File size: 8,529 Bytes
3978e51 |
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 |
from typing import Dict, List, Optional, Tuple, Type
import torch
from torch import nn
from torch.nn.modules import activation
from torch.utils.checkpoint import checkpoint_sequential
from .utils import (
band_widths_from_specs,
check_no_gap,
check_no_overlap,
check_nonzero_bandwidth,
)
class BaseNormMLP(nn.Module):
def __init__(
self,
emb_dim: int,
mlp_dim: int,
bandwidth: int,
in_channels: Optional[int],
hidden_activation: str = "Tanh",
hidden_activation_kwargs=None,
complex_mask: bool = True,
):
super().__init__()
if hidden_activation_kwargs is None:
hidden_activation_kwargs = {}
self.hidden_activation_kwargs = hidden_activation_kwargs
self.norm = nn.LayerNorm(emb_dim)
self.hidden = nn.Sequential(
nn.Linear(in_features=emb_dim, out_features=mlp_dim),
activation.__dict__[hidden_activation](**self.hidden_activation_kwargs),
)
self.bandwidth = bandwidth
self.in_channels = in_channels
self.complex_mask = complex_mask
self.reim = 2 if complex_mask else 1
self.glu_mult = 2
class NormMLP(BaseNormMLP):
def __init__(
self,
emb_dim: int,
mlp_dim: int,
bandwidth: int,
in_channels: Optional[int],
hidden_activation: str = "Tanh",
hidden_activation_kwargs=None,
complex_mask: bool = True,
) -> None:
super().__init__(
emb_dim=emb_dim,
mlp_dim=mlp_dim,
bandwidth=bandwidth,
in_channels=in_channels,
hidden_activation=hidden_activation,
hidden_activation_kwargs=hidden_activation_kwargs,
complex_mask=complex_mask,
)
self.output = nn.Sequential(
nn.Linear(
in_features=mlp_dim,
out_features=bandwidth * in_channels * self.reim * 2,
),
nn.GLU(dim=-1),
)
try:
self.combined = torch.compile(
nn.Sequential(self.norm, self.hidden, self.output), disable=True
)
except Exception as e:
self.combined = nn.Sequential(self.norm, self.hidden, self.output)
def reshape_output(self, mb):
# print(mb.shape)
batch, n_time, _ = mb.shape
if self.complex_mask:
mb = mb.reshape(
batch, n_time, self.in_channels, self.bandwidth, self.reim
).contiguous()
# print(mb.shape)
mb = torch.view_as_complex(mb) # (batch, n_time, in_channels, bandwidth)
else:
mb = mb.reshape(batch, n_time, self.in_channels, self.bandwidth)
mb = torch.permute(mb, (0, 2, 3, 1)) # (batch, in_channels, bandwidth, n_time)
return mb
def forward(self, qb):
# qb = (batch, n_time, emb_dim)
# qb = self.norm(qb) # (batch, n_time, emb_dim)
# qb = self.hidden(qb) # (batch, n_time, mlp_dim)
# mb = self.output(qb) # (batch, n_time, bandwidth * in_channels * reim)
mb = checkpoint_sequential(self.combined, 2, qb, use_reentrant=False)
mb = self.reshape_output(mb) # (batch, in_channels, bandwidth, n_time)
return mb
class MaskEstimationModuleSuperBase(nn.Module):
pass
class MaskEstimationModuleBase(MaskEstimationModuleSuperBase):
def __init__(
self,
band_specs: List[Tuple[float, float]],
emb_dim: int,
mlp_dim: int,
in_channels: Optional[int],
hidden_activation: str = "Tanh",
hidden_activation_kwargs: Dict = None,
complex_mask: bool = True,
norm_mlp_cls: Type[nn.Module] = NormMLP,
norm_mlp_kwargs: Dict = None,
) -> None:
super().__init__()
self.band_widths = band_widths_from_specs(band_specs)
self.n_bands = len(band_specs)
if hidden_activation_kwargs is None:
hidden_activation_kwargs = {}
if norm_mlp_kwargs is None:
norm_mlp_kwargs = {}
self.norm_mlp = nn.ModuleList(
[
norm_mlp_cls(
bandwidth=self.band_widths[b],
emb_dim=emb_dim,
mlp_dim=mlp_dim,
in_channels=in_channels,
hidden_activation=hidden_activation,
hidden_activation_kwargs=hidden_activation_kwargs,
complex_mask=complex_mask,
**norm_mlp_kwargs,
)
for b in range(self.n_bands)
]
)
def compute_masks(self, q):
batch, n_bands, n_time, emb_dim = q.shape
masks = []
for b, nmlp in enumerate(self.norm_mlp):
# print(f"maskestim/{b:02d}")
qb = q[:, b, :, :]
mb = nmlp(qb)
masks.append(mb)
return masks
def compute_mask(self, q, b):
batch, n_bands, n_time, emb_dim = q.shape
qb = q[:, b, :, :]
mb = self.norm_mlp[b](qb)
return mb
class OverlappingMaskEstimationModule(MaskEstimationModuleBase):
def __init__(
self,
in_channels: int,
band_specs: List[Tuple[float, float]],
freq_weights: List[torch.Tensor],
n_freq: int,
emb_dim: int,
mlp_dim: int,
cond_dim: int = 0,
hidden_activation: str = "Tanh",
hidden_activation_kwargs: Dict = None,
complex_mask: bool = True,
norm_mlp_cls: Type[nn.Module] = NormMLP,
norm_mlp_kwargs: Dict = None,
use_freq_weights: bool = False,
) -> None:
check_nonzero_bandwidth(band_specs)
check_no_gap(band_specs)
if cond_dim > 0:
raise NotImplementedError
super().__init__(
band_specs=band_specs,
emb_dim=emb_dim + cond_dim,
mlp_dim=mlp_dim,
in_channels=in_channels,
hidden_activation=hidden_activation,
hidden_activation_kwargs=hidden_activation_kwargs,
complex_mask=complex_mask,
norm_mlp_cls=norm_mlp_cls,
norm_mlp_kwargs=norm_mlp_kwargs,
)
self.n_freq = n_freq
self.band_specs = band_specs
self.in_channels = in_channels
if freq_weights is not None and use_freq_weights:
for i, fw in enumerate(freq_weights):
self.register_buffer(f"freq_weights/{i}", fw)
self.use_freq_weights = use_freq_weights
else:
self.use_freq_weights = False
def forward(self, q):
# q = (batch, n_bands, n_time, emb_dim)
batch, n_bands, n_time, emb_dim = q.shape
masks = torch.zeros(
(batch, self.in_channels, self.n_freq, n_time),
device=q.device,
dtype=torch.complex64,
)
for im in range(n_bands):
fstart, fend = self.band_specs[im]
mask = self.compute_mask(q, im)
if self.use_freq_weights:
fw = self.get_buffer(f"freq_weights/{im}")[:, None]
mask = mask * fw
masks[:, :, fstart:fend, :] += mask
return masks
class MaskEstimationModule(OverlappingMaskEstimationModule):
def __init__(
self,
band_specs: List[Tuple[float, float]],
emb_dim: int,
mlp_dim: int,
in_channels: Optional[int],
hidden_activation: str = "Tanh",
hidden_activation_kwargs: Dict = None,
complex_mask: bool = True,
**kwargs,
) -> None:
check_nonzero_bandwidth(band_specs)
check_no_gap(band_specs)
check_no_overlap(band_specs)
super().__init__(
in_channels=in_channels,
band_specs=band_specs,
freq_weights=None,
n_freq=None,
emb_dim=emb_dim,
mlp_dim=mlp_dim,
hidden_activation=hidden_activation,
hidden_activation_kwargs=hidden_activation_kwargs,
complex_mask=complex_mask,
)
def forward(self, q, cond=None):
# q = (batch, n_bands, n_time, emb_dim)
masks = self.compute_masks(
q
) # [n_bands * (batch, in_channels, bandwidth, n_time)]
# TODO: currently this requires band specs to have no gap and no overlap
masks = torch.concat(masks, dim=2) # (batch, in_channels, n_freq, n_time)
return masks
|