Spaces:
Running
Running
File size: 14,032 Bytes
851751e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 |
#!/usr/bin/env python3
# This file is covered by the LICENSE file in the root of this project.
from collections import OrderedDict
import torch
import torch.nn as nn
import torch.nn.functional as F
class BasicBlock(nn.Module):
def __init__(self, inplanes, planes, bn_d=0.1):
super(BasicBlock, self).__init__()
self.conv1 = nn.Conv2d(inplanes, planes[0], kernel_size=1,
stride=1, padding=0, bias=False)
self.bn1 = nn.BatchNorm2d(planes[0], momentum=bn_d)
self.relu1 = nn.LeakyReLU(0.1)
self.conv2 = nn.Conv2d(planes[0], planes[1], kernel_size=3,
stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes[1], momentum=bn_d)
self.relu2 = nn.LeakyReLU(0.1)
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu1(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu2(out)
out += residual
return out
# ******************************************************************************
# number of layers per model
model_blocks = {
21: [1, 1, 2, 2, 1],
53: [1, 2, 8, 8, 4],
}
class Backbone(nn.Module):
"""
Class for DarknetSeg. Subclasses PyTorch's own "nn" module
"""
def __init__(self, params):
super(Backbone, self).__init__()
self.use_range = params["input_depth"]["range"]
self.use_xyz = params["input_depth"]["xyz"]
self.use_remission = params["input_depth"]["remission"]
self.drop_prob = params["dropout"]
self.bn_d = params["bn_d"]
self.OS = params["OS"]
self.layers = params["extra"]["layers"]
# input depth calc
self.input_depth = 0
self.input_idxs = []
if self.use_range:
self.input_depth += 1
self.input_idxs.append(0)
if self.use_xyz:
self.input_depth += 3
self.input_idxs.extend([1, 2, 3])
if self.use_remission:
self.input_depth += 1
self.input_idxs.append(4)
# stride play
self.strides = [2, 2, 2, 2, 2]
# check current stride
current_os = 1
for s in self.strides:
current_os *= s
# make the new stride
if self.OS > current_os:
print("Can't do OS, ", self.OS,
" because it is bigger than original ", current_os)
else:
# redo strides according to needed stride
for i, stride in enumerate(reversed(self.strides), 0):
if int(current_os) != self.OS:
if stride == 2:
current_os /= 2
self.strides[-1 - i] = 1
if int(current_os) == self.OS:
break
# check that darknet exists
assert self.layers in model_blocks.keys()
# generate layers depending on darknet type
self.blocks = model_blocks[self.layers]
# input layer
self.conv1 = nn.Conv2d(self.input_depth, 32, kernel_size=3,
stride=1, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(32, momentum=self.bn_d)
self.relu1 = nn.LeakyReLU(0.1)
# encoder
self.enc1 = self._make_enc_layer(BasicBlock, [32, 64], self.blocks[0],
stride=self.strides[0], bn_d=self.bn_d)
self.enc2 = self._make_enc_layer(BasicBlock, [64, 128], self.blocks[1],
stride=self.strides[1], bn_d=self.bn_d)
self.enc3 = self._make_enc_layer(BasicBlock, [128, 256], self.blocks[2],
stride=self.strides[2], bn_d=self.bn_d)
self.enc4 = self._make_enc_layer(BasicBlock, [256, 512], self.blocks[3],
stride=self.strides[3], bn_d=self.bn_d)
self.enc5 = self._make_enc_layer(BasicBlock, [512, 1024], self.blocks[4],
stride=self.strides[4], bn_d=self.bn_d)
# for a bit of fun
self.dropout = nn.Dropout2d(self.drop_prob)
# last channels
self.last_channels = 1024
# make layer useful function
def _make_enc_layer(self, block, planes, blocks, stride, bn_d=0.1):
layers = []
# downsample
layers.append(("conv", nn.Conv2d(planes[0], planes[1],
kernel_size=3,
stride=[1, stride], dilation=1,
padding=1, bias=False)))
layers.append(("bn", nn.BatchNorm2d(planes[1], momentum=bn_d)))
layers.append(("relu", nn.LeakyReLU(0.1)))
# blocks
inplanes = planes[1]
for i in range(0, blocks):
layers.append(("residual_{}".format(i),
block(inplanes, planes, bn_d)))
return nn.Sequential(OrderedDict(layers))
def run_layer(self, x, layer, skips, os):
y = layer(x)
if y.shape[2] < x.shape[2] or y.shape[3] < x.shape[3]:
skips[os] = x.detach()
os *= 2
x = y
return x, skips, os
def forward(self, x, return_logits=False, return_list=None):
# filter input
x = x[:, self.input_idxs]
# run cnn
# store for skip connections
skips = {}
out_dict = {}
os = 1
# first layer
x, skips, os = self.run_layer(x, self.conv1, skips, os)
x, skips, os = self.run_layer(x, self.bn1, skips, os)
x, skips, os = self.run_layer(x, self.relu1, skips, os)
if return_list and 'enc_0' in return_list:
out_dict['enc_0'] = x.detach().cpu() # 32, 64, 1024
# all encoder blocks with intermediate dropouts
x, skips, os = self.run_layer(x, self.enc1, skips, os)
if return_list and 'enc_1' in return_list:
out_dict['enc_1'] = x.detach().cpu() # 64, 64, 512
x, skips, os = self.run_layer(x, self.dropout, skips, os)
x, skips, os = self.run_layer(x, self.enc2, skips, os)
if return_list and 'enc_2' in return_list:
out_dict['enc_2'] = x.detach().cpu() # 128, 64, 256
x, skips, os = self.run_layer(x, self.dropout, skips, os)
x, skips, os = self.run_layer(x, self.enc3, skips, os)
if return_list and 'enc_3' in return_list:
out_dict['enc_3'] = x.detach().cpu() # 256, 64, 128
x, skips, os = self.run_layer(x, self.dropout, skips, os)
x, skips, os = self.run_layer(x, self.enc4, skips, os)
if return_list and 'enc_4' in return_list:
out_dict['enc_4'] = x.detach().cpu() # 512, 64, 64
x, skips, os = self.run_layer(x, self.dropout, skips, os)
x, skips, os = self.run_layer(x, self.enc5, skips, os)
if return_list and 'enc_5' in return_list:
out_dict['enc_5'] = x.detach().cpu() # 1024, 64, 32
if return_logits:
return x
x, skips, os = self.run_layer(x, self.dropout, skips, os)
if return_list is not None:
return x, skips, out_dict
return x, skips
def get_last_depth(self):
return self.last_channels
def get_input_depth(self):
return self.input_depth
class Decoder(nn.Module):
"""
Class for DarknetSeg. Subclasses PyTorch's own "nn" module
"""
def __init__(self, params, OS=32, feature_depth=1024):
super(Decoder, self).__init__()
self.backbone_OS = OS
self.backbone_feature_depth = feature_depth
self.drop_prob = params["dropout"]
self.bn_d = params["bn_d"]
self.index = 0
# stride play
self.strides = [2, 2, 2, 2, 2]
# check current stride
current_os = 1
for s in self.strides:
current_os *= s
# redo strides according to needed stride
for i, stride in enumerate(self.strides):
if int(current_os) != self.backbone_OS:
if stride == 2:
current_os /= 2
self.strides[i] = 1
if int(current_os) == self.backbone_OS:
break
# decoder
self.dec5 = self._make_dec_layer(BasicBlock,
[self.backbone_feature_depth, 512],
bn_d=self.bn_d,
stride=self.strides[0])
self.dec4 = self._make_dec_layer(BasicBlock, [512, 256], bn_d=self.bn_d,
stride=self.strides[1])
self.dec3 = self._make_dec_layer(BasicBlock, [256, 128], bn_d=self.bn_d,
stride=self.strides[2])
self.dec2 = self._make_dec_layer(BasicBlock, [128, 64], bn_d=self.bn_d,
stride=self.strides[3])
self.dec1 = self._make_dec_layer(BasicBlock, [64, 32], bn_d=self.bn_d,
stride=self.strides[4])
# layer list to execute with skips
self.layers = [self.dec5, self.dec4, self.dec3, self.dec2, self.dec1]
# for a bit of fun
self.dropout = nn.Dropout2d(self.drop_prob)
# last channels
self.last_channels = 32
def _make_dec_layer(self, block, planes, bn_d=0.1, stride=2):
layers = []
# downsample
if stride == 2:
layers.append(("upconv", nn.ConvTranspose2d(planes[0], planes[1],
kernel_size=[1, 4], stride=[1, 2],
padding=[0, 1])))
else:
layers.append(("conv", nn.Conv2d(planes[0], planes[1],
kernel_size=3, padding=1)))
layers.append(("bn", nn.BatchNorm2d(planes[1], momentum=bn_d)))
layers.append(("relu", nn.LeakyReLU(0.1)))
# blocks
layers.append(("residual", block(planes[1], planes, bn_d)))
return nn.Sequential(OrderedDict(layers))
def run_layer(self, x, layer, skips, os):
feats = layer(x) # up
if feats.shape[-1] > x.shape[-1]:
os //= 2 # match skip
feats = feats + skips[os].detach() # add skip
x = feats
return x, skips, os
def forward(self, x, skips, return_logits=False, return_list=None):
os = self.backbone_OS
out_dict = {}
# run layers
x, skips, os = self.run_layer(x, self.dec5, skips, os)
if return_list and 'dec_4' in return_list:
out_dict['dec_4'] = x.detach().cpu() # 512, 64, 64
x, skips, os = self.run_layer(x, self.dec4, skips, os)
if return_list and 'dec_3' in return_list:
out_dict['dec_3'] = x.detach().cpu() # 256, 64, 128
x, skips, os = self.run_layer(x, self.dec3, skips, os)
if return_list and 'dec_2' in return_list:
out_dict['dec_2'] = x.detach().cpu() # 128, 64, 256
x, skips, os = self.run_layer(x, self.dec2, skips, os)
if return_list and 'dec_1' in return_list:
out_dict['dec_1'] = x.detach().cpu() # 64, 64, 512
x, skips, os = self.run_layer(x, self.dec1, skips, os)
if return_list and 'dec_0' in return_list:
out_dict['dec_0'] = x.detach().cpu() # 32, 64, 1024
logits = torch.clone(x).detach()
x = self.dropout(x)
if return_logits:
return x, logits
if return_list is not None:
return out_dict
return x
def get_last_depth(self):
return self.last_channels
class Model(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.backbone = Backbone(params=self.config["backbone"])
self.decoder = Decoder(params=self.config["decoder"], OS=self.config["backbone"]["OS"],
feature_depth=self.backbone.get_last_depth())
def load_pretrained_weights(self, path):
w_dict = torch.load(path + "/backbone",
map_location=lambda storage, loc: storage)
self.backbone.load_state_dict(w_dict, strict=True)
w_dict = torch.load(path + "/segmentation_decoder",
map_location=lambda storage, loc: storage)
self.decoder.load_state_dict(w_dict, strict=True)
def forward(self, x, return_logits=False, return_final_logits=False, return_list=None, agg_type='depth'):
if return_logits:
logits = self.backbone(x, return_logits)
logits = F.adaptive_avg_pool2d(logits, (1, 1)).squeeze()
logits = torch.clone(logits).detach().cpu().numpy()
return logits
elif return_list is not None:
x, skips, enc_dict = self.backbone(x, return_list=return_list)
dec_dict = self.decoder(x, skips, return_list=return_list)
out_dict = {**enc_dict, **dec_dict}
return out_dict
elif return_final_logits:
assert agg_type in ['all', 'sector', 'depth']
y, skips = self.backbone(x)
y, logits = self.decoder(y, skips, True)
B, C, H, W = logits.shape
N = 16
# avg all
if agg_type == 'all':
logits = logits.mean([2, 3])
# avg in patch
elif agg_type == 'sector':
logits = logits.view(B, C, H, N, W // N).mean([2, 4]).reshape(B, -1)
# avg in row
elif agg_type == 'depth':
logits = logits.view(B, C, N, H // N, W).mean([3, 4]).reshape(B, -1)
logits = torch.clone(logits).detach().cpu().numpy()
return logits
else:
y, skips = self.backbone(x)
y = self.decoder(y, skips, False)
return y
|