File size: 11,363 Bytes
d3dbf03 |
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 |
import numpy as np
import torch
import torch.nn as nn
from mmcv.cnn import build_activation_layer
from mmengine.model import BaseModule, ModuleList, Sequential
from mmaction.models.utils import unit_tcn
from mmaction.models.utils.graph import k_adjacency, normalize_digraph
class MLP(BaseModule):
def __init__(self,
in_channels,
out_channels,
act_cfg=dict(type='ReLU'),
dropout=0):
super().__init__()
channels = [in_channels] + out_channels
self.layers = ModuleList()
for i in range(1, len(channels)):
if dropout > 1e-3:
self.layers.append(nn.Dropout(p=dropout))
self.layers.append(
nn.Conv2d(channels[i - 1], channels[i], kernel_size=1))
self.layers.append(nn.BatchNorm2d(channels[i]))
if act_cfg:
self.layers.append(build_activation_layer(act_cfg))
def forward(self, x):
for layer in self.layers:
x = layer(x)
return x
class MSGCN(BaseModule):
def __init__(self,
num_scales,
in_channels,
out_channels,
A,
dropout=0,
act_cfg=dict(type='ReLU')):
super().__init__()
self.num_scales = num_scales
A_powers = [
k_adjacency(A, k, with_self=True) for k in range(num_scales)
]
A_powers = np.stack([normalize_digraph(g) for g in A_powers])
# K, V, V
self.register_buffer('A', torch.Tensor(A_powers))
self.PA = nn.Parameter(self.A.clone())
nn.init.uniform_(self.PA, -1e-6, 1e-6)
self.mlp = MLP(
in_channels * num_scales, [out_channels],
dropout=dropout,
act_cfg=act_cfg)
def forward(self, x):
N, C, T, V = x.shape
A = self.A
A = A + self.PA
support = torch.einsum('kvu,nctv->nkctu', A, x)
support = support.reshape(N, self.num_scales * C, T, V)
out = self.mlp(support)
return out
# ! Notice: The implementation of MSTCN in
# MS-G3D is not the same as our implementation.
class MSTCN(BaseModule):
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
dilations=[1, 2, 3, 4],
residual=True,
act_cfg=dict(type='ReLU'),
init_cfg=[
dict(type='Constant', layer='BatchNorm2d', val=1),
dict(type='Kaiming', layer='Conv2d', mode='fan_out')
],
tcn_dropout=0):
super().__init__(init_cfg=init_cfg)
# Multiple branches of temporal convolution
self.num_branches = len(dilations) + 2
branch_channels = out_channels // self.num_branches
branch_channels_rem = out_channels - branch_channels * (
self.num_branches - 1)
if type(kernel_size) == list:
assert len(kernel_size) == len(dilations)
else:
kernel_size = [kernel_size] * len(dilations)
self.branches = ModuleList([
Sequential(
nn.Conv2d(
in_channels, branch_channels, kernel_size=1, padding=0),
nn.BatchNorm2d(branch_channels),
build_activation_layer(act_cfg),
unit_tcn(
branch_channels,
branch_channels,
kernel_size=ks,
stride=stride,
dilation=dilation),
) for ks, dilation in zip(kernel_size, dilations)
])
# Additional Max & 1x1 branch
self.branches.append(
Sequential(
nn.Conv2d(
in_channels, branch_channels, kernel_size=1, padding=0),
nn.BatchNorm2d(branch_channels),
build_activation_layer(act_cfg),
nn.MaxPool2d(
kernel_size=(3, 1), stride=(stride, 1), padding=(1, 0)),
nn.BatchNorm2d(branch_channels)))
self.branches.append(
Sequential(
nn.Conv2d(
in_channels,
branch_channels_rem,
kernel_size=1,
padding=0,
stride=(stride, 1)), nn.BatchNorm2d(branch_channels_rem)))
# Residual connection
if not residual:
self.residual = lambda x: 0
elif (in_channels == out_channels) and (stride == 1):
self.residual = lambda x: x
else:
self.residual = unit_tcn(
in_channels, out_channels, kernel_size=1, stride=stride)
self.act = build_activation_layer(act_cfg)
self.drop = nn.Dropout(tcn_dropout)
def forward(self, x):
# Input dim: (N,C,T,V)
res = self.residual(x)
branch_outs = []
for tempconv in self.branches:
out = tempconv(x)
branch_outs.append(out)
out = torch.cat(branch_outs, dim=1)
out += res
out = self.act(out)
out = self.drop(out)
return out
class UnfoldTemporalWindows(BaseModule):
def __init__(self, window_size, window_stride, window_dilation=1):
super().__init__()
self.window_size = window_size
self.window_stride = window_stride
self.window_dilation = window_dilation
self.padding = (window_size + (window_size - 1) *
(window_dilation - 1) - 1) // 2
self.unfold = nn.Unfold(
kernel_size=(self.window_size, 1),
dilation=(self.window_dilation, 1),
stride=(self.window_stride, 1),
padding=(self.padding, 0))
def forward(self, x):
# Input shape: (N,C,T,V), out: (N,C,T,V*window_size)
N, C, T, V = x.shape
x = self.unfold(x)
# Permute extra channels from window size to the graph dimension;
# -1 for number of windows
x = x.reshape(N, C, self.window_size, -1, V).permute(0, 1, 3, 2,
4).contiguous()
x = x.reshape(N, C, -1, self.window_size * V)
return x
class ST_MSGCN(BaseModule):
def __init__(self,
in_channels,
out_channels,
A,
num_scales,
window_size,
residual=False,
dropout=0,
act_cfg=dict(type='ReLU')):
super().__init__()
self.num_scales = num_scales
self.window_size = window_size
A = self.build_st_graph(A, window_size)
A_scales = [
k_adjacency(A, k, with_self=True) for k in range(num_scales)
]
A_scales = np.stack([normalize_digraph(g) for g in A_scales])
self.register_buffer('A', torch.Tensor(A_scales))
self.V = len(A)
self.PA = nn.Parameter(self.A.clone())
nn.init.uniform_(self.PA, -1e-6, 1e-6)
self.mlp = MLP(
in_channels * num_scales, [out_channels],
dropout=dropout,
act_cfg=act_cfg)
# Residual connection
if not residual:
self.residual = lambda x: 0
elif (in_channels == out_channels):
self.residual = lambda x: x
else:
self.residual = MLP(in_channels, [out_channels], act_cfg=None)
self.act = build_activation_layer(act_cfg)
def build_st_graph(self, A, window_size):
if not isinstance(A, np.ndarray):
A = A.data.cpu().numpy()
assert len(A.shape) == 2 and A.shape[0] == A.shape[1]
V = len(A)
A_with_I = A + np.eye(V, dtype=A.dtype)
A_large = np.tile(A_with_I, (window_size, window_size)).copy()
return A_large
def forward(self, x):
N, C, T, V = x.shape # T = number of windows, V = self.V * window_size
A = self.A + self.PA
# Perform Graph Convolution
res = self.residual(x)
agg = torch.einsum('kvu,nctv->nkctu', A, x)
agg = agg.reshape(N, self.num_scales * C, T, V)
out = self.mlp(agg)
if res == 0:
return self.act(out)
else:
return self.act(out + res)
class MSG3DBlock(BaseModule):
def __init__(self,
in_channels,
out_channels,
A,
num_scales,
window_size,
window_stride,
window_dilation,
embed_factor=1,
activation='relu'):
super().__init__()
self.window_size = window_size
self.out_channels = out_channels
self.embed_channels_in = out_channels // embed_factor
self.embed_channels_out = out_channels // embed_factor
if embed_factor == 1:
self.in1x1 = nn.Identity()
self.embed_channels_in = self.embed_channels_out = in_channels
# The first STGC block changes channels right away;
# others change at collapse
if in_channels == 3:
self.embed_channels_out = out_channels
else:
self.in1x1 = MLP(in_channels, [self.embed_channels_in])
self.gcn3d = Sequential(
UnfoldTemporalWindows(window_size, window_stride, window_dilation),
ST_MSGCN(
in_channels=self.embed_channels_in,
out_channels=self.embed_channels_out,
A=A,
num_scales=num_scales,
window_size=window_size))
self.out_conv = nn.Conv3d(
self.embed_channels_out,
out_channels,
kernel_size=(1, self.window_size, 1))
self.out_bn = nn.BatchNorm2d(out_channels)
def forward(self, x):
N, _, T, V = x.shape
x = self.in1x1(x)
# Construct temporal windows and apply MS-GCN
x = self.gcn3d(x)
# Collapse the window dimension
x = x.reshape(N, self.embed_channels_out, -1, self.window_size, V)
x = self.out_conv(x).squeeze(dim=3)
x = self.out_bn(x)
# no activation
return x
class MW_MSG3DBlock(BaseModule):
def __init__(self,
in_channels,
out_channels,
A,
num_scales,
window_sizes=[3, 5],
window_stride=1,
window_dilations=[1, 1]):
super().__init__()
self.gcn3d = ModuleList([
MSG3DBlock(in_channels, out_channels, A, num_scales, window_size,
window_stride, window_dilation) for window_size,
window_dilation in zip(window_sizes, window_dilations)
])
def forward(self, x):
out_sum = 0
for gcn3d in self.gcn3d:
out_sum += gcn3d(x)
return out_sum
|