Spaces:
Sleeping
Sleeping
File size: 17,078 Bytes
d9e647a 56e1924 d9e647a 56e1924 d9e647a 56e1924 d9e647a 56e1924 d9e647a 56e1924 d9e647a 56e1924 d9e647a |
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 |
from collections import OrderedDict
from typing import List, Literal, Optional, Dict, Any, Union
import torch
from torch import nn
from flamo import dsp, system
from flamo.auxiliary.reverb import (
parallelFDNAccurateGEQ,
parallelFirstOrderShelving,
)
from flamo.functional import signal_gallery
from flareverb.config.config import (
BaseConfig,
FDNAttenuation,
FDNMixing,
FDNConfig,
)
from flareverb.utils import ms_to_samps, rt2slope
from flareverb.reverb import MapGamma
class BaseFDN(nn.Module):
"""Base Feedback Delay Network (FDN) class for reverberation modeling.
"""
def __init__(
self,
config: FDNConfig,
nfft: int,
alias_decay_db: float,
delay_lengths: List[int],
device: Literal["cpu", "cuda"] = "cuda",
requires_grad: bool = True,
output_layer: Literal["freq_complex", "freq_mag", "time"] = "time",
) -> None:
"""
"""
super().__init__()
self._validate_delays(config, delay_lengths)
self._initialize_parameters(
config, nfft, alias_decay_db, delay_lengths, device, requires_grad
)
self._setup_fdn_system(config, output_layer)
def forward(
self,
inputs: torch.Tensor,
ext_params: List[Dict[str, Any]],
) -> torch.Tensor:
"""
Forward pass through the FDN.
Processes input signals through the Feedback Delay Network to generate
reverberated output. Each input can have its own set of external parameters
for dynamic control of the FDN characteristics.
Parameters
----------
inputs : torch.Tensor
Input tensor of shape (batch_size, signal_length).
ext_params : List[Dict[str, Any]]
List of external parameters for each input signal. Each dictionary
can contain parameters to modify the FDN behavior during processing.
The length must match the batch size.
Returns
-------
torch.Tensor
Processed output tensor. Contains the reverberated signals.
"""
outputs = []
for x, ext_param in zip(inputs, ext_params):
# Apply the FDN with the external parameters
output = self.shell(x[..., None], ext_param)
outputs.append(output)
return torch.stack(outputs).squeeze(-1)
def get_params(self) -> OrderedDict[str, Any]:
"""
Get the current parameters of the FDN.
Extracts all learnable and configurable parameters from the FDN system
for analysis, storage, or parameter transfer. All parameters are converted
to CPU NumPy arrays for compatibility.
Returns
-------
OrderedDict[str, Any]
Dictionary containing all FDN parameters:
- 'delays': List of delay lengths in samples
- 'onset_time': List of onset times in milliseconds
- 'early_reflections': Direct path gain values
- 'input_gains': Input gain coefficients for each delay line
- 'output_gains': Output gain coefficients for each delay line
- 'feedback_matrix': Mixing (feedback) matrix coefficients
- 'attenuation': Attenuation coefficients for each delay line
Notes
-----
- All parameters are detached from the computation graph and moved to CPU
- The returned parameters can be used to recreate or modify the FDN
"""
core = self.shell.get_core()
map_matrix = core.branchA.feedback_loop.feedback.mixing_matrix.map
params = OrderedDict()
params["delays"] = self.delay_lengths.cpu().numpy().tolist()
params["onset_time"] = self.onset
params["early_reflections"] = (
core.branchB.early_reflections.param.cpu().detach().numpy().tolist()
)
params["input_gains"] = (
core.branchA.input_gain.param.cpu().squeeze().detach().numpy().tolist()
)
params["output_gains"] = (
core.branchA.output_gain.param[0].cpu().squeeze().detach().numpy().tolist()
)
params["feedback_matrix"] = (
map_matrix(core.branchA.feedback_loop.feedback.mixing_matrix.param).cpu()
.detach()
.squeeze()
.numpy()
.tolist()
)
# params["attenuation"] = (
# core.branchA.feedback_loop.feedback.attenuation.param.cpu()
# .detach()
# .numpy()
# .tolist()
# )
return params
def _validate_delays(self, config: BaseConfig, delay_lengths: List[int]) -> None:
"""Validate delay lengths."""
if config.N != len(delay_lengths):
raise ValueError(
f"N ({config.N}) must match the length of delay_lengths ({len(delay_lengths)})"
)
def _initialize_parameters(
self,
config: FDNConfig,
nfft: int,
alias_decay_db: float,
delay_lengths: List[int],
device: str,
requires_grad: bool,
) -> None:
"""Initialize FDN parameters."""
self.device = torch.device(device)
# Core FDN parameters
self.N = config.N
self.fs = config.fs
self.nfft = nfft
self.alias_decay_db = alias_decay_db
self.requires_grad = requires_grad
# Onset configuration
self.early_reflections_type = config.early_reflections_type
self.onset = ms_to_samps(torch.tensor(config.onset_time), config.fs)
# Channel configuration
self.in_ch = config.in_ch
self.out_ch = config.out_ch
# Delay configuration
self.delay_lengths = torch.tensor(
delay_lengths, device=self.device, dtype=torch.int64
)
def _setup_fdn_system(self, config: BaseConfig, output_layer: str) -> None:
"""Setup the complete FDN system."""
# Create FDN branches
branch_a = self._create_fdn_branch(
config.attenuation_config, config.mixing_matrix_config
)
branch_b = self._create_direct_path(config)
# Combine branches
fdn_core = system.Parallel(brA=branch_a, brB=branch_b, sum_output=True)
# Setup I/O layers
input_layer = dsp.FFT(self.nfft)
output_layer = self._create_output_layer(output_layer)
# Create shell
self.shell = system.Shell(
core=fdn_core,
input_layer=input_layer,
output_layer=output_layer,
)
def _create_output_layer(self, output_type: str):
"""Create the appropriate output layer based on type."""
if output_type == "time":
return dsp.iFFTAntiAlias(nfft=self.nfft, alias_decay_db=self.alias_decay_db)
elif output_type == "freq_complex":
return dsp.Transform(transform=lambda x: x)
elif output_type == "freq_mag":
return dsp.Transform(transform=lambda x: torch.abs(x))
else:
raise ValueError(f"Unsupported output layer type: {output_type}")
def _create_fdn_branch(
self, attenuation_config: FDNAttenuation, mixing_matrix_config: FDNMixing
):
"""Create the main FDN branch (branch A)."""
# Input and output gains
input_gain = dsp.Gain(
size=(self.N, self.in_ch),
nfft=self.nfft,
requires_grad=self.requires_grad,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
output_gain = dsp.Gain(
size=(self.out_ch, self.N),
nfft=self.nfft,
requires_grad=self.requires_grad,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
# Feedback loop components
delays = self._create_delay_lines()
mixing_matrix = self._create_mixing_matrix(mixing_matrix_config)
attenuation = self._create_attenuation(attenuation_config)
# Feedback path
feedback = system.Series(
OrderedDict({"mixing_matrix": mixing_matrix, "attenuation": attenuation})
)
# Recursion
feedback_loop = system.Recursion(fF=delays, fB=feedback)
# Complete FDN branch
return system.Series(
OrderedDict(
{
"input_gain": input_gain,
"feedback_loop": feedback_loop,
"output_gain": output_gain,
}
)
)
def _create_delay_lines(self):
"""Create parallel delay lines."""
delays = dsp.parallelDelay(
size=(self.N,),
max_len=self.delay_lengths.max(),
nfft=self.nfft,
isint=True,
requires_grad=False,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
delays.assign_value(delays.sample2s(self.delay_lengths))
return delays
def _create_mixing_matrix(self, config: FDNMixing):
"""Create orthogonal mixing matrix."""
if config.is_scattering or config.is_velvet_noise:
m_L = torch.randint(
low=1,
high=int(torch.floor(min(self.delay_lengths) / 10)),
size=[self.N],
)
m_R = torch.randint(
low=1,
high=int(torch.floor(min(self.delay_lengths) / 10)),
size=[self.N],
)
if config.is_scattering:
mixing = dsp.ScatteringMatrix(
size=(config.n_stages, self.N, self.N),
nfft=self.nfft,
sparsity=config.sparsity,
gain_per_sample=1.0,
m_L=m_L,
m_R=m_R,
requires_grad=self.requires_grad,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
else:
mixing = dsp.VelvetNoiseMatrix(
size=(config.n_stages, self.N, self.N),
nfft=self.nfft,
density=1 / config.sparsity,
gain_per_sample=1.0,
m_L=m_L,
m_R=m_R,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
elif config.mixing_type == "householder":
mixing = dsp.HouseholderMatrix(
size=(self.N, self.N),
nfft=self.nfft,
requires_grad=self.requires_grad,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
else:
try:
mixing = dsp.Matrix(
size=(self.N, self.N),
nfft=self.nfft,
matrix_type=config.mixing_type,
requires_grad=self.requires_grad,
alias_decay_db=self.alias_decay_db,
device=self.device,
) # TODO add hadamard, tiny rotation
except:
raise ValueError(f"Unsupported mixing type: {config.mixing_type}")
return mixing
def _create_direct_path(self, config: BaseConfig):
"""Create the direct path branch (branch B)."""
onset_delay = dsp.parallelDelay(
size=(self.in_ch,),
max_len=self.onset,
nfft=self.nfft,
isint=True,
requires_grad=False,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
if config.early_reflections_type == "FIR":
L = self.delay_lengths.min()
early_reflections = dsp.parallelFilter(
size=(L-self.onset, self.in_ch),
nfft=self.nfft,
requires_grad=False,
map=lambda x: x,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
else:
early_reflections = dsp.Gain(
size=(self.in_ch, self.out_ch),
nfft=self.nfft,
requires_grad=False,
map=lambda x: x,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
self._configure_onset(onset_delay, early_reflections)
return system.Series(
OrderedDict(
{
"onset_delay": onset_delay,
"early_reflections": early_reflections,
}
)
)
def _configure_onset(self, onset_delay, early_reflections):
"""Configure onset behavior based on early_reflections_type."""
# Ensure onset has correct number of values
if len(self.onset) != self.in_ch:
self.onset = self.onset.repeat(self.in_ch)
if self.early_reflections_type is None:
onset_delay.assign_value(
onset_delay.sample2s(torch.zeros((self.in_ch,), device=self.device))
)
early_reflections.assign_value(torch.zeros((self.in_ch, 1)))
elif self.early_reflections_type == "gain":
onset_delay.assign_value(onset_delay.sample2s(torch.tensor(self.onset)))
early_reflections.assign_value(torch.randn((self.in_ch, 1)))
elif self.early_reflections_type == "FIR":
velvet_noise = signal_gallery(
batch_size=1,
n_samples=early_reflections.size[0],
n=self.in_ch,
signal_type="velvet",
fs=self.fs,
rate=max(int(torch.rand(1,) / 100 * self.fs), self.fs / early_reflections.size[0] + 1),
).squeeze(0)
early_reflections.assign_value(velvet_noise)
else:
raise ValueError(f"Unsupported onset type: {self.early_reflections_type}")
def _create_attenuation(self, config: FDNAttenuation):
"""Create attenuation based on configuration type."""
if config.attenuation_type == "homogeneous":
return self._create_homogeneous_attenuation(config)
elif config.attenuation_type == "geq":
return self._create_geq_attenuation(config)
elif config.attenuation_type == "first_order_lp":
return self._create_first_order_attenuation(config)
else:
raise ValueError(f"Unsupported attenuation type: {config.attenuation_type}")
def _create_homogeneous_attenuation(self, config: FDNAttenuation):
"""Create homogeneous attenuation."""
attenuation = dsp.parallelGain(
size=(self.N,),
nfft=self.nfft,
requires_grad=False,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
attenuation.map = MapGamma(self.delay_lengths)
if config.attenuation_param == None:
# Random attenuation within range
random_rt = (
torch.rand((1,), device=self.device)
* (config.attenuation_range[1] - config.attenuation_range[0])
+ config.attenuation_range[0]
)
attenuation_value = self._calculate_attenuation_value(random_rt)
else:
# Use specific attenuation parameter
attenuation_value = self._calculate_attenuation_value(
torch.tensor(config.attenuation_param, device=self.device)
)
attenuation.assign_value(attenuation_value)
return attenuation
def _calculate_attenuation_value(self, rt_value: torch.Tensor) -> torch.Tensor:
"""Calculate attenuation value from RT value."""
return 10 ** (
(rt2slope(rt_value, self.fs) * torch.ones((self.N,), device=self.device))
/ 20
)
def _create_geq_attenuation(self, config: FDNAttenuation):
"""Create GEQ-based attenuation."""
attenuation = parallelFDNAccurateGEQ(
octave_interval=config.t60_octave_interval,
nfft=self.nfft,
fs=self.fs,
delays=self.delay_lengths,
alias_decay_db=self.alias_decay_db,
start_freq=config.t60_center_freq[0],
end_freq=config.t60_center_freq[-1],
device=None,
)
attenuation.assign_value(
torch.tensor(config.attenuation_param[0], device=self.device)
)
return attenuation
def _create_first_order_attenuation(self, config: FDNAttenuation):
"""Create first-order shelving attenuation."""
attenuation = parallelFirstOrderShelving(
nfft=self.nfft,
fs=self.fs,
rt_nyquist=config.rt_nyquist,
delays=self.delay_lengths,
alias_decay_db=self.alias_decay_db,
device=self.device,
)
attenuation.assign_value(
torch.tensor(config.attenuation_param[0], device=self.device)
)
return attenuation |