File size: 14,911 Bytes
f348660 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 | import torch, torchvision
from safetensors.torch import load_file
from utils import log_utils
import networks
class RCNetModel(object):
'''
Image radar fusion to determine correspondence of radar to image
Arg(s):
input_channels_image : int
number of channels in the image
input_channels_depth : int
number of channels in depth map
input_patch_size_image : int
patch of image to consider for radar point
encoder_type : str
encoder type
n_filters_encoder_image : list[int]
list of filters for each layer in image encoder
n_neurons_encoder_image : list[int]
list of neurons for each layer in depth encoder
decoder_type : str
decoder type
n_filters_decoder : list[int]
list of filters for each layer in decoder
weight_initializer : str
kaiming_normal, kaiming_uniform, xavier_normal, xavier_uniform
activation_func : str
activation function for network
device : torch.device
device for running model
'''
def __init__(self,
input_channels_image,
input_channels_depth,
input_patch_size_image,
encoder_type,
n_filters_encoder_image,
n_neurons_encoder_depth,
decoder_type,
n_filters_decoder,
weight_initializer='kaiming_uniform',
activation_func='leaky_relu',
device=torch.device('cuda')):
self.input_patch_size_image = input_patch_size_image
self.device = device
# height, width = input_patch_size_image
# latent_height = np.ceil(height / 32.0).astype(int)
# latent_width = np.ceil(width / 32.0).astype(int)
height, width = input_patch_size_image
latent_height = int((height // 32.0))
latent_width = int((width // 32.0))
latent_size_depth = latent_height * latent_width * n_neurons_encoder_depth[-1]
# Build encoder
if 'rcnet' in encoder_type:
self.encoder = networks.RCNetEncoder(
input_channels_image=input_channels_image,
input_channels_depth=input_channels_depth,
input_patch_size_image=input_patch_size_image,
n_filters_encoder_image=n_filters_encoder_image,
n_neurons_encoder_depth=n_neurons_encoder_depth,
latent_size_depth=latent_size_depth,
weight_initializer=weight_initializer,
activation_func=activation_func,
use_batch_norm='batch_norm' in encoder_type)
else:
raise ValueError('Encoder type {} not supported.'.format(encoder_type))
# Calculate number of channels for latent and skip connections combining image + depth
n_skips = n_filters_encoder_image[:-1]
n_skips = n_skips[::-1] + [0]
latent_channels = n_filters_encoder_image[-1] + n_neurons_encoder_depth[-1]
# Build decoder
if 'multiscale' in decoder_type:
self.decoder = networks.MultiScaleDecoder(
input_channels=latent_channels,
output_channels=1,
n_resolution=1,
n_filters=n_filters_decoder,
n_skips=n_skips,
weight_initializer=weight_initializer,
activation_func=activation_func,
output_func='linear',
use_batch_norm='batch_norm' in decoder_type,
deconv_type='up')
else:
raise ValueError('Decoder type {} not supported.'.format(decoder_type))
# Move to device
self.to(self.device)
def forward(self, image, point, bounding_boxes, return_logits=True):
'''
Forwards the inputs through the network
Arg(s):
image : torch.Tensor[float32]
N x 3 x H x W image
point : torch.Tensor[float32]
N x 3 input point
return_logits : bool
if set, then return logits otherwise sigmoid
Returns:
torch.Tensor[float32] : N x 1 x H x W logits (correspondence map)
'''
latent, skips = self.encoder(image, point, bounding_boxes)
logits = self.decoder(x=latent, skips=skips, shape=self.input_patch_size_image)[-1]
if return_logits:
return logits
else:
return torch.sigmoid(logits)
def parameters(self):
'''
Returns the list of parameters in the model
Returns:
list[torch.Tensor[float32]] : list of parameters
'''
parameters = \
list(self.encoder.parameters()) + \
list(self.decoder.parameters())
return parameters
def eval(self):
'''
Sets model to evaluation mode
'''
self.encoder.eval()
self.decoder.eval()
def to(self, device):
'''
Moves model to specified device
Arg(s):
device : torch.device
device for running model
'''
# Move to device
self.encoder.to(device)
self.decoder.to(device)
def restore_model(self, checkpoint_path):
'''
Restore weights of the model
Arg(s):
checkpoint_path : str
path to checkpoint
Returns:
int : current step in optimization
None : retained for compatibility with the baseline inference call
'''
checkpoint = load_file(checkpoint_path, device='cpu')
encoder_prefix = 'radarnet_encoder.'
decoder_prefix = 'radarnet_decoder.'
encoder_state = {
key[len(encoder_prefix):]: value
for key, value in checkpoint.items()
if key.startswith(encoder_prefix)
}
decoder_state = {
key[len(decoder_prefix):]: value
for key, value in checkpoint.items()
if key.startswith(decoder_prefix)
}
self.encoder.load_state_dict(encoder_state)
self.decoder.load_state_dict(decoder_state)
return 0, None
def data_parallel(self):
'''
Allows multi-gpu split along batch
'''
self.encoder = torch.nn.DataParallel(self.encoder)
self.decoder = torch.nn.DataParallel(self.decoder)
def log_summary(self,
summary_writer,
tag,
step,
image=None,
output_response=None,
output_label=None,
output_depth=None,
validity_map=None,
ground_truth_label=None,
ground_truth_depth=None,
scalars={},
n_display=4):
'''
Logs summary to Tensorboard
Arg(s):
summary_writer : SummaryWriter
Tensorboard summary writer
tag : str
tag that prefixes names to log
step : int
current step in training
image : torch.Tensor[float32]
N x 3 x H x W image
output_response : torch.Tensor[float32]
N x 1 x H x W soft correspondence map
output_label : torch.Tensor[float32]
N x 1 x H x W binary correspondence map
output_depth : torch.Tensor[float32]
N x 1 x H x W depth map
validity_map : torch.Tensor[float32]
N x 1 x H x W validity map
ground_truth_label : torch.Tensor[float32]
N x 1 x H x W ground truth label
ground_truth_depth : torch.Tensor[float32]
N x 1 x H x W ground truth depth map
scalars : dict[str, float]
dictionary of scalars to log
n_display : int
number of images to display
'''
with torch.no_grad():
display_summary_image = []
display_summary_image_text = tag
if image is not None:
image_summary = image[0:n_display, ...]
display_summary_image_text += '-image'
# Add to list of images to log
display_summary_image.append(image_summary.cpu())
if output_response is not None:
output_response_summary = output_response[0:n_display, ...]
display_summary_image_text += '-output_response'
# Add to list of images to log
display_summary_image.append(
log_utils.colorize(
output_response_summary.cpu(),
colormap='inferno'))
# Log distribution of output response
summary_writer.add_histogram(
tag + '-output_response_distro',
output_response_summary,
global_step=step)
if output_label is not None:
output_label_summary = output_label[0:n_display, ...]
display_summary_image_text += '-output_label'
# Add to list of images to log
display_summary_image.append(
log_utils.colorize(
output_label_summary.cpu(),
colormap='inferno'))
# Log distribution of output and label
summary_writer.add_histogram(
tag + '-output_label_distro',
output_label_summary,
global_step=step)
if ground_truth_label is not None:
ground_truth_label_summary = ground_truth_label[0:n_display, ...]
validity_map_label_summary = torch.where(
ground_truth_label_summary > 0,
torch.ones_like(ground_truth_label_summary),
torch.zeros_like(ground_truth_label_summary))
display_summary_image_text += '-ground_truth_label'
if output_label is not None:
display_summary_image_text += '-error'
# Compute output error w.r.t. ground truth
ground_truth_label_error_summary = \
torch.abs(output_label_summary - ground_truth_label_summary)
ground_truth_label_error_summary = torch.where(
validity_map_label_summary == 1.0,
(ground_truth_label_error_summary + 1e-8) / (ground_truth_label_summary + 1e-8),
validity_map_label_summary)
display_summary_image.append(
log_utils.colorize(
ground_truth_label_error_summary.cpu(),
colormap='inferno'))
# Add to list of images to log
display_summary_image.append(
log_utils.colorize(
ground_truth_label_summary.cpu(),
colormap='inferno'))
# Log distribution of ground truth
summary_writer.add_histogram(
tag + '_ground_truth_label_distro',
ground_truth_label,
global_step=step)
if validity_map is not None:
validity_map_summary = validity_map[0:n_display, ...]
display_summary_image_text += '-validity_map'
# Add to list of images to log
display_summary_image.append(
log_utils.colorize(
validity_map_summary.cpu(),
colormap='inferno'))
if output_depth is not None:
output_depth_summary = output_depth[0:n_display, ...]
display_summary_image_text += '-output_depth'
# Add to list of images to log
display_summary_image.append(
log_utils.colorize(
(output_depth_summary / 100.0).cpu(),
colormap='viridis'))
# Log distribution of output depth
summary_writer.add_histogram(
tag + '-output_depth_distro',
output_depth,
global_step=step)
if ground_truth_depth is not None:
ground_truth_depth = torch.unsqueeze(ground_truth_depth[:, 0, :, :], dim=1)
ground_truth_depth_summary = ground_truth_depth[0:n_display, ...]
validity_map_summary = torch.where(
ground_truth_depth_summary > 0,
torch.ones_like(ground_truth_depth_summary),
torch.zeros_like(ground_truth_depth_summary))
display_summary_image_text += '-ground_truth_label'
if output_depth is not None:
display_summary_image_text += '-error'
# Compute output error w.r.t. ground truth
ground_truth_depth_error_summary = \
torch.abs(output_depth_summary - ground_truth_depth_summary)
ground_truth_depth_error_summary = torch.where(
validity_map_summary == 1.0,
(ground_truth_depth_error_summary + 1e-8) / (ground_truth_depth_summary + 1e-8),
validity_map_summary)
display_summary_image.append(
log_utils.colorize(
(ground_truth_depth_error_summary / 0.05).cpu(),
colormap='inferno'))
# Add to list of images to log
display_summary_image.append(
log_utils.colorize(
(ground_truth_depth_summary / 100.0).cpu(),
colormap='viridis'))
# Log distribution of ground truth
summary_writer.add_histogram(
tag + '-ground_truth_distro',
ground_truth_depth,
global_step=step)
# Log scalars to tensorboard
for (name, value) in scalars.items():
summary_writer.add_scalar(tag + '-' + name, value, global_step=step)
# Log image summaries to tensorboard
if len(display_summary_image) > 1:
display_summary_image = torch.cat(display_summary_image, dim=2)
summary_writer.add_image(
display_summary_image_text,
torchvision.utils.make_grid(display_summary_image, nrow=n_display),
global_step=step)
|