Spaces:
Running
Running
File size: 13,902 Bytes
27763e5 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import numpy as np
class SkeletonConv(nn.Module):
def __init__(self, neighbour_list, in_channels, out_channels, kernel_size, joint_num, stride=1, padding=0,
bias=True, padding_mode='zeros', add_offset=False, in_offset_channel=0):
super(SkeletonConv, self).__init__()
if in_channels % joint_num != 0 or out_channels % joint_num != 0:
raise Exception('in/out channels should be divided by joint_num')
self.in_channels_per_joint = in_channels // joint_num
self.out_channels_per_joint = out_channels // joint_num
if padding_mode == 'zeros': padding_mode = 'constant'
self.expanded_neighbour_list = []
self.expanded_neighbour_list_offset = []
self.neighbour_list = neighbour_list
self.add_offset = add_offset
self.joint_num = joint_num
self.stride = stride
self.dilation = 1
self.groups = 1
self.padding = padding
self.padding_mode = padding_mode
self._padding_repeated_twice = (padding, padding)
for neighbour in neighbour_list:
expanded = []
for k in neighbour:
for i in range(self.in_channels_per_joint):
expanded.append(k * self.in_channels_per_joint + i)
self.expanded_neighbour_list.append(expanded)
if self.add_offset:
self.offset_enc = SkeletonLinear(neighbour_list, in_offset_channel * len(neighbour_list), out_channels)
for neighbour in neighbour_list:
expanded = []
for k in neighbour:
for i in range(add_offset):
expanded.append(k * in_offset_channel + i)
self.expanded_neighbour_list_offset.append(expanded)
self.weight = torch.zeros(out_channels, in_channels, kernel_size)
if bias:
self.bias = torch.zeros(out_channels)
else:
self.register_parameter('bias', None)
self.mask = torch.zeros_like(self.weight)
for i, neighbour in enumerate(self.expanded_neighbour_list):
self.mask[self.out_channels_per_joint * i: self.out_channels_per_joint * (i + 1), neighbour, ...] = 1
self.mask = nn.Parameter(self.mask, requires_grad=False)
self.description = 'SkeletonConv(in_channels_per_armature={}, out_channels_per_armature={}, kernel_size={}, ' \
'joint_num={}, stride={}, padding={}, bias={})'.format(
in_channels // joint_num, out_channels // joint_num, kernel_size, joint_num, stride, padding, bias
)
self.reset_parameters()
def reset_parameters(self):
for i, neighbour in enumerate(self.expanded_neighbour_list):
""" Use temporary variable to avoid assign to copy of slice, which might lead to un expected result """
tmp = torch.zeros_like(self.weight[self.out_channels_per_joint * i: self.out_channels_per_joint * (i + 1),
neighbour, ...])
nn.init.kaiming_uniform_(tmp, a=math.sqrt(5))
self.weight[self.out_channels_per_joint * i: self.out_channels_per_joint * (i + 1),
neighbour, ...] = tmp
if self.bias is not None:
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(
self.weight[self.out_channels_per_joint * i: self.out_channels_per_joint * (i + 1), neighbour, ...])
bound = 1 / math.sqrt(fan_in)
tmp = torch.zeros_like(
self.bias[self.out_channels_per_joint * i: self.out_channels_per_joint * (i + 1)])
nn.init.uniform_(tmp, -bound, bound)
self.bias[self.out_channels_per_joint * i: self.out_channels_per_joint * (i + 1)] = tmp
self.weight = nn.Parameter(self.weight)
if self.bias is not None:
self.bias = nn.Parameter(self.bias)
def set_offset(self, offset):
if not self.add_offset: raise Exception('Wrong Combination of Parameters')
self.offset = offset.reshape(offset.shape[0], -1)
def forward(self, input):
weight_masked = self.weight * self.mask
res = F.conv1d(F.pad(input, self._padding_repeated_twice, mode=self.padding_mode),
weight_masked, self.bias, self.stride,
0, self.dilation, self.groups)
if self.add_offset:
offset_res = self.offset_enc(self.offset)
offset_res = offset_res.reshape(offset_res.shape + (1, ))
res += offset_res / 100
return res
def __repr__(self):
return self.description
class SkeletonLinear(nn.Module):
def __init__(self, neighbour_list, in_channels, out_channels, extra_dim1=False):
super(SkeletonLinear, self).__init__()
self.neighbour_list = neighbour_list
self.in_channels = in_channels
self.out_channels = out_channels
self.in_channels_per_joint = in_channels // len(neighbour_list)
self.out_channels_per_joint = out_channels // len(neighbour_list)
self.extra_dim1 = extra_dim1
self.expanded_neighbour_list = []
for neighbour in neighbour_list:
expanded = []
for k in neighbour:
for i in range(self.in_channels_per_joint):
expanded.append(k * self.in_channels_per_joint + i)
self.expanded_neighbour_list.append(expanded)
self.weight = torch.zeros(out_channels, in_channels)
self.mask = torch.zeros(out_channels, in_channels)
self.bias = nn.Parameter(torch.Tensor(out_channels))
self.reset_parameters()
def reset_parameters(self):
for i, neighbour in enumerate(self.expanded_neighbour_list):
tmp = torch.zeros_like(
self.weight[i*self.out_channels_per_joint: (i + 1)*self.out_channels_per_joint, neighbour]
)
self.mask[i*self.out_channels_per_joint: (i + 1)*self.out_channels_per_joint, neighbour] = 1
nn.init.kaiming_uniform_(tmp, a=math.sqrt(5))
self.weight[i*self.out_channels_per_joint: (i + 1)*self.out_channels_per_joint, neighbour] = tmp
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self.bias, -bound, bound)
self.weight = nn.Parameter(self.weight)
self.mask = nn.Parameter(self.mask, requires_grad=False)
def forward(self, input):
input = input.reshape(input.shape[0], -1)
weight_masked = self.weight * self.mask
res = F.linear(input, weight_masked, self.bias)
if self.extra_dim1: res = res.reshape(res.shape + (1,))
return res
class SkeletonPoolJoint(nn.Module):
def __init__(self, topology, pooling_mode, channels_per_joint, last_pool=False):
super(SkeletonPoolJoint, self).__init__()
if pooling_mode != 'mean':
raise Exception('Unimplemented pooling mode in matrix_implementation')
self.joint_num = len(topology)
self.parent = topology
self.pooling_list = []
self.pooling_mode = pooling_mode
self.pooling_map = [-1 for _ in range(len(self.parent))]
self.child = [-1 for _ in range(len(self.parent))]
children_cnt = [0 for _ in range(len(self.parent))]
for x, pa in enumerate(self.parent):
if pa < 0: continue
children_cnt[pa] += 1
self.child[pa] = x
self.pooling_map[0] = 0
for x in range(len(self.parent)):
if children_cnt[x] == 0 or (children_cnt[x] == 1 and children_cnt[self.child[x]] > 1):
while children_cnt[x] <= 1:
pa = self.parent[x]
if last_pool:
seq = [x]
while pa != -1 and children_cnt[pa] == 1:
seq = [pa] + seq
x = pa
pa = self.parent[x]
self.pooling_list.append(seq)
break
else:
if pa != -1 and children_cnt[pa] == 1:
self.pooling_list.append([pa, x])
x = self.parent[pa]
else:
self.pooling_list.append([x, ])
break
elif children_cnt[x] > 1:
self.pooling_list.append([x, ])
self.description = 'SkeletonPool(in_joint_num={}, out_joint_num={})'.format(
len(topology), len(self.pooling_list),
)
self.pooling_list.sort(key=lambda x:x[0])
for i, a in enumerate(self.pooling_list):
for j in a:
self.pooling_map[j] = i
self.output_joint_num = len(self.pooling_list)
self.new_topology = [-1 for _ in range(len(self.pooling_list))]
for i, x in enumerate(self.pooling_list):
if i < 1: continue
self.new_topology[i] = self.pooling_map[self.parent[x[0]]]
self.weight = torch.zeros(len(self.pooling_list) * channels_per_joint, self.joint_num * channels_per_joint)
for i, pair in enumerate(self.pooling_list):
for j in pair:
for c in range(channels_per_joint):
self.weight[i * channels_per_joint + c, j * channels_per_joint + c] = 1.0 / len(pair)
self.weight = nn.Parameter(self.weight, requires_grad=False)
def forward(self, input: torch.Tensor):
return torch.matmul(self.weight, input.unsqueeze(-1)).squeeze(-1)
class SkeletonPool(nn.Module):
def __init__(self, edges, pooling_mode, channels_per_edge, last_pool=False):
super(SkeletonPool, self).__init__()
if pooling_mode != 'mean':
raise Exception('Unimplemented pooling mode in matrix_implementation')
self.channels_per_edge = channels_per_edge
self.pooling_mode = pooling_mode
self.edge_num = len(edges) + 1
self.seq_list = []
self.pooling_list = []
self.new_edges = []
degree = [0] * 100
for edge in edges:
degree[edge[0]] += 1
degree[edge[1]] += 1
def find_seq(j, seq):
nonlocal self, degree, edges
if degree[j] > 2 and j != 0:
self.seq_list.append(seq)
seq = []
if degree[j] == 1:
self.seq_list.append(seq)
return
for idx, edge in enumerate(edges):
if edge[0] == j:
find_seq(edge[1], seq + [idx])
find_seq(0, [])
for seq in self.seq_list:
if last_pool:
self.pooling_list.append(seq)
continue
if len(seq) % 2 == 1:
self.pooling_list.append([seq[0]])
self.new_edges.append(edges[seq[0]])
seq = seq[1:]
for i in range(0, len(seq), 2):
self.pooling_list.append([seq[i], seq[i + 1]])
self.new_edges.append([edges[seq[i]][0], edges[seq[i + 1]][1]])
# add global position
self.pooling_list.append([self.edge_num - 1])
self.description = 'SkeletonPool(in_edge_num={}, out_edge_num={})'.format(
len(edges), len(self.pooling_list)
)
self.weight = torch.zeros(len(self.pooling_list) * channels_per_edge, self.edge_num * channels_per_edge)
for i, pair in enumerate(self.pooling_list):
for j in pair:
for c in range(channels_per_edge):
self.weight[i * channels_per_edge + c, j * channels_per_edge + c] = 1.0 / len(pair)
self.weight = nn.Parameter(self.weight, requires_grad=False)
def forward(self, input: torch.Tensor):
return torch.matmul(self.weight, input)
class SkeletonUnpool(nn.Module):
def __init__(self, pooling_list, channels_per_edge):
super(SkeletonUnpool, self).__init__()
self.pooling_list = pooling_list
self.input_joint_num = len(pooling_list)
self.output_joint_num = 0
self.channels_per_edge = channels_per_edge
for t in self.pooling_list:
self.output_joint_num += len(t)
self.description = 'SkeletonUnpool(in_joint_num={}, out_joint_num={})'.format(
self.input_joint_num, self.output_joint_num,
)
self.weight = torch.zeros(self.output_joint_num * channels_per_edge, self.input_joint_num * channels_per_edge)
for i, pair in enumerate(self.pooling_list):
for j in pair:
for c in range(channels_per_edge):
self.weight[j * channels_per_edge + c, i * channels_per_edge + c] = 1
self.weight = nn.Parameter(self.weight)
self.weight.requires_grad_(False)
def forward(self, input: torch.Tensor):
return torch.matmul(self.weight, input.unsqueeze(-1)).squeeze(-1)
def find_neighbor_joint(parents, threshold):
n_joint = len(parents)
dist_mat = np.empty((n_joint, n_joint), dtype=np.int)
dist_mat[:, :] = 100000
for i, p in enumerate(parents):
dist_mat[i, i] = 0
if i != 0:
dist_mat[i, p] = dist_mat[p, i] = 1
"""
Floyd's algorithm
"""
for k in range(n_joint):
for i in range(n_joint):
for j in range(n_joint):
dist_mat[i, j] = min(dist_mat[i, j], dist_mat[i, k] + dist_mat[k, j])
neighbor_list = []
for i in range(n_joint):
neighbor = []
for j in range(n_joint):
if dist_mat[i, j] <= threshold:
neighbor.append(j)
neighbor_list.append(neighbor)
return neighbor_list
|