Spaces:
Running
on
Zero
Running
on
Zero
File size: 27,408 Bytes
8fdb145 bb51a02 1ea42dc cfa70b3 fd012a7 9958d06 7406325 9868b75 1ea42dc fd012a7 1ea42dc 9d7618f 1ea42dc fd012a7 239bed9 ed10990 1f0da43 a3e7293 9868b75 fd012a7 294c6ec 9d7618f fd012a7 1ea42dc fd012a7 1ea42dc 294c6ec 1ea42dc 294c6ec 1ea42dc 294c6ec 1ea42dc fd012a7 1ea42dc fd012a7 9d7618f 1ea42dc 294c6ec fd012a7 294c6ec 1ea42dc 294c6ec fd012a7 294c6ec 1ea42dc 294c6ec 1ea42dc 294c6ec 1ea42dc fd012a7 1ea42dc 294c6ec 1ea42dc 6481a43 9958d06 1ea42dc 9958d06 5825808 294c6ec b3f8835 294c6ec c0d7f61 a021cba a158362 a021cba b3f8835 c0d7f61 8d6cb47 c0d7f61 294c6ec 183a87e 2b86b7e fd012a7 c51a1c9 239bed9 6481a43 1f0da43 bb51a02 fd012a7 1ea42dc 239bed9 c51a1c9 cfa70b3 1ea42dc 743aa2c 1ea42dc 1752002 1ea42dc 239bed9 294c6ec ed10990 1ea42dc cfa70b3 d5353b5 1ea42dc 294c6ec cfa70b3 294c6ec cfa70b3 256bea9 294c6ec cfa70b3 294c6ec cfa70b3 256bea9 294c6ec 9958d06 1ea42dc 294c6ec 1ea42dc 1fd2f8b 183a87e 294c6ec 1fd2f8b 294c6ec 256bea9 1fd2f8b 183a87e 1fd2f8b fd012a7 256bea9 183a87e 294c6ec 256bea9 6481a43 e9642f9 6481a43 1b97a00 98bc719 c51a1c9 98bc719 3fe9868 1ea42dc 239bed9 1ea42dc fd012a7 e9642f9 a52dad5 1ea42dc e9642f9 98bc719 1fd2f8b 239bed9 294c6ec 183a87e 294c6ec 573d12d fd012a7 5c45beb fd012a7 183a87e fd012a7 8fdb145 fd012a7 183a87e 94ac77e 183a87e 1ea42dc 6481a43 1ea42dc 22dc587 7406325 9438ab2 942f170 38cc15f ed10990 1b97a00 f88cef7 743aa2c 1ea42dc 256bea9 c51a1c9 ed10990 c51a1c9 1ea42dc c51a1c9 3fe9868 239bed9 ed10990 c51a1c9 fe820fd ed10990 1ea42dc d5353b5 3fe9868 d5353b5 ed10990 cfa70b3 294c6ec 256bea9 1ea42dc cfa70b3 1f0da43 a52dad5 d660a99 94ac77e a23f2ef d660a99 1fd2f8b a23f2ef d25994f 1ea42dc cd46b84 743aa2c 256bea9 1fd2f8b 294c6ec 1fd2f8b 256bea9 294c6ec 239bed9 cfa70b3 294c6ec 1fd2f8b cd46b84 fd012a7 256bea9 c0d7f61 1fd2f8b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 |
import spaces
import random
import argparse
import glob
import json
import os
import time
from concurrent.futures import ThreadPoolExecutor
import gradio as gr
import numpy as np
import torch
import torch.nn.functional as F
import tqdm
from huggingface_hub import hf_hub_download
import MIDI
from midi_model import MIDIModel, MIDIModelConfig
from midi_synthesizer import MidiSynthesizer
MAX_SEED = np.iinfo(np.int32).max
in_space = os.getenv("SYSTEM") == "spaces"
@torch.inference_mode()
def generate(model: MIDIModel, prompt=None, batch_size=1, max_len=512, temp=1.0, top_p=0.98, top_k=20,
disable_patch_change=False, disable_control_change=False, disable_channels=None, generator=None):
tokenizer = model.tokenizer
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
input_tensor = input_tensor.unsqueeze(0)
input_tensor = torch.cat([input_tensor] * batch_size, dim=0)
else:
if len(prompt.shape) == 2:
prompt = prompt[None, :]
prompt = np.repeat(prompt, repeats=batch_size, axis=0)
elif prompt.shape[0] == 1:
prompt = np.repeat(prompt, repeats=batch_size, axis=0)
elif len(prompt.shape) != 3 or prompt.shape[0] != batch_size:
raise ValueError(f"invalid shape for prompt, {prompt.shape}")
prompt = prompt[..., :max_token_seq]
if prompt.shape[-1] < max_token_seq:
prompt = np.pad(prompt, ((0, 0), (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)
cur_len = input_tensor.shape[1]
bar = tqdm.tqdm(desc="generating", total=max_len - cur_len)
with bar:
while cur_len < max_len:
end = [False] * batch_size
hidden = model.forward(input_tensor)[:, -1]
next_token_seq = None
event_names = [""] * batch_size
for i in range(max_token_seq):
mask = torch.zeros((batch_size, tokenizer.vocab_size), dtype=torch.int64, device=model.device)
for b in range(batch_size):
if end[b]:
mask[b, tokenizer.pad_id] = 1
continue
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[b, mask_ids] = 1
else:
param_names = tokenizer.events[event_names[b]]
if i > len(param_names):
mask[b, tokenizer.pad_id] = 1
continue
param_name = param_names[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[b, mask_ids] = 1
mask = mask.unsqueeze(1)
logits = model.forward_token(hidden, next_token_seq)[:, -1:]
scores = torch.softmax(logits / temp, dim=-1) * mask
samples = model.sample_top_p_k(scores, top_p, top_k, generator=generator)
if i == 0:
next_token_seq = samples
for b in range(batch_size):
if end[b]:
continue
eid = samples[b].item()
if eid == tokenizer.eos_id:
end[b] = True
else:
event_names[b] = tokenizer.id_events[eid]
else:
next_token_seq = torch.cat([next_token_seq, samples], dim=1)
if all([len(tokenizer.events[event_names[b]]) == i for b in range(batch_size) if not end[b]]):
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[:, 0].cpu().numpy()
if all(end):
break
def create_msg(name, data):
return {"name": name, "data": data}
def send_msgs(msgs):
return json.dumps(msgs)
def calc_time(x):
return 5.849e-5*x**2 + 0.04781*x + 0.1168
def get_duration(model_name, tab, mid_seq, continuation_state, continuation_select, instruments, drum_kit, bpm,
time_sig, key_sig, mid, midi_events, reduce_cc_st, remap_track_channel, add_default_instr,
remove_empty_channels, seed, seed_rand, gen_events, temp, top_p, top_k, allow_cc):
if tab == 0:
start_events = 1
elif tab == 1 and mid is not None:
start_events = midi_events
elif tab == 2 and mid_seq is not None:
start_events = len(mid_seq[0])
else:
start_events = 1
t = calc_time(start_events + gen_events) - calc_time(start_events) + 5
if "large" in model_name:
t *= 2
return t
@spaces.GPU(duration=get_duration)
def run(model_name, tab, mid_seq, continuation_state, continuation_select, instruments, drum_kit, bpm, time_sig,
key_sig, mid, midi_events, reduce_cc_st, remap_track_channel, add_default_instr, remove_empty_channels,
seed, seed_rand, gen_events, temp, top_p, top_k, allow_cc):
model, lora_name = models[model_name]
if lora_name is None and model.peft_loaded():
model.disable_adapters()
elif lora_name is not None:
model.enable_adapters()
model.set_adapter(lora_name)
model.to(device=opt.device)
tokenizer = model.tokenizer
bpm = int(bpm)
if time_sig == "auto":
time_sig = None
time_sig_nn = 4
time_sig_dd = 2
else:
time_sig_nn, time_sig_dd = time_sig.split('/')
time_sig_nn = int(time_sig_nn)
time_sig_dd = {2: 1, 4: 2, 8: 3}[int(time_sig_dd)]
if key_sig == 0:
key_sig = None
key_sig_sf = 0
key_sig_mi = 0
else:
key_sig = (key_sig - 1)
key_sig_sf = key_sig // 2 - 7
key_sig_mi = key_sig % 2
gen_events = int(gen_events)
max_len = gen_events
if seed_rand:
seed = random.randint(0, MAX_SEED)
generator = torch.Generator(opt.device).manual_seed(seed)
disable_patch_change = False
disable_channels = None
if tab == 0:
i = 0
mid = [[tokenizer.bos_id] + [tokenizer.pad_id] * (tokenizer.max_token_seq - 1)]
if tokenizer.version == "v2":
if time_sig is not None:
mid.append(tokenizer.event2tokens(["time_signature", 0, 0, 0, time_sig_nn - 1, time_sig_dd - 1]))
if key_sig is not None:
mid.append(tokenizer.event2tokens(["key_signature", 0, 0, 0, key_sig_sf + 7, key_sig_mi]))
if bpm != 0:
mid.append(tokenizer.event2tokens(["set_tempo", 0, 0, 0, bpm]))
patches = {}
if instruments is None:
instruments = []
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 + 1, c, p]))
mid = np.asarray([mid] * OUTPUT_BATCH_SIZE, dtype=np.int64)
mid_seq = mid.tolist()
if len(instruments) > 0:
disable_patch_change = True
disable_channels = [i for i in range(16) if i not in patches]
elif tab == 1 and mid is not None:
eps = 4 if reduce_cc_st else 0
mid = tokenizer.tokenize(MIDI.midi2score(mid), cc_eps=eps, tempo_eps=eps,
remap_track_channel=remap_track_channel,
add_default_instr=add_default_instr,
remove_empty_channels=remove_empty_channels)
mid = mid[:int(midi_events)]
mid = np.asarray([mid] * OUTPUT_BATCH_SIZE, dtype=np.int64)
mid_seq = mid.tolist()
elif tab == 2 and mid_seq is not None:
mid = np.asarray(mid_seq, dtype=np.int64)
if continuation_select > 0:
continuation_state.append(mid_seq)
mid = np.repeat(mid[continuation_select - 1:continuation_select], repeats=OUTPUT_BATCH_SIZE, axis=0)
mid_seq = mid.tolist()
else:
continuation_state.append(mid.shape[1])
else:
continuation_state = [0]
mid = [[tokenizer.bos_id] + [tokenizer.pad_id] * (tokenizer.max_token_seq - 1)]
mid = np.asarray([mid] * OUTPUT_BATCH_SIZE, dtype=np.int64)
mid_seq = mid.tolist()
if mid is not None:
max_len += mid.shape[1]
init_msgs = [create_msg("progress", [0, gen_events])]
if not (tab == 2 and continuation_select == 0):
for i in range(OUTPUT_BATCH_SIZE):
events = [tokenizer.tokens2event(tokens) for tokens in mid_seq[i]]
init_msgs += [create_msg("visualizer_clear", [i, tokenizer.version]),
create_msg("visualizer_append", [i, events])]
yield mid_seq, continuation_state, seed, send_msgs(init_msgs)
midi_generator = generate(model, mid, batch_size=OUTPUT_BATCH_SIZE, 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,
generator=generator)
events = [list() for i in range(OUTPUT_BATCH_SIZE)]
t = time.time()
for i, token_seqs in enumerate(midi_generator):
token_seqs = token_seqs.tolist()
for j in range(OUTPUT_BATCH_SIZE):
token_seq = token_seqs[j]
mid_seq[j].append(token_seq)
events[j].append(tokenizer.tokens2event(token_seq))
if time.time() - t > 0.2:
msgs = [create_msg("progress", [i + 1, gen_events])]
for j in range(OUTPUT_BATCH_SIZE):
msgs += [create_msg("visualizer_append", [j, events[j]])]
events[j] = list()
yield mid_seq, continuation_state, seed, send_msgs(msgs)
t = time.time()
yield mid_seq, continuation_state, seed, send_msgs([])
def finish_run(model_name, mid_seq):
if mid_seq is None:
outputs = [None] * OUTPUT_BATCH_SIZE
return *outputs, []
tokenizer = models[model_name][0].tokenizer
outputs = []
end_msgs = [create_msg("progress", [0, 0])]
if not os.path.exists("outputs"):
os.mkdir("outputs")
for i in range(OUTPUT_BATCH_SIZE):
events = [tokenizer.tokens2event(tokens) for tokens in mid_seq[i]]
mid = tokenizer.detokenize(mid_seq[i])
with open(f"outputs/output{i + 1}.mid", 'wb') as f:
f.write(MIDI.score2midi(mid))
outputs.append(f"outputs/output{i + 1}.mid")
end_msgs += [create_msg("visualizer_clear", [i, tokenizer.version]),
create_msg("visualizer_append", [i, events]),
create_msg("visualizer_end", i)]
return *outputs, send_msgs(end_msgs)
def synthesis_task(mid):
return synthesizer.synthesis(MIDI.score2opus(mid))
def render_audio(model_name, mid_seq, should_render_audio):
if (not should_render_audio) or mid_seq is None:
outputs = [None] * OUTPUT_BATCH_SIZE
return tuple(outputs)
tokenizer = models[model_name][0].tokenizer
outputs = []
if not os.path.exists("outputs"):
os.mkdir("outputs")
audio_futures = []
for i in range(OUTPUT_BATCH_SIZE):
mid = tokenizer.detokenize(mid_seq[i])
audio_future = thread_pool.submit(synthesis_task, mid)
audio_futures.append(audio_future)
for future in audio_futures:
outputs.append((44100, future.result()))
return tuple(outputs)
def undo_continuation(model_name, mid_seq, continuation_state):
if mid_seq is None or len(continuation_state) < 2:
return mid_seq, continuation_state, send_msgs([])
tokenizer = models[model_name][0].tokenizer
if isinstance(continuation_state[-1], list):
mid_seq = continuation_state[-1]
else:
mid_seq = [ms[:continuation_state[-1]] for ms in mid_seq]
continuation_state = continuation_state[:-1]
end_msgs = [create_msg("progress", [0, 0])]
for i in range(OUTPUT_BATCH_SIZE):
events = [tokenizer.tokens2event(tokens) for tokens in mid_seq[i]]
end_msgs += [create_msg("visualizer_clear", [i, tokenizer.version]),
create_msg("visualizer_append", [i, events]),
create_msg("visualizer_end", i)]
return mid_seq, continuation_state, send_msgs(end_msgs)
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:
js_content = jsfile.read()
js_content = js_content.replace("const MIDI_OUTPUT_BATCH_SIZE=4;",
f"const MIDI_OUTPUT_BATCH_SIZE={OUTPUT_BATCH_SIZE};")
javascript += f"\n<!-- {path} --><script>{js_content}</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
def hf_hub_download_retry(repo_id, filename):
print(f"downloading {repo_id} {filename}")
retry = 0
err = None
while retry < 30:
try:
return hf_hub_download(repo_id=repo_id, filename=filename)
except Exception as e:
err = e
retry += 1
if err:
raise err
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()}
key_signatures = ['C♭', 'A♭m', 'G♭', 'E♭m', 'D♭', 'B♭m', 'A♭', 'Fm', 'E♭', 'Cm', 'B♭', 'Gm', 'F', 'Dm',
'C', 'Am', 'G', 'Em', 'D', 'Bm', 'A', 'F♯m', 'E', 'C♯m', 'B', 'G♯m', 'F♯', 'D♯m', 'C♯', 'A♯m']
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="cuda", help="device to run model")
parser.add_argument("--batch", type=int, default=8, help="batch size")
parser.add_argument("--max-gen", type=int, default=1024, help="max")
opt = parser.parse_args()
OUTPUT_BATCH_SIZE = opt.batch
soundfont_path = hf_hub_download_retry(repo_id="skytnt/midi-model", filename="soundfont.sf2")
thread_pool = ThreadPoolExecutor(max_workers=OUTPUT_BATCH_SIZE)
synthesizer = MidiSynthesizer(soundfont_path)
models_info = {
"generic pretrain model (tv2o-medium) by skytnt": [
"skytnt/midi-model-tv2o-medium", "", "tv2o-medium", {
"jpop": "skytnt/midi-model-tv2om-jpop-lora",
"touhou": "skytnt/midi-model-tv2om-touhou-lora"
}
],
"generic pretrain model (tv2o-large) by asigalov61": [
"asigalov61/Music-Llama", "", "tv2o-large", {}
],
"generic pretrain model (tv2o-medium) by asigalov61": [
"asigalov61/Music-Llama-Medium", "", "tv2o-medium", {}
],
"generic pretrain model (tv1-medium) by skytnt": [
"skytnt/midi-model", "", "tv1-medium", {}
]
}
models = {}
if opt.device == "cuda":
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_flash_sdp(True)
for name, (repo_id, path, config, loras) in models_info.items():
model_path = hf_hub_download_retry(repo_id=repo_id, filename=f"{path}model.ckpt")
model = MIDIModel(config=MIDIModelConfig.from_name(config))
ckpt = torch.load(model_path, map_location="cpu", weights_only=True)
state_dict = ckpt.get("state_dict", ckpt)
model.load_state_dict(state_dict, strict=False)
for lora_name, lora_repo in loras.items():
model.load_adapter(lora_repo, lora_name)
if loras:
model.disable_adapters()
model.to(device="cpu", dtype=torch.float32).eval()
models[name] = model, None
for lora_name, lora_repo in loras.items():
models[f"{name} with {lora_name} lora"] = model, lora_name
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 symbolic 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)"
" or [download windows app](https://github.com/SkyTNT/midi-model/releases)"
" for unlimited generation\n\n"
"**Update v1.3**: MIDITokenizerV2 and new MidiVisualizer"
)
js_msg = gr.Textbox(elem_id="msg_receiver", visible=False)
js_msg.change(None, [js_msg], [], js="""
(msg_json) =>{
let msgs = JSON.parse(msg_json);
executeCallbacks(msgReceiveCallbacks, msgs);
return [];
}
""")
input_model = gr.Dropdown(label="select model", choices=list(models.keys()),
type="value", value=list(models.keys())[0])
tab_select = gr.State(value=0)
with gr.Tabs():
with gr.TabItem("custom 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")
input_bpm = gr.Slider(label="BPM (beats per minute, auto if 0)", minimum=0, maximum=255,
step=1,
value=0)
input_time_sig = gr.Radio(label="time signature (only for tv2 models)",
value="auto",
choices=["auto", "4/4", "2/4", "3/4", "6/4", "7/4",
"2/2", "3/2", "4/2", "3/8", "5/8", "6/8", "7/8", "9/8", "12/8"]
)
input_key_sig = gr.Radio(label="key signature (only for tv2 models)",
value="auto",
choices=["auto"] + key_signatures,
type="index"
)
example1 = gr.Examples([
[[], "None"],
[["Acoustic Grand"], "None"],
[['Acoustic Grand', 'SynthStrings 2', 'SynthStrings 1', 'Pizzicato Strings',
'Pad 2 (warm)', 'Tremolo Strings', 'String Ensemble 1'], "Orchestra"],
[['Trumpet', 'Oboe', 'Trombone', 'String Ensemble 1', 'Clarinet',
'French Horn', 'Pad 4 (choir)', 'Bassoon', 'Flute'], "None"],
[['Flute', 'French Horn', 'Clarinet', 'String Ensemble 2', 'English Horn', 'Bassoon',
'Oboe', 'Pizzicato Strings'], "Orchestra"],
[['Electric Piano 2', 'Lead 5 (charang)', 'Electric Bass(pick)', 'Lead 2 (sawtooth)',
'Pad 1 (new age)', 'Orchestra Hit', 'Cello', 'Electric Guitar(clean)'], "Standard"],
[["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)
input_reduce_cc_st = gr.Checkbox(label="reduce control_change and set_tempo events", value=True)
input_remap_track_channel = gr.Checkbox(
label="remap tracks and channels so each track has only one channel and in order", value=True)
input_add_default_instr = gr.Checkbox(
label="add a default instrument to channels that don't have an instrument", value=True)
input_remove_empty_channels = gr.Checkbox(label="remove channels without notes", value=False)
example2 = gr.Examples([[file, 128] for file in glob.glob("example/*.mid")],
[input_midi, input_midi_events])
with gr.TabItem("last output prompt") as tab3:
gr.Markdown("Continue generating on the last output.")
input_continuation_select = gr.Radio(label="select output to continue generating", value="all",
choices=["all"] + [f"output{i + 1}" for i in
range(OUTPUT_BATCH_SIZE)],
type="index"
)
undo_btn = gr.Button("undo the last continuation")
tab1.select(lambda: 0, None, tab_select, queue=False)
tab2.select(lambda: 1, None, tab_select, queue=False)
tab3.select(lambda: 2, None, tab_select, queue=False)
input_seed = gr.Slider(label="seed", minimum=0, maximum=2 ** 31 - 1,
step=1, value=0)
input_seed_rand = gr.Checkbox(label="random seed", value=True)
input_gen_events = gr.Slider(label="generate max 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=0.95)
input_top_k = gr.Slider(label="top k", minimum=1, maximum=128, step=1, value=20)
input_allow_cc = gr.Checkbox(label="allow midi cc event", value=True)
input_render_audio = gr.Checkbox(label="render audio after generation", value=True)
example3 = gr.Examples([[1, 0.94, 128], [1, 0.98, 20], [1, 0.98, 12]],
[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.State()
output_continuation_state = gr.State([0])
midi_outputs = []
audio_outputs = []
with gr.Tabs(elem_id="output_tabs"):
for i in range(OUTPUT_BATCH_SIZE):
with gr.TabItem(f"output {i + 1}") as tab1:
output_midi_visualizer = gr.HTML(elem_id=f"midi_visualizer_container_{i}")
output_audio = gr.Audio(label="output audio", format="mp3", elem_id=f"midi_audio_{i}")
output_midi = gr.File(label="output midi", file_types=[".mid"])
midi_outputs.append(output_midi)
audio_outputs.append(output_audio)
run_event = run_btn.click(run, [input_model, tab_select, output_midi_seq, output_continuation_state,
input_continuation_select, input_instruments, input_drum_kit, input_bpm,
input_time_sig, input_key_sig, input_midi, input_midi_events,
input_reduce_cc_st, input_remap_track_channel,
input_add_default_instr, input_remove_empty_channels,
input_seed, input_seed_rand, input_gen_events, input_temp, input_top_p,
input_top_k, input_allow_cc],
[output_midi_seq, output_continuation_state, input_seed, js_msg],
concurrency_limit=10, queue=True)
finish_run_event = run_event.then(fn=finish_run,
inputs=[input_model, output_midi_seq],
outputs=midi_outputs + [js_msg],
queue=False)
finish_run_event.then(fn=render_audio,
inputs=[input_model, output_midi_seq, input_render_audio],
outputs=audio_outputs,
queue=False)
# stop_btn.click(None, [], [], cancels=run_event,
# queue=False)
undo_btn.click(undo_continuation, [input_model, output_midi_seq, output_continuation_state],
[output_midi_seq, output_continuation_state, js_msg], queue=False)
app.queue().launch(server_port=opt.port, share=opt.share, inbrowser=True)
thread_pool.shutdown()
|