Spaces:
Running
Running
File size: 25,688 Bytes
bb70eb3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 |
# Third-party
import torch
import torch.nn as nn
# Local
from vocal_isolation.short_time_fourier_transform import STFT
from vocal_isolation.constants import COMPUTATION_DEVICE
dim_s = 4
dim_c = 4
class TFCC(nn.Module):
def __init__(self, c, l, k):
super(TFCC, self).__init__()
self.H = nn.ModuleList()
for i in range(l):
self.H.append(
nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=c, kernel_size=k, stride=1, padding=k // 2),
nn.GroupNorm(4, c),
nn.ReLU(),
)
)
def forward(self, x):
for h in self.H:
x = h(x)
return x
class DenseTFCC(nn.Module):
def __init__(self, c, l, k):
super(DenseTFCC, self).__init__()
self.conv = nn.ModuleList()
for i in range(l):
self.conv.append(
nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=c, kernel_size=k, stride=1, padding=k // 2),
nn.GroupNorm(4, c),
nn.ReLU(),
)
)
def forward(self, x):
for layer in self.conv[:-1]:
x = torch.cat([layer(x), x], 1)
return self.conv[-1](x)
class TFC_TDFF(nn.Module):
def __init__(self, c, l, f, k, bn, dense=False, bias=True):
super(TFC_TDFF, self).__init__()
self.use_tdf = bn is not None
self.tfc = DenseTFCC(c, l, k) if dense else TFCC(c, l, k)
if self.use_tdf:
if bn == 0:
self.tdf = nn.Sequential(
nn.Linear(f, f, bias=bias),
nn.GroupNorm(4, c),
nn.ReLU()
)
else:
self.tdf = nn.Sequential(
nn.Linear(f, f // bn, bias=bias),
nn.GroupNorm(4, c),
nn.ReLU(),
nn.Linear(f // bn, f, bias=bias),
nn.GroupNorm(4, c),
nn.ReLU()
)
def forward(self, x):
x = self.tfc(x)
return x + self.tdf(x) if self.use_tdf else x
class TFC(nn.Module):
def __init__(self, c, l, k):
super(TFC, self).__init__()
self.H = nn.ModuleList()
for i in range(l):
self.H.append(
nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=c, kernel_size=k, stride=1, padding=k // 2),
nn.GroupNorm(2, c),
nn.ReLU(),
)
)
def forward(self, x):
for h in self.H:
x = h(x)
return x
class DenseTFC(nn.Module):
def __init__(self, c, l, k):
super(DenseTFC, self).__init__()
self.conv = nn.ModuleList()
for i in range(l):
self.conv.append(
nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=c, kernel_size=k, stride=1, padding=k // 2),
nn.GroupNorm(2, c),
nn.ReLU(),
)
)
def forward(self, x):
for layer in self.conv[:-1]:
x = torch.cat([layer(x), x], 1)
return self.conv[-1](x)
class TFC_TDF(nn.Module):
def __init__(self, c, l, f, k, bn, dense=False, bias=True):
super(TFC_TDF, self).__init__()
self.use_tdf = bn is not None
self.tfc = DenseTFC(c, l, k) if dense else TFC(c, l, k)
if self.use_tdf:
if bn == 0:
self.tdf = nn.Sequential(
nn.Linear(f, f, bias=bias),
nn.GroupNorm(2, c),
nn.ReLU()
)
else:
self.tdf = nn.Sequential(
nn.Linear(f, f // bn, bias=bias),
nn.GroupNorm(2, c),
nn.ReLU(),
nn.Linear(f // bn, f, bias=bias),
nn.GroupNorm(2, c),
nn.ReLU()
)
def forward(self, x):
x = self.tfc(x)
return x + self.tdf(x) if self.use_tdf else x
class Conv_TDF(nn.Module):
"""
Convolutional Time-Domain Filter (TDF) Module.
Args:
c (int): The number of input and output channels for the convolutional layers.
l (int): The number of convolutional layers within the module.
f (int): The number of features (or units) in the time-domain filter.
k (int): The size of the convolutional kernels (filters).
bn (int or None): Batch normalization factor (controls TDF behavior). If None, TDF is not used.
bias (bool): A boolean flag indicating whether bias terms are included in the linear layers.
Attributes:
use_tdf (bool): Flag indicating whether TDF is used.
Methods:
forward(x): Forward pass through the TDF module.
"""
def __init__(self, c, l, f, k, bn, bias=True):
super(Conv_TDF, self).__init__()
# Determine whether to use TDF (Time-Domain Filter)
self.use_tdf = bn is not None
# Define a list of convolutional layers within the module
self.H = nn.ModuleList()
for i in range(l):
self.H.append(
nn.Sequential(
nn.Conv2d(
in_channels=c,
out_channels=c,
kernel_size=k,
stride=1,
padding=k // 2,
),
nn.GroupNorm(2, c),
nn.ReLU(),
)
)
# Define the Time-Domain Filter (TDF) layers if enabled
if self.use_tdf:
if bn == 0:
self.tdf = nn.Sequential(
nn.Linear(f, f, bias=bias), nn.GroupNorm(2, c), nn.ReLU()
)
else:
self.tdf = nn.Sequential(
nn.Linear(f, f // bn, bias=bias),
nn.GroupNorm(2, c),
nn.ReLU(),
nn.Linear(f // bn, f, bias=bias),
nn.GroupNorm(2, c),
nn.ReLU(),
)
def forward(self, x):
# Apply the convolutional layers sequentially
for h in self.H:
x = h(x)
# Apply the Time-Domain Filter (TDF) if enabled, and add the result to the orignal input
return x + self.tdf(x) if self.use_tdf else x
class Conv_TDF_net_trimm(nn.Module):
"""
Convolutional Time-Domain Filter (TDF) Network with Trimming.
Used for Vocals.
Args:
L (int): This parameter controls the number of down-sampling (DS) blocks in the network.
It's divided by 2 to determine how many DS blocks should be created.
l (int): This parameter represents the number of convolutional layers (or filters) within each dense (fully connected) block.
g (int): This parameter specifies the number of output channels for the first convolutional layer and is also used to determine the number of channels for subsequent layers in the network.
dim_f (int): This parameter represents the number of frequency bins (spectrogram columns) in the input audio data.
dim_t (int): This parameter represents the number of time frames (spectrogram rows) in the input audio data.
k (int): This parameter specifies the size of convolutional kernels (filters) used in the network's convolutional layers.
bn (int or None): This parameter controls whether batch normalization is used in the network.
If it's None, batch normalization may or may not be used based on other conditions in the code.
bias (bool): This parameter is a boolean flag that controls whether bias terms are included in the convolutional layers.
overlap (int): This parameter specifies the amount of overlap between consecutive chunks of audio data during processing.
Attributes:
n (int): The calculated number of down-sampling (DS) blocks.
dim_f (int): The number of frequency bins (spectrogram columns) in the input audio data.
dim_t (int): The number of time frames (spectrogram rows) in the input audio data.
n_fft (int): The size of the Fast Fourier Transform (FFT) window.
hop (int): The hop size used in the STFT calculations.
n_bins (int): The number of bins in the frequency domain.
chunk_size (int): The size of each chunk of audio data.
target_name (str): The name of the target instrument being separated.
overlap (int): The amount of overlap between consecutive chunks of audio data during processing.
Methods:
forward(x): Forward pass through the Conv_TDF_net_trimm network.
"""
def __init__(
self,
model_path,
use_onnx,
target_name,
L,
l,
g,
dim_f,
dim_t,
k=3,
hop=1024,
bn=None,
bias=True,
overlap=1500,
):
super(Conv_TDF_net_trimm, self).__init__()
# Dictionary specifying the scale for the number of FFT bins for different target names
n_fft_scale = {"vocals": 3, "*": 2}
# Number of input and output channels for the initial and final convolutional layers
out_c = in_c = 4
# Number of down-sampling (DS) blocks
self.n = L // 2
# Dimensions of the frequency and time axes of the input data
self.dim_f = 3072
self.dim_t = 256
# Number of FFT bins (frequencies) and hop size for the Short-Time Fourier Transform (STFT)
self.n_fft = 7680
self.hop = hop
self.n_bins = self.n_fft // 2 + 1
# Chunk size used for processing
self.chunk_size = hop * (self.dim_t - 1)
# Target name for the model
self.target_name = target_name
# Overlap between consecutive chunks of audio data during processing
self.overlap = overlap
# STFT module for audio processing
self.stft = STFT(self.n_fft, self.hop, self.dim_f)
# Check if ONNX representation of the model should be used
if not use_onnx:
# First convolutional layer
self.first_conv = nn.Sequential(
nn.Conv2d(in_channels=in_c, out_channels=g, kernel_size=1, stride=1),
nn.BatchNorm2d(g),
nn.ReLU(),
)
# Initialize variables for dense (fully connected) blocks and downsampling (DS) blocks
f = self.dim_f
c = g
self.ds_dense = nn.ModuleList()
self.ds = nn.ModuleList()
# Loop through down-sampling (DS) blocks
for i in range(self.n):
# Create dense (fully connected) block for down-sampling
self.ds_dense.append(Conv_TDF(c, l, f, k, bn, bias=bias))
# Create down-sampling (DS) block
scale = (2, 2)
self.ds.append(
nn.Sequential(
nn.Conv2d(
in_channels=c,
out_channels=c + g,
kernel_size=scale,
stride=scale,
),
nn.BatchNorm2d(c + g),
nn.ReLU(),
)
)
f = f // 2
c += g
# Middle dense (fully connected block)
self.mid_dense = Conv_TDF(c, l, f, k, bn, bias=bias)
# If batch normalization is not specified and mid_tdf is True, use Conv_TDF with bn=0 and bias=False
if bn is None and mid_tdf:
self.mid_dense = Conv_TDF(c, l, f, k, bn=0, bias=False)
# Initialize variables for up-sampling (US) blocks
self.us_dense = nn.ModuleList()
self.us = nn.ModuleList()
# Loop through up-sampling (US) blocks
for i in range(self.n):
scale = (2, 2)
# Create up-sampling (US) block
self.us.append(
nn.Sequential(
nn.ConvTranspose2d(
in_channels=c,
out_channels=c - g,
kernel_size=scale,
stride=scale,
),
nn.BatchNorm2d(c - g),
nn.ReLU(),
)
)
f = f * 2
c -= g
# Create dense (fully connected) block for up-sampling
self.us_dense.append(Conv_TDF(c, l, f, k, bn, bias=bias))
# Final convolutional layer
self.final_conv = nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=out_c, kernel_size=1, stride=1),
)
try:
# Load model state from a file
self.load_state_dict(
torch.load(
f"{model_path}/{target_name}.pt",
map_location=COMPUTATION_DEVICE,
)
)
print(f"Loading model ({target_name})")
except FileNotFoundError:
print(f"Random init ({target_name})")
def forward(self, x):
"""
Forward pass through the Conv_TDF_net_trimm network.
Args:
x (torch.Tensor): Input tensor.
Returns:
torch.Tensor: Output tensor after passing through the network.
"""
x = self.first_conv(x)
x = x.transpose(-1, -2)
ds_outputs = []
for i in range(self.n):
x = self.ds_dense[i](x)
ds_outputs.append(x)
x = self.ds[i](x)
x = self.mid_dense(x)
for i in range(self.n):
x = self.us[i](x)
x *= ds_outputs[-i - 1]
x = self.us_dense[i](x)
x = x.transpose(-1, -2)
x = self.final_conv(x)
return x
class Conv_TDF_net_trim(nn.Module):
"""
Convolutional Time-Domain Filter (TDF) Network with Trimming.
Used for drums and other.
Args:
L (int): This parameter controls the number of down-sampling (DS) blocks in the network.
It's divided by 2 to determine how many DS blocks should be created.
l (int): This parameter represents the number of convolutional layers (or filters) within each dense (fully connected) block.
g (int): This parameter specifies the number of output channels for the first convolutional layer and is also used to determine the number of channels for subsequent layers in the network.
dim_f (int): This parameter represents the number of frequency bins (spectrogram columns) in the input audio data.
dim_t (int): This parameter represents the number of time frames (spectrogram rows) in the input audio data.
k (int): This parameter specifies the size of convolutional kernels (filters) used in the network's convolutional layers.
bn (int or None): This parameter controls whether batch normalization is used in the network.
If it's None, batch normalization may or may not be used based on other conditions in the code.
bias (bool): This parameter is a boolean flag that controls whether bias terms are included in the convolutional layers.
overlap (int): This parameter specifies the amount of overlap between consecutive chunks of audio data during processing.
Attributes:
n (int): The calculated number of down-sampling (DS) blocks.
dim_f (int): The number of frequency bins (spectrogram columns) in the input audio data.
dim_t (int): The number of time frames (spectrogram rows) in the input audio data.
n_fft (int): The size of the Fast Fourier Transform (FFT) window.
hop (int): The hop size used in the STFT calculations.
n_bins (int): The number of bins in the frequency domain.
chunk_size (int): The size of each chunk of audio data.
target_name (str): The name of the target instrument being separated.
overlap (int): The amount of overlap between consecutive chunks of audio data during processing.
Methods:
forward(x): Forward pass through the Conv_TDF_net_trimm network.
"""
def __init__(self, model_path, use_onnx, target_name,
L, l, g, dim_f, dim_t, k=3, hop=1024, bn=None, bias=False, overlap=1754):
super(Conv_TDF_net_trim, self).__init__()
n_fft_scale = {'drums', 'other'}
out_c = in_c = 4
self.n = L//2
self.dim_f = 3072
self.dim_t = 256
self.n_fft = 7680
self.hop = hop
self.n_bins = self.n_fft//2+1
self.chunk_size = hop * (self.dim_t-1)
self.use_onnx = use_onnx
self.target_name = target_name
self.overlap = overlap
self.window = nn.Parameter(torch.hann_window(window_length=self.n_fft, periodic=True), requires_grad=False)
self.freq_pad = nn.Parameter(torch.zeros([1, dim_c, self.n_bins - self.dim_f, self.dim_t]), requires_grad=False)
self.stft = STFT(self.n_fft, self.hop, self.dim_f)
if not use_onnx:
scale = (2, 2)
self.first_conv = nn.Sequential(
nn.Conv2d(in_channels=4, out_channels=g, kernel_size=(1, 1)),
nn.GroupNorm(2, g),
nn.ReLU(),
)
f = self.dim_f
c = g
self.encoding_blocks = nn.ModuleList()
self.ds = nn.ModuleList()
for i in range(self.n):
self.encoding_blocks.append(TFC_TDF(c, l, f, k, bn, bias=bias))
self.ds.append(
nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=c + g, kernel_size=scale, stride=scale),
nn.GroupNorm(2, c + g),
nn.ReLU()
)
)
f = f // 2
c += g
self.bottleneck_block = TFC_TDF(c, l, f, k, bn, bias=bias)
self.decoding_blocks = nn.ModuleList()
self.us = nn.ModuleList()
for i in range(self.n):
self.us.append(
nn.Sequential(
nn.ConvTranspose2d(in_channels=c, out_channels=c - g, kernel_size=scale, stride=scale),
nn.GroupNorm(2, c - g),
nn.ReLU()
)
)
f = f * 2
c -= g
self.decoding_blocks.append(TFC_TDF(c, l, f, k, bn, bias=bias))
self.final_conv = nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=4, kernel_size=(1, 1)),
)
try:
self.load_state_dict(
torch.load(f'{model_path}/{target_name}.pt', map_location=lambda storage, loc: storage)
)
print(f'Loading model ({target_name})')
except FileNotFoundError:
print(f'Random init ({target_name})')
def forward(self, x):
x = self.first_conv(x)
x = x.transpose(-1, -2)
ds_outputs = []
for i in range(self.n):
x = self.encoding_blocks[i](x)
ds_outputs.append(x)
x = self.ds[i](x)
x = self.bottleneck_block(x)
for i in range(self.n):
x = self.us[i](x)
x *= ds_outputs[-i - 1]
x = self.decoding_blocks[i](x)
x = x.transpose(-1, -2)
x = self.final_conv(x)
return x
class Conv_TDF_net_trimmmm(nn.Module):
"""
Convolutional Time-Domain Filter (TDF) Network with Trimming.
Used for bass.
Args:
L (int): This parameter controls the number of down-sampling (DS) blocks in the network.
It's divided by 2 to determine how many DS blocks should be created.
l (int): This parameter represents the number of convolutional layers (or filters) within each dense (fully connected) block.
g (int): This parameter specifies the number of output channels for the first convolutional layer and is also used to determine the number of channels for subsequent layers in the network.
dim_f (int): This parameter represents the number of frequency bins (spectrogram columns) in the input audio data.
dim_t (int): This parameter represents the number of time frames (spectrogram rows) in the input audio data.
k (int): This parameter specifies the size of convolutional kernels (filters) used in the network's convolutional layers.
bn (int or None): This parameter controls whether batch normalization is used in the network.
If it's None, batch normalization may or may not be used based on other conditions in the code.
bias (bool): This parameter is a boolean flag that controls whether bias terms are included in the convolutional layers.
overlap (int): This parameter specifies the amount of overlap between consecutive chunks of audio data during processing.
Attributes:
n (int): The calculated number of down-sampling (DS) blocks.
dim_f (int): The number of frequency bins (spectrogram columns) in the input audio data.
dim_t (int): The number of time frames (spectrogram rows) in the input audio data.
n_fft (int): The size of the Fast Fourier Transform (FFT) window.
hop (int): The hop size used in the STFT calculations.
n_bins (int): The number of bins in the frequency domain.
chunk_size (int): The size of each chunk of audio data.
target_name (str): The name of the target instrument being separated.
overlap (int): The amount of overlap between consecutive chunks of audio data during processing.
Methods:
forward(x): Forward pass through the Conv_TDF_net_trimm network.
"""
def __init__(self, model_path, use_onnx, target_name,
L, l, g, dim_f, dim_t, k=3, hop=1024, bn=None, bias=False, overlap=1350):
super(Conv_TDF_net_trimmmm, self).__init__()
n_fft_scale = {'bass'}
out_c = in_c = 4
self.n = L//2
self.dim_f = 2048
self.dim_t = 256
self.n_fft = 16384
self.hop = hop
self.n_bins = self.n_fft//2+1
self.chunk_size = hop * (self.dim_t-1)
self.use_onnx = use_onnx
self.target_name = target_name
self.overlap = overlap
self.window = nn.Parameter(torch.hann_window(window_length=self.n_fft, periodic=True), requires_grad=False)
self.freq_pad = nn.Parameter(torch.zeros([1, dim_c, self.n_bins - self.dim_f, self.dim_t]), requires_grad=False)
self.stft = STFT(self.n_fft, self.hop, self.dim_f)
if not use_onnx:
scale = (2, 2)
self.first_conv = nn.Sequential(
nn.Conv2d(in_channels=4, out_channels=g, kernel_size=(1, 1)),
nn.GroupNorm(4, g),
nn.ReLU(),
)
f = self.dim_f
c = g
self.encoding_blocks = nn.ModuleList()
self.ds = nn.ModuleList()
for i in range(self.n):
self.encoding_blocks.append(TFC_TDFF(c, l, f, k, bn, bias=bias))
self.ds.append(
nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=c + g, kernel_size=scale, stride=scale),
nn.GroupNorm(4, c + g),
nn.ReLU()
)
)
f = f // 2
c += g
self.bottleneck_block = TFC_TDFF(c, l, f, k, bn, bias=bias)
self.decoding_blocks = nn.ModuleList()
self.us = nn.ModuleList()
for i in range(self.n):
self.us.append(
nn.Sequential(
nn.ConvTranspose2d(in_channels=c, out_channels=c - g, kernel_size=scale, stride=scale),
nn.GroupNorm(4, c - g),
nn.ReLU()
)
)
f = f * 2
c -= g
self.decoding_blocks.append(TFC_TDFF(c, l, f, k, bn, bias=bias))
self.final_conv = nn.Sequential(
nn.Conv2d(in_channels=c, out_channels=4, kernel_size=(1, 1)),
)
try:
self.load_state_dict(
torch.load(f'{model_path}/{target_name}.pt', map_location=lambda storage, loc: storage)
)
print(f'Loading model ({target_name})')
except FileNotFoundError:
print(f'Random init ({target_name})')
def forward(self, x):
x = self.first_conv(x)
x = x.transpose(-1, -2)
ds_outputs = []
for i in range(self.n):
x = self.encoding_blocks[i](x)
ds_outputs.append(x)
x = self.ds[i](x)
x = self.bottleneck_block(x)
for i in range(self.n):
x = self.us[i](x)
x *= ds_outputs[-i - 1]
x = self.decoding_blocks[i](x)
x = x.transpose(-1, -2)
x = self.final_conv(x)
return x
|