Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,857 Bytes
8e8cd3e |
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 |
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchaudio
import math
from .mossformer.utils.one_path_flash_fsmn import Dual_Path_Model, SBFLASHBlock_DualA
from models.av_mossformer2_tse.visual_frontend import VisualFrontend
EPS = 1e-8
class avMossformer(nn.Module):
def __init__(self, args):
super(avMossformer, self).__init__()
N, L, = args.network_audio.encoder_out_nchannels, args.network_audio.encoder_kernel_size
self.encoder = Encoder(L, N)
self.separator = Separator(args)
self.decoder = Decoder(args, N, L)
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_normal_(p)
def forward(self, mixture, visual):
"""
Args:
mixture: [M, T], M is batch size, T is #samples
Returns:
est_source: [M, C, T]
"""
mixture_w = self.encoder(mixture)
est_mask = self.separator(mixture_w, visual)
est_source = self.decoder(mixture_w, est_mask)
# T changed after conv1d in encoder, fix it here
T_origin = mixture.size(-1)
T_conv = est_source.size(-1)
est_source = F.pad(est_source, (0, T_origin - T_conv))
return est_source
class Encoder(nn.Module):
def __init__(self, L, N):
super(Encoder, self).__init__()
self.L, self.N = L, N
self.conv1d_U = nn.Conv1d(1, N, kernel_size=L, stride=L // 2, bias=False)
def forward(self, mixture):
"""
Args:
mixture: [M, T], M is batch size, T is #samples
Returns:
mixture_w: [M, N, K], where K = (T-L)/(L/2)+1 = 2T/L-1
"""
mixture = torch.unsqueeze(mixture, 1) # [M, 1, T]
mixture_w = F.relu(self.conv1d_U(mixture)) # [M, N, K]
return mixture_w
class Decoder(nn.Module):
def __init__(self, args, N, L):
super(Decoder, self).__init__()
self.N, self.L, self.args = N, L, args
self.basis_signals = nn.Linear(N, L, bias=False)
def forward(self, mixture_w, est_mask):
"""
Args:
mixture_w: [M, N, K]
est_mask: [M, C, N, K]
Returns:
est_source: [M, C, T]
"""
est_source = mixture_w * est_mask
est_source = torch.transpose(est_source, 2, 1) # [M, K, N]
est_source = self.basis_signals(est_source) # [M, K, L]
est_source = overlap_and_add(est_source, self.L//2) # M x C x T
return est_source
class Separator(nn.Module):
def __init__(self, args):
super(Separator, self).__init__()
self.layer_norm = nn.GroupNorm(1, args.network_audio.encoder_out_nchannels, eps=1e-8)
self.bottleneck_conv1x1 = nn.Conv1d(args.network_audio.encoder_out_nchannels, args.network_audio.encoder_out_nchannels, 1, bias=False)
# mossformer 2
intra_model = SBFLASHBlock_DualA(
num_layers=args.network_audio.intra_numlayers,
d_model=args.network_audio.encoder_out_nchannels,
nhead=args.network_audio.intra_nhead,
d_ffn=args.network_audio.intra_dffn,
dropout=args.network_audio.intra_dropout,
use_positional_encoding=args.network_audio.intra_use_positional,
norm_before=args.network_audio.intra_norm_before
)
self.masknet = Dual_Path_Model(
in_channels=args.network_audio.encoder_out_nchannels,
out_channels=args.network_audio.encoder_out_nchannels,
intra_model=intra_model,
num_layers=args.network_audio.masknet_numlayers,
norm=args.network_audio.masknet_norm,
K=args.network_audio.masknet_chunksize,
num_spks=args.network_audio.masknet_numspks,
skip_around_intra=args.network_audio.masknet_extraskipconnection,
linear_layer_after_inter_intra=args.network_audio.masknet_useextralinearlayer
)
# reference
# visual
stacks = []
for x in range(5):
stacks +=[VisualConv1D(V=256, H=512)]
self.visual_conv = nn.Sequential(*stacks)
self.v_ds = nn.Conv1d(512, 256, 1, bias=False)
self.av_conv = nn.Conv1d(args.network_audio.encoder_out_nchannels+args.network_reference.emb_size, args.network_audio.encoder_out_nchannels, 1, bias=True)
def forward(self, x, visual):
"""
Keep this API same with TasNet
Args:
mixture_w: [M, N, K], M is batch size
returns:
est_mask: [M, C, N, K]
"""
M, N, D = x.size()
x = self.layer_norm(x)
x = self.bottleneck_conv1x1(x)
visual = visual.transpose(1,2)
visual = self.v_ds(visual)
visual = self.visual_conv(visual)
visual = F.interpolate(visual, (D), mode='linear')
x = torch.cat((x, visual),1)
x = self.av_conv(x)
x = self.masknet(x)
x = x.squeeze(0)
return x
def overlap_and_add(signal, frame_step):
"""Reconstructs a signal from a framed representation.
Adds potentially overlapping frames of a signal with shape
`[..., frames, frame_length]`, offsetting subsequent frames by `frame_step`.
The resulting tensor has shape `[..., output_size]` where
output_size = (frames - 1) * frame_step + frame_length
Args:
signal: A [..., frames, frame_length] Tensor. All dimensions may be unknown, and rank must be at least 2.
frame_step: An integer denoting overlap offsets. Must be less than or equal to frame_length.
Returns:
A Tensor with shape [..., output_size] containing the overlap-added frames of signal's inner-most two dimensions.
output_size = (frames - 1) * frame_step + frame_length
Based on https://github.com/tensorflow/tensorflow/blob/r1.12/tensorflow/contrib/signal/python/ops/reconstruction_ops.py
"""
outer_dimensions = signal.size()[:-2]
frames, frame_length = signal.size()[-2:]
subframe_length = math.gcd(frame_length, frame_step) # gcd=Greatest Common Divisor
subframe_step = frame_step // subframe_length
subframes_per_frame = frame_length // subframe_length
output_size = frame_step * (frames - 1) + frame_length
output_subframes = output_size // subframe_length
subframe_signal = signal.view(*outer_dimensions, -1, subframe_length)
frame = torch.arange(0, output_subframes).unfold(0, subframes_per_frame, subframe_step)
frame = signal.new_tensor(frame).long().cuda() # signal may in GPU or CPU
frame = frame.contiguous().view(-1)
result = signal.new_zeros(*outer_dimensions, output_subframes, subframe_length)
result.index_add_(-2, frame, subframe_signal)
result = result.view(*outer_dimensions, -1)
return result
class av_mossformer_tmp(nn.Module):
def __init__(self, args):
super(av_mossformer_tmp, self).__init__()
args.causal=0
self.sep_network = avMossformer(args)
self.v_front_end = VisualFrontend(args)
def forward(self, mixture, ref):
ref = self.v_front_end(ref.unsqueeze(1)).transpose(1,2)
return self.sep_network(mixture, ref)
class VisualConv1D(nn.Module):
def __init__(self, V=256, H=512):
super(VisualConv1D, self).__init__()
relu_0 = nn.ReLU()
norm_0 = GlobalLayerNorm(V)
conv1x1 = nn.Conv1d(V, H, 1, bias=False)
relu = nn.ReLU()
norm_1 = GlobalLayerNorm(H)
dsconv = nn.Conv1d(H, H, 3, stride=1, padding=1,dilation=1, groups=H, bias=False)
prelu = nn.PReLU()
norm_2 = GlobalLayerNorm(H)
pw_conv = nn.Conv1d(H, V, 1, bias=False)
self.net = nn.Sequential(relu_0, norm_0, conv1x1, relu, norm_1 ,dsconv, prelu, norm_2, pw_conv)
def forward(self, x):
out = self.net(x)
return out + x
class GlobalLayerNorm(nn.Module):
"""Global Layer Normalization (gLN)"""
def __init__(self, channel_size):
super(GlobalLayerNorm, self).__init__()
self.gamma = nn.Parameter(torch.Tensor(1, channel_size, 1)) # [1, N, 1]
self.beta = nn.Parameter(torch.Tensor(1, channel_size,1 )) # [1, N, 1]
self.reset_parameters()
def reset_parameters(self):
self.gamma.data.fill_(1)
self.beta.data.zero_()
def forward(self, y):
"""
Args:
y: [M, N, K], M is batch size, N is channel size, K is length
Returns:
gLN_y: [M, N, K]
"""
# TODO: in torch 1.0, torch.mean() support dim list
mean = y.mean(dim=1, keepdim=True).mean(dim=2, keepdim=True) #[M, 1, 1]
var = (torch.pow(y-mean, 2)).mean(dim=1, keepdim=True).mean(dim=2, keepdim=True)
gLN_y = self.gamma * (y - mean) / torch.pow(var + EPS, 0.5) + self.beta
return gLN_y
|