_token_seq.unsqueeze(1)
            input_tensor = torch.cat([input_tensor, next_token_seq], dim=1)
            cur_len += 1
            bar.update(1)
            yield next_token_seq.reshape(-1).cpu().numpy()
            if end:
                break


def run(tab, instruments, drum_kit, mid, midi_events, gen_events, temp, top_p, top_k, allow_cc, amp):
    mid_seq = []
    max_len = int(gen_events)
    img_len = 1024
    img = np.full((128 * 2, img_len, 3), 255, dtype=np.uint8)
    state = {"t1": 0, "t": 0, "cur_pos": 0}
    rand = np.random.RandomState(0)
    colors = {(i, j): rand.randint(0, 200, 3) for i in range(128) for j in range(16)}

    def draw_event(tokens):
        if tokens[0] in tokenizer.id_events:
            name = tokenizer.id_events[tokens[0]]
            if len(tokens) <= len(tokenizer.events[name]):
                return
            params = tokens[1:]
            params = [params[i] - tokenizer.parameter_ids[p][0] for i, p in enumerate(tokenizer.events[name])]
            if not all([0 <= params[i] < tokenizer.event_parameters[p] for i, p in enumerate(tokenizer.events[name])]):
                return
            event = [name] + params
            state["t1"] += event[1]
            t = state["t1"] * 16 + event[2]
            state["t"] = t
            if name == "note":
                tr, d, c, p = event[3:7]
                shift = t + d - (state["cur_pos"] + img_len)
                if shift > 0:
                    img[:, :-shift] = img[:, shift:]
                    img[:, -shift:] = 255
                    state["cur_pos"] += shift
                t = t - state["cur_pos"]
                img[p * 2:(p + 1) * 2, t: t + d] = colors[(tr, c)]

    def get_img():
        t = state["t"] - state["cur_pos"]
        img_new = img.copy()
        img_new[:, t: t + 2] = 0
        return PIL.Image.fromarray(np.flip(img_new, 0))

    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 != 9 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 or drum_kit != "None":
            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)
            draw_event(token_seq)
    generator = generate(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, amp=amp)
    for token_seq in generator:
        mid_seq.append(token_seq)
        draw_event(token_seq)
        yield mid_seq, get_img(), 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)
    yield mid_seq, get_img(), "output.mid", (44100, audio)


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)


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("--device", type=str, default="cpu", help="device to run model")
    soundfont_path = hf_hub_download(repo_id="skytnt/midi-model", filename="soundfont.sf2")
    model_path = hf_hub_download(repo_id="skytnt/midi-model", filename="model.ckpt")
    opt = parser.parse_args()
    tokenizer = MIDITokenizer()
    model = MIDIModel(tokenizer).to(device=opt.device)
    ckpt = torch.load(model_path, map_location="cpu")
    state_dict = ckpt.get("state_dict", ckpt)
    model.load_state_dict(state_dict, strict=False)
    model.eval()

    app = gr.Blocks()
    with app:
        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=10, type="value")
                input_drum_kit = gr.Dropdown(label="drum kit", choices=list(drum_kits2number.keys()), type="value",
                                             value="None")
            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)

        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=512, step=1, value=512)
        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=0.97)
        input_top_k = gr.Slider(label="top k", minimum=1, maximum=50, step=1, value=20)
        input_allow_cc = gr.Checkbox(label="allow control change event", value=True)
        input_amp = gr.Checkbox(label="enable amp", value=True)
        run_btn = gr.Button("generate", variant="primary")
        stop_btn = gr.Button("stop")
        output_midi_seq = gr.Variable()
        output_midi_img = gr.Image(label="output image")
        output_midi = gr.File(label="output midi", file_types=[".mid"])
        output_audio = gr.Audio(label="output audio", format="mp3")
        run_event = run_btn.click(run, [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, input_amp],
                                  [output_midi_seq, output_midi_img, output_midi, output_audio])
        stop_btn.click(cancel_run, output_midi_seq, [output_midi, output_audio], cancels=run_event, queue=False)
    app.queue(1).launch(server_port=opt.port, share=opt.share)