Spaces:
Running
on
Zero
Running
on
Zero
File size: 13,292 Bytes
9e15541 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 |
"""
Implements image encoders
"""
from collections import OrderedDict
from torch import profiler
from scenedino.models.prediction_heads.layers import *
import numpy as np
import torch
import torch.nn as nn
import torchvision.models as models
import torch.utils.model_zoo as model_zoo
# Code taken from https://github.com/nianticlabs/monodepth2
#
# Godard, Clément, et al.
# "Digging into self-supervised monocular depth estimation."
# Proceedings of the IEEE/CVF international conference on computer vision.
# 2019.
class ResNetMultiImageInput(models.ResNet):
"""Constructs a resnet model with varying number of input images.
Adapted from https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
"""
def __init__(self, block, layers, num_classes=1000, num_input_images=1):
super(ResNetMultiImageInput, self).__init__(block, layers)
self.inplanes = 64
self.conv1 = nn.Conv2d(
num_input_images * 3, 64, kernel_size=7, stride=2, padding=3, bias=False
)
self.bn1 = nn.BatchNorm2d(64)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def resnet_multiimage_input(num_layers, pretrained=False, num_input_images=1):
"""Constructs a ResNet model.
Args:
num_layers (int): Number of resnet layers. Must be 18 or 50
pretrained (bool): If True, returns a model pre-trained on ImageNet
num_input_images (int): Number of frames stacked as input
"""
assert num_layers in [18, 50], "Can only run with 18 or 50 layer resnet"
blocks = {18: [2, 2, 2, 2], 50: [3, 4, 6, 3]}[num_layers]
block_type = {18: models.resnet.BasicBlock, 50: models.resnet.Bottleneck}[
num_layers
]
model = ResNetMultiImageInput(block_type, blocks, num_input_images=num_input_images)
if pretrained:
loaded = model_zoo.load_url(
models.resnet.model_urls["resnet{}".format(num_layers)]
)
loaded["conv1.weight"] = (
torch.cat([loaded["conv1.weight"]] * num_input_images, 1) / num_input_images
)
model.load_state_dict(loaded)
return model
class ResnetEncoder(nn.Module):
"""Pytorch module for a resnet encoder"""
def __init__(self, num_layers, pretrained, num_input_images=1):
super(ResnetEncoder, self).__init__()
self.num_ch_enc = np.array([64, 64, 128, 256, 512])
resnets = {
18: models.resnet18,
34: models.resnet34,
50: models.resnet50,
101: models.resnet101,
152: models.resnet152,
}
weights = {
18: models.resnet.ResNet18_Weights.IMAGENET1K_V1,
34: models.resnet.ResNet34_Weights.IMAGENET1K_V1,
50: models.resnet.ResNet50_Weights.IMAGENET1K_V1,
101: models.resnet.ResNet101_Weights.IMAGENET1K_V1,
152: models.resnet.ResNet152_Weights.IMAGENET1K_V1,
}
if num_layers not in resnets:
raise ValueError(
"{} is not a valid number of resnet layers".format(num_layers)
)
if num_input_images > 1:
self.encoder = resnet_multiimage_input(
num_layers, pretrained, num_input_images
)
else:
# TODO: currently hardcoded make adaptable
self.encoder = resnets[num_layers](
weights=weights[num_layers]
)
if num_layers > 34:
self.num_ch_enc[1:] *= 4
def forward(self, input_image):
self.features = []
x = (input_image - 0.45) / 0.225
x = self.encoder.conv1(x)
x = self.encoder.bn1(x)
self.features.append(self.encoder.relu(x))
self.features.append(
self.encoder.layer1(self.encoder.maxpool(self.features[-1]))
)
self.features.append(self.encoder.layer2(self.features[-1]))
self.features.append(self.encoder.layer3(self.features[-1]))
self.features.append(self.encoder.layer4(self.features[-1]))
return self.features
class DepthDecoder(nn.Module):
def __init__(
self, num_ch_enc, scales=range(4), num_output_channels=1, use_skips=True
):
super(DepthDecoder, self).__init__()
self.num_output_channels = num_output_channels
self.use_skips = use_skips
self.upsample_mode = "nearest"
self.scales = scales
self.num_ch_enc = num_ch_enc
self.num_ch_dec = np.array([16, 32, 64, 128, 256])
# decoder
self.convs = OrderedDict()
for i in range(4, -1, -1):
# upconv_0
num_ch_in = self.num_ch_enc[-1] if i == 4 else self.num_ch_dec[i + 1]
num_ch_out = self.num_ch_dec[i]
self.convs[("upconv", i, 0)] = ConvBlock(num_ch_in, num_ch_out)
# upconv_1
num_ch_in = self.num_ch_dec[i]
if self.use_skips and i > 0:
num_ch_in += self.num_ch_enc[i - 1]
num_ch_out = self.num_ch_dec[i]
self.convs[("upconv", i, 1)] = ConvBlock(num_ch_in, num_ch_out)
for s in self.scales:
self.convs[("dispconv", s)] = Conv3x3(
self.num_ch_dec[s], self.num_output_channels
)
self.decoder_keys = {k: i for i, k in enumerate(self.convs.keys())}
self.decoder = nn.ModuleList(list(self.convs.values()))
self.sigmoid = nn.Sigmoid()
def forward(self, input_features):
self.outputs = {}
# decoder
x = input_features[-1]
for i in range(4, -1, -1):
# x = self.convs[("upconv", i, 0)](x)
x = self.decoder[self.decoder_keys[("upconv", i, 0)]](x)
x = [upsample(x)]
if self.use_skips and i > 0:
if x[0].shape[2] > input_features[i - 1].shape[2]:
x[0] = x[0][:, :, : input_features[i - 1].shape[2], :]
if x[0].shape[3] > input_features[i - 1].shape[3]:
x[0] = x[0][:, :, :, : input_features[i - 1].shape[3]]
x += [input_features[i - 1]]
x = torch.cat(x, 1)
# x = self.convs[("upconv", i, 1)](x)
x = self.decoder[self.decoder_keys[("upconv", i, 1)]](x)
self.outputs[("features", i)] = x
if i in self.scales:
# self.outputs[("disp", i)] = self.sigmoid(self.convs[("dispconv", i)](x))
self.outputs[("disp", i)] = self.sigmoid(
self.decoder[self.decoder_keys[("dispconv", i)]](x)
)
return self.outputs
class Decoder(nn.Module):
def __init__(
self, num_ch_enc, num_ch_dec=None, d_out=1, scales=range(4), use_skips=True, extra_outs=0
):
super(Decoder, self).__init__()
self.use_skips = use_skips
self.upsample_mode = "nearest"
self.num_ch_enc = num_ch_enc
if num_ch_dec is None:
self.num_ch_dec = np.array([128, 128, 256, 256, 512])
else:
self.num_ch_dec = num_ch_dec
self.d_out = d_out
self.scales = scales
self.extra_outs = extra_outs
self.num_ch_dec = [max(self.d_out, chns) for chns in self.num_ch_dec]
# decoder
self.convs = OrderedDict()
for i in range(4, -1, -1):
# upconv_0
num_ch_in = self.num_ch_enc[-1] if i == 4 else self.num_ch_dec[i + 1]
num_ch_out = self.num_ch_dec[i]
self.convs[("upconv", i, 0)] = ConvBlock(num_ch_in, num_ch_out)
# upconv_1
num_ch_in = self.num_ch_dec[i]
if self.use_skips and i > 0:
num_ch_in += self.num_ch_enc[i - 1]
num_ch_out = self.num_ch_dec[i]
self.convs[("upconv", i, 1)] = ConvBlock(num_ch_in, num_ch_out)
for s in self.scales:
self.convs[("dispconv", s)] = Conv3x3(self.num_ch_dec[s], self.d_out)
if self.extra_outs > 0:
for s in self.scales:
self.convs[("extra_outs", s)] = Conv3x3(self.num_ch_dec[s], self.extra_outs)
self.d_out += self.extra_outs
self.decoder_keys = {k: i for i, k in enumerate(self.convs.keys())}
self.decoder = nn.ModuleList(list(self.convs.values()))
self.sigmoid = nn.Sigmoid()
def forward(self, input_features):
with profiler.record_function("encoder_forward"):
self.outputs = {}
# decoder
x = input_features[-1]
for i in range(4, -1, -1):
x = self.decoder[self.decoder_keys[("upconv", i, 0)]](x)
x = [F.interpolate(x, scale_factor=(2, 2), mode="nearest")]
if self.use_skips and i > 0:
feats = input_features[i - 1]
if x[0].shape[2] > feats.shape[2]:
x[0] = x[0][:, :, : feats.shape[2], :]
if x[0].shape[3] > feats.shape[3]:
x[0] = x[0][:, :, :, : feats.shape[3]]
x += [feats]
x = torch.cat(x, 1)
x = self.decoder[self.decoder_keys[("upconv", i, 1)]](x)
self.outputs[("features", i)] = x
if i in self.scales:
self.outputs[("disp", i)] = self.decoder[
self.decoder_keys[("dispconv", i)]
](x)
if self.extra_outs > 0:
self.outputs[("extra_outs", i)] = self.decoder[
self.decoder_keys[("extra_outs", i)]
](x)
return self.outputs
class Monodepth2(nn.Module):
"""
2D (Spatial/Pixel-aligned/local) image encoder
"""
def __init__(
self,
resnet_layers=18,
cp_location=None,
freeze=False,
num_ch_dec=None,
d_out=128,
scales=range(4),
pointwise_convs=None,
extra_outs=0
):
super().__init__()
self.encoder = ResnetEncoder(resnet_layers, True, 1)
self.num_ch_enc = self.encoder.num_ch_enc
self.upsample_mode = "nearest"
self.d_out = d_out
self.extra_outs = extra_outs
self.scales = scales
if pointwise_convs is None:
decoder_out = self.d_out
else:
decoder_out = pointwise_convs[0]
# decoder
self.decoder = Decoder(
num_ch_enc=self.num_ch_enc,
d_out=decoder_out,
num_ch_dec=num_ch_dec,
scales=self.scales,
extra_outs=extra_outs,
)
self.num_ch_dec = self.decoder.num_ch_dec
# pointwise_convs
if pointwise_convs is None:
self.pointwise_convs = None
else:
pointwise_convs.append(self.d_out)
self.pointwise_convs = nn.Sequential(
*[nn.Sequential(
nn.ReLU(),
nn.Conv2d(pointwise_convs[i], pointwise_convs[i+1], 1, 1, 0),
) for i in range(len(pointwise_convs)-1)]
)
self.latent_size = self.d_out
if cp_location is not None:
cp = torch.load(cp_location)
self.load_state_dict(cp["model"])
if freeze:
for p in self.parameters(True):
p.requires_grad = False
def forward(self, x):
"""
For extracting ResNet's features.
:param x image (B, C, H, W)
:return latent (B, latent_size, H, W)
"""
with profiler.record_function("encoder_forward"):
x = torch.cat([x * 0.5 + 0.5], dim=1)
image_features = self.encoder(x) ## output encoder from Monodepth2
outputs = self.decoder(image_features)
x = [outputs[("disp", i)] for i in self.scales]
if self.decoder.extra_outs > 0:
x = [torch.cat((x[i], outputs[("extra_outs", i)]), dim=1) for i in self.scales]
if self.pointwise_convs is not None:
x = [
self.pointwise_convs(x_) for x_ in x
]
return x ## default: x ## note: we need img_feat for feeding into NeuRay
@classmethod
def from_conf(cls, conf):
return cls(
cp_location=conf.get("cp_location", None),
freeze=conf.get("freeze", False),
num_ch_dec=conf.get("num_ch_dec", None),
d_out=conf.get("d_out", 128),
resnet_layers=conf.get("resnet_layers", 18),
pointwise_convs=conf.get("pointwise_convs", None),
extra_outs=conf.get("extra_outs", 0),
)
|