File size: 11,607 Bytes
b6c45cb |
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 |
import torch
from scipy.signal import get_window
# from asteroid_test.losses import PITLossWrapper
from torch import nn
'''
class LambdaOverlapAdd(torch.nn.Module):
"""Overlap-add with lambda transform on segments.
Segment input signal, apply lambda function (a neural network for example)
and combine with OLA.
Args:
nnet (callable): Function to apply to each segment.
n_src (int): Number of sources in the output of nnet.
window_size (int): Size of segmenting window.
hop_size (int): Segmentation hop size.
window (str): Name of the window (see scipy.signal.get_window) used
for the synthesis.
reorder_chunks (bool): Whether to reorder each consecutive segment.
This might be useful when `nnet` is permutation invariant, as
source assignements might change output channel from one segment
to the next (in classic speech separation for example).
Reordering is performed based on the correlation between
the overlapped part of consecutive segment.
Examples:
>>> from asteroid_test import ConvTasNet
>>> nnet = ConvTasNet(n_src=2)
>>> continuous_nnet = LambdaOverlapAdd(
>>> nnet=nnet,
>>> n_src=2,
>>> window_size=64000,
>>> hop_size=None,
>>> window="hanning",
>>> reorder_chunks=True,
>>> enable_grad=False,
>>> )
>>> wav = torch.randn(1, 1, 500000)
>>> out_wavs = continuous_nnet.forward(wav)
"""
def __init__(
self,
nnet,
n_src,
window_size,
hop_size=None,
window="hanning",
reorder_chunks=True,
enable_grad=False,
):
super().__init__()
assert window_size % 2 == 0, "Window size must be even"
self.nnet = nnet
self.window_size = window_size
self.hop_size = hop_size if hop_size is not None else window_size // 2
self.n_src = n_src
if window:
window = get_window(window, self.window_size).astype("float32")
window = torch.from_numpy(window)
self.use_window = True
else:
self.use_window = False
self.register_buffer("window", window)
self.reorder_chunks = reorder_chunks
self.enable_grad = enable_grad
def ola_forward(self, x):
"""Heart of the class: segment signal, apply func, combine with OLA."""
assert x.ndim == 3
batch, channels, n_frames = x.size()
# Overlap and add:
# [batch, chans, n_frames] -> [batch, chans, win_size, n_chunks]
unfolded = torch.nn.functional.unfold(
x.unsqueeze(-1),
kernel_size=(self.window_size, 1),
padding=(self.window_size, 0),
stride=(self.hop_size, 1),
)
out = []
n_chunks = unfolded.shape[-1]
for frame_idx in range(n_chunks): # for loop to spare memory
frame = self.nnet(unfolded[..., frame_idx])
# user must handle multichannel by reshaping to batch
if frame_idx == 0:
assert frame.ndim == 3, "nnet should return (batch, n_src, time)"
assert frame.shape[1] == self.n_src, "nnet should return (batch, n_src, time)"
frame = frame.reshape(batch * self.n_src, -1)
if frame_idx != 0 and self.reorder_chunks:
# we determine best perm based on xcorr with previous sources
frame = _reorder_sources(
frame, out[-1], self.n_src, self.window_size, self.hop_size
)
if self.use_window:
frame = frame * self.window
else:
frame = frame / (self.window_size / self.hop_size)
out.append(frame)
out = torch.stack(out).reshape(n_chunks, batch * self.n_src, self.window_size)
out = out.permute(1, 2, 0)
out = torch.nn.functional.fold(
out,
(n_frames, 1),
kernel_size=(self.window_size, 1),
padding=(self.window_size, 0),
stride=(self.hop_size, 1),
)
return out.squeeze(-1).reshape(batch, self.n_src, -1)
def forward(self, x):
"""Forward module: segment signal, apply func, combine with OLA.
Args:
x (:class:`torch.Tensor`): waveform signal of shape (batch, 1, time).
Returns:
:class:`torch.Tensor`: The output of the lambda OLA.
"""
# Here we can do the reshaping
with torch.autograd.set_grad_enabled(self.enable_grad):
olad = self.ola_forward(x)
return olad
def _reorder_sources(
current: torch.FloatTensor,
previous: torch.FloatTensor,
n_src: int,
window_size: int,
hop_size: int,
):
"""
Reorder sources in current chunk to maximize correlation with previous chunk.
Used for Continuous Source Separation. Standard dsp correlation is used
for reordering.
Args:
current (:class:`torch.Tensor`): current chunk, tensor
of shape (batch, n_src, window_size)
previous (:class:`torch.Tensor`): previous chunk, tensor
of shape (batch, n_src, window_size)
n_src (:class:`int`): number of sources.
window_size (:class:`int`): window_size, equal to last dimension of
both current and previous.
hop_size (:class:`int`): hop_size between current and previous tensors.
Returns:
current:
"""
batch, frames = current.size()
current = current.reshape(-1, n_src, frames)
previous = previous.reshape(-1, n_src, frames)
overlap_f = window_size - hop_size
def reorder_func(x, y):
x = x[..., :overlap_f]
y = y[..., -overlap_f:]
# Mean normalization
x = x - x.mean(-1, keepdim=True)
y = y - y.mean(-1, keepdim=True)
# Negative mean Correlation
return -torch.sum(x.unsqueeze(1) * y.unsqueeze(2), dim=-1)
# We maximize correlation-like between previous and current.
pit = PITLossWrapper(reorder_func)
current = pit(current, previous, return_est=True)[1]
return current.reshape(batch, frames)
'''
class DualPathProcessing(nn.Module):
"""Perform Dual-Path processing via overlap-add as in DPRNN [1].
Args:
chunk_size (int): Size of segmenting window.
hop_size (int): segmentation hop size.
References:
[1] "Dual-path RNN: efficient long sequence modeling for
time-domain single-channel speech separation", Yi Luo, Zhuo Chen
and Takuya Yoshioka. https://arxiv.org/abs/1910.06379
"""
def __init__(self, chunk_size, hop_size):
super(DualPathProcessing, self).__init__()
self.chunk_size = chunk_size
self.hop_size = hop_size
self.n_orig_frames = None
def unfold(self, x):
"""Unfold the feature tensor from
(batch, channels, time) to (batch, channels, chunk_size, n_chunks).
Args:
x: (:class:`torch.Tensor`): feature tensor of shape (batch, channels, time).
Returns:
x: (:class:`torch.Tensor`): spliced feature tensor of shape
(batch, channels, chunk_size, n_chunks).
"""
# x is (batch, chan, frames)
batch, chan, frames = x.size()
assert x.ndim == 3
self.n_orig_frames = x.shape[-1]
unfolded = torch.nn.functional.unfold(
x.unsqueeze(-1),
kernel_size=(self.chunk_size, 1),
padding=(self.chunk_size, 0),
stride=(self.hop_size, 1),
)
return unfolded.reshape(
batch, chan, self.chunk_size, -1
) # (batch, chan, chunk_size, n_chunks)
def fold(self, x, output_size=None):
"""Folds back the spliced feature tensor.
Input shape (batch, channels, chunk_size, n_chunks) to original shape
(batch, channels, time) using overlap-add.
Args:
x: (:class:`torch.Tensor`): spliced feature tensor of shape
(batch, channels, chunk_size, n_chunks).
output_size: (int, optional): sequence length of original feature tensor.
If None, the original length cached by the previous call of `unfold`
will be used.
Returns:
x: (:class:`torch.Tensor`): feature tensor of shape (batch, channels, time).
.. note:: `fold` caches the original length of the pr
"""
output_size = output_size if output_size is not None else self.n_orig_frames
# x is (batch, chan, chunk_size, n_chunks)
batch, chan, chunk_size, n_chunks = x.size()
to_unfold = x.reshape(batch, chan * self.chunk_size, n_chunks)
x = torch.nn.functional.fold(
to_unfold,
(output_size, 1),
kernel_size=(self.chunk_size, 1),
padding=(self.chunk_size, 0),
stride=(self.hop_size, 1),
)
x /= self.chunk_size / self.hop_size
return x.reshape(batch, chan, self.n_orig_frames)
@staticmethod
def intra_process(x, module):
"""Performs intra-chunk processing.
Args:
x (:class:`torch.Tensor`): spliced feature tensor of shape
(batch, channels, chunk_size, n_chunks).
module (:class:`torch.nn.Module`): module one wish to apply to each chunk
of the spliced feature tensor.
Returns:
x (:class:`torch.Tensor`): processed spliced feature tensor of shape
(batch, channels, chunk_size, n_chunks).
.. note:: the module should have the channel first convention and accept
a 3D tensor of shape (batch, channels, time).
"""
# x is (batch, channels, chunk_size, n_chunks)
batch, channels, chunk_size, n_chunks = x.size()
# we reshape to batch*chunk_size, channels, n_chunks
x = x.transpose(1, -1).reshape(batch * n_chunks, chunk_size, channels).transpose(1, -1)
x = module(x)
x = x.reshape(batch, n_chunks, channels, chunk_size).transpose(1, -1).transpose(1, 2)
return x
@staticmethod
def inter_process(x, module):
"""Performs inter-chunk processing.
Args:
x (:class:`torch.Tensor`): spliced feature tensor of shape
(batch, channels, chunk_size, n_chunks).
module (:class:`torch.nn.Module`): module one wish to apply between
each chunk of the spliced feature tensor.
Returns:
x (:class:`torch.Tensor`): processed spliced feature tensor of shape
(batch, channels, chunk_size, n_chunks).
.. note:: the module should have the channel first convention and accept
a 3D tensor of shape (batch, channels, time).
"""
batch, channels, chunk_size, n_chunks = x.size()
x = x.transpose(1, 2).reshape(batch * chunk_size, channels, n_chunks)
x = module(x)
x = x.reshape(batch, chunk_size, channels, n_chunks).transpose(1, 2)
return x
|