File size: 20,321 Bytes
abd2a81 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 |
import math
from functools import partial
import scipy.interpolate
import numpy as np
import torch
import torch.distributed
from torch.nn import functional as F
from . import measures
from . import frame_field_utils
import torch_lydorn.kornia
from lydorn_utils import math_utils, print_utils
# --- Base classes --- #
class Loss(torch.nn.Module):
def __init__(self, name):
"""
Attribute extra_info can be used in self.compute() to add intermediary results of loss computation for
visualization for example.
It is the second output of self.__call__()
:param name:
"""
super(Loss, self).__init__()
self.name = name
self.norm_meter = None
self.norm = torch.nn.parameter.Parameter(torch.Tensor(1), requires_grad=False)
self.reset_norm()
self.extra_info = {} #
def reset_norm(self):
self.norm_meter = math_utils.AverageMeter("{}_norm".format(self.name), init_val=1)
self.norm[0] = self.norm_meter.val
def update_norm(self, pred_batch, gt_batch, nums):
loss = self.compute(pred_batch, gt_batch)
self.norm_meter.update(loss, nums)
self.norm[0] = self.norm_meter.val
def sync(self, world_size):
"""
This method should be used to synchronize loss norms across GPUs when using distributed training
:return:
"""
torch.distributed.all_reduce(self.norm)
self.norm /= world_size
def compute(self, pred_batch, gt_batch):
raise NotImplementedError
def forward(self, pred_batch, gt_batch, normalize=True):
loss = self.compute(pred_batch, gt_batch)
if normalize:
assert 1e-9 < self.norm[0], "self.norm[0] <= 1e-9 -> this might lead to numerical instabilities."
loss = loss / self.norm[0]
extra_info = self.extra_info
self.extra_info = {} # Re-init extra_info
# contains_nan = bool(torch.sum(torch.isnan(loss)).item())
# assert not contains_nan, f"loss {str(self)} is Nan!"
return loss, extra_info
def __repr__(self):
return "{} (name={}, norm={:0.06})".format(self.__class__.__name__, self.name, self.norm[0])
class MultiLoss(torch.nn.Module):
def __init__(self, loss_funcs, weights, epoch_thresholds=None, pre_processes=None):
"""
@param loss_funcs:
@param weights:
@param pre_processes: List of functions to call with 2 arguments (which are updated): pred_batch, gt_batch to compute only one values used by several losses.
"""
super(MultiLoss, self).__init__()
assert len(loss_funcs) == len(weights), \
"Should have the same amount of loss_funcs ({}) and weights ({})".format(len(loss_funcs), len(weights))
self.loss_funcs = torch.nn.ModuleList(loss_funcs)
self.weights = []
for weight in weights:
if isinstance(weight, list):
# Weight is a list of coefs corresponding to epoch_thresholds, they will be interpolated in-between
self.weights.append(scipy.interpolate.interp1d(epoch_thresholds, weight, bounds_error=False, fill_value=(weight[0], weight[-1])))
elif isinstance(weight, float) or isinstance(weight, int):
self.weights.append(float(weight))
else:
raise TypeError(f"Type {type(weight)} not supported as a loss coef weight.")
self.pre_processes = pre_processes
for loss_func, weight in zip(self.loss_funcs, self.weights):
if weight == 0:
print_utils.print_info(f"INFO: loss '{loss_func.name}' has a weight of zero and thus won't affect grad update.")
def reset_norm(self):
for loss_func in self.loss_funcs:
loss_func.reset_norm()
def update_norm(self, pred_batch, gt_batch, nums):
if self.pre_processes is not None:
for pre_process in self.pre_processes:
pred_batch, gt_batch = pre_process(pred_batch, gt_batch)
for loss_func in self.loss_funcs:
loss_func.update_norm(pred_batch, gt_batch, nums)
def sync(self, world_size):
"""
This method should be used to synchronize loss norms across GPUs when using distributed training
:return:
"""
for loss_func in self.loss_funcs:
loss_func.sync(world_size)
def forward(self, pred_batch, gt_batch, normalize=True, epoch=None):
if self.pre_processes is not None:
for pre_process in self.pre_processes:
pred_batch, gt_batch = pre_process(pred_batch, gt_batch)
total_loss = 0
# total_weight = 0
individual_losses_dict = {}
extra_dict = {}
for loss_func_i, weight_i in zip(self.loss_funcs, self.weights):
loss_i, extra_dict_i = loss_func_i(pred_batch, gt_batch, normalize=normalize)
if isinstance(weight_i, scipy.interpolate.interpolate.interp1d) and epoch is not None:
current_weight = float(weight_i(epoch))
else:
current_weight = weight_i
total_loss += current_weight * loss_i
# total_weight += weight_i
individual_losses_dict[loss_func_i.name] = loss_i
extra_dict[loss_func_i.name] = extra_dict_i
# total_loss /= total_weight
return total_loss, individual_losses_dict, extra_dict
def __repr__(self):
ret = "\n\t".join([str(loss_func) for loss_func in self.loss_funcs])
return "{}:\n\t{}".format(self.__class__.__name__, ret)
# --- Build combined loss: --- #
def compute_seg_loss_weigths(pred_batch, gt_batch, config):
"""
Combines distances (from U-Net paper) with sizes (from https://github.com/neptune-ai/open-solution-mapping-challenge).
@param pred_batch:
@param gt_batch:
@return:
"""
device = gt_batch["distances"].device
use_dist = config["loss_params"]["seg_loss_params"]["use_dist"]
use_size = config["loss_params"]["seg_loss_params"]["use_size"]
w0 = config["loss_params"]["seg_loss_params"]["w0"]
sigma = config["loss_params"]["seg_loss_params"]["sigma"]
height = gt_batch["image"].shape[2]
width = gt_batch["image"].shape[3]
im_radius = math.sqrt(height * width) / 2
# --- Class imbalance weight (not forgetting background):
gt_polygons_mask = (0 < gt_batch["gt_polygons_image"]).float()
background_freq = 1 - torch.sum(gt_batch["class_freq"], dim=1)
pixel_class_freq = gt_polygons_mask * gt_batch["class_freq"][:, :, None, None] + \
(1 - gt_polygons_mask) * background_freq[:, None, None, None]
if pixel_class_freq.min() == 0:
print_utils.print_error("ERROR: pixel_class_freq has some zero values, can't divide by zero!")
raise ZeroDivisionError
freq_weights = 1 / pixel_class_freq
# print("freq_weights:", freq_weights.min().item(), freq_weights.max().item())
# Compute size weights
# print("sizes:", gt_batch["sizes"].min().item(), gt_batch["sizes"].max().item())
# print("distances:", gt_batch["distances"].min().item(), gt_batch["distances"].max().item())
# print("im_radius:", im_radius)
size_weights = None
if use_size:
if gt_batch["sizes"].min() == 0:
print_utils.print_error(("ERROR: sizes tensor has zero values, can't divide by zero!"))
raise ZeroDivisionError
size_weights = 1 + 1 / (im_radius * gt_batch["sizes"])
distance_weights = None
if use_dist:
# print("distances:", gt_batch["distances"].min().item(), gt_batch["distances"].max().item())
distance_weights = gt_batch["distances"] * (height + width) # Denormalize distances
distance_weights = w0 * torch.exp(-(distance_weights ** 2) / (sigma ** 2))
# print("sum(distances == 0):", torch.sum(gt_batch["distances"] == 0).item())
# print("distance_weights:", distance_weights.min().item(), distance_weights.max().item())
# print(distance_weights.shape, distance_weights.min().item(), distance_weights.max().item())
# print(size_weights.shape, size_weights.min().item(), size_weights.max().item())
# print(freq_weights.shape, freq_weights.min().item(), freq_weights.max().item())
gt_batch["seg_loss_weights"] = freq_weights
if use_dist:
gt_batch["seg_loss_weights"] += distance_weights
if use_size:
gt_batch["seg_loss_weights"] *= size_weights
# print(gt_batch["seg_loss_weights"].shape, gt_batch["seg_loss_weights"].min().item(), gt_batch["seg_loss_weights"].max().item())
# print("seg_loss_weights:", size_weights.min().item(), size_weights.max().item())
# print("freq_weights:", freq_weights.min().item(), freq_weights.max().item())
# print("size_weights:", size_weights.min().item(), size_weights.max().item())
# print("distance_weights:", distance_weights.min().item(), distance_weights.max().item())
# Display:
# display_seg_loss_weights = gt_batch["seg_loss_weights"][0].cpu().detach().numpy()
# display_distance_weights = distance_weights[0].cpu().detach().numpy()
# skimage.io.imsave("seg_loss_dist_weights.png", display_distance_weights[0])
# display_size_weights = size_weights[0].cpu().detach().numpy()
# skimage.io.imsave("seg_loss_size_weights.png", display_size_weights[0])
# display_freq_weights = freq_weights[0].cpu().detach().numpy()
# display_freq_weights = display_freq_weights - display_freq_weights.min()
# display_freq_weights /= display_freq_weights.max()
# skimage.io.imsave("seg_loss_freq_weights.png", np.moveaxis(display_freq_weights, 0, -1))
# for i in range(3):
# skimage.io.imsave(f"seg_loss_weights_{i}.png", display_seg_loss_weights[i])
# skimage.io.imsave(f"freq_weights_{i}.png", display_freq_weights[i])
return pred_batch, gt_batch
def compute_gt_field(pred_batch, gt_batch):
gt_crossfield_angle = gt_batch["gt_crossfield_angle"]
gt_field = torch.cat([torch.cos(gt_crossfield_angle),
torch.sin(gt_crossfield_angle)], dim=1)
gt_batch["gt_field"] = gt_field
return pred_batch, gt_batch
class ComputeSegGrads:
def __init__(self, device):
self.spatial_gradient = torch_lydorn.kornia.filters.SpatialGradient(mode="scharr", coord="ij", normalized=True, device=device)
def __call__(self, pred_batch, gt_batch):
seg = pred_batch["seg"] # (b, c, h, w)
seg_grads = 2 * self.spatial_gradient(seg) # (b, c, 2, h, w), Normalize (kornia normalizes to -0.5, 0.5 for input in [0, 1])
seg_grad_norm = seg_grads.norm(dim=2) # (b, c, h, w)
seg_grads_normed = seg_grads / (seg_grad_norm[:, :, None, ...] + 1e-6) # (b, c, 2, h, w)
pred_batch["seg_grads"] = seg_grads
pred_batch["seg_grad_norm"] = seg_grad_norm
pred_batch["seg_grads_normed"] = seg_grads_normed
return pred_batch, gt_batch
def build_combined_loss(config):
pre_processes = []
loss_funcs = []
weights = []
if config["compute_seg"]:
partial_compute_seg_loss_weigths = partial(compute_seg_loss_weigths, config=config)
pre_processes.append(partial_compute_seg_loss_weigths)
gt_channel_selector = [config["seg_params"]["compute_interior"], config["seg_params"]["compute_edge"], config["seg_params"]["compute_vertex"]]
loss_funcs.append(SegLoss(name="seg",
gt_channel_selector=gt_channel_selector,
bce_coef=config["loss_params"]["seg_loss_params"]["bce_coef"],
dice_coef=config["loss_params"]["seg_loss_params"]["dice_coef"]))
weights.append(config["loss_params"]["multiloss"]["coefs"]["seg"])
if config["compute_crossfield"]:
pre_processes.append(compute_gt_field)
loss_funcs.append(CrossfieldAlignLoss(name="crossfield_align"))
weights.append(config["loss_params"]["multiloss"]["coefs"]["crossfield_align"])
loss_funcs.append(CrossfieldAlign90Loss(name="crossfield_align90"))
weights.append(config["loss_params"]["multiloss"]["coefs"]["crossfield_align90"])
loss_funcs.append(CrossfieldSmoothLoss(name="crossfield_smooth"))
weights.append(config["loss_params"]["multiloss"]["coefs"]["crossfield_smooth"])
# --- Coupling losses:
if config["compute_seg"]:
need_seg_grads = False
pred_channel = -1
# Seg interior <-> Crossfield coupling:
if config["seg_params"]["compute_interior"] and config["compute_crossfield"]:
need_seg_grads = True
pred_channel += 1
loss_funcs.append(SegCrossfieldLoss(name="seg_interior_crossfield", pred_channel=pred_channel))
weights.append(config["loss_params"]["multiloss"]["coefs"]["seg_interior_crossfield"])
# Seg edge <-> Crossfield coupling:
if config["seg_params"]["compute_edge"] and config["compute_crossfield"]:
need_seg_grads = True
pred_channel += 1
loss_funcs.append(SegCrossfieldLoss(name="seg_edge_crossfield", pred_channel=pred_channel))
weights.append(config["loss_params"]["multiloss"]["coefs"]["seg_edge_crossfield"])
# Seg edge <-> seg interior coupling:
if config["seg_params"]["compute_interior"] and config["seg_params"]["compute_edge"]:
need_seg_grads = True
loss_funcs.append(SegEdgeInteriorLoss(name="seg_edge_interior"))
weights.append(config["loss_params"]["multiloss"]["coefs"]["seg_edge_interior"])
if need_seg_grads:
pre_processes.append(ComputeSegGrads(config["device"]))
combined_loss = MultiLoss(loss_funcs, weights, epoch_thresholds=config["loss_params"]["multiloss"]["coefs"]["epoch_thresholds"], pre_processes=pre_processes)
return combined_loss
# --- Specific losses --- #
class SegLoss(Loss):
def __init__(self, name, gt_channel_selector, bce_coef=0.5, dice_coef=0.5):
"""
:param name:
:param gt_channel_selector: used to select which channels gt_polygons_image to use to compare to predicted seg
(see docstring of method compute() for more details).
"""
super(SegLoss, self).__init__(name)
self.gt_channel_selector = gt_channel_selector
self.bce_coef = bce_coef
self.dice_coef = dice_coef
def compute(self, pred_batch, gt_batch):
"""
seg and gt_polygons_image do not necessarily have the same channel count.
gt_selector is used to select which channels of gt_polygons_image to use.
For example, if seg has C_pred=2 (interior and edge) and
gt_polygons_image has C_gt=3 (interior, edge and vertex), use gt_channel_selector=slice(0, 2)
@param pred_batch: key "seg" is shape (N, C_pred, H, W)
@param gt_batch: key "gt_polygons_image" is shape (N, C_gt, H, W)
@return:
"""
# print(self.name)
pred_seg = pred_batch["seg"]
gt_seg = gt_batch["gt_polygons_image"][:, self.gt_channel_selector, ...]
weights = gt_batch["seg_loss_weights"][:, self.gt_channel_selector, ...]
dice = measures.dice_loss(pred_seg, gt_seg)
mean_dice = torch.mean(dice)
mean_cross_entropy = F.binary_cross_entropy(pred_seg, gt_seg, weight=weights, reduction="mean")
# Display:
# dispaly_pred_seg = pred_seg[0, 0].cpu().detach().numpy()
# print(f'{self.name}_pred:', dispaly_pred_seg.shape, dispaly_pred_seg.min(), dispaly_pred_seg.max())
# skimage.io.imsave(f'{self.name}_pred.png', dispaly_pred_seg)
# dispaly_gt_seg = gt_seg[0].cpu().detach().numpy()
# skimage.io.imsave(f'{self.name}_gt.png', dispaly_gt_seg)
return self.bce_coef * mean_cross_entropy + self.dice_coef * mean_dice
class CrossfieldAlignLoss(Loss):
def __init__(self, name):
super(CrossfieldAlignLoss, self).__init__(name)
def compute(self, pred_batch, gt_batch):
c0 = pred_batch["crossfield"][:, :2]
c2 = pred_batch["crossfield"][:, 2:]
z = gt_batch["gt_field"]
gt_polygons_image = gt_batch["gt_polygons_image"]
assert 2 <= gt_polygons_image.shape[1], \
"gt_polygons_image should have at least 2 channels for interior and edges"
gt_edges = gt_polygons_image[:, 1, ...]
align_loss = frame_field_utils.framefield_align_error(c0, c2, z, complex_dim=1)
avg_align_loss = torch.mean(align_loss * gt_edges)
self.extra_info["gt_field"] = gt_batch["gt_field"]
return avg_align_loss
class CrossfieldAlign90Loss(Loss):
def __init__(self, name):
super(CrossfieldAlign90Loss, self).__init__(name)
def compute(self, pred_batch, gt_batch):
c0 = pred_batch["crossfield"][:, :2]
c2 = pred_batch["crossfield"][:, 2:]
z = gt_batch["gt_field"]
z_90deg = torch.cat((- z[:, 1:2, ...], z[:, 0:1, ...]), dim=1)
gt_polygons_image = gt_batch["gt_polygons_image"]
assert gt_polygons_image.shape[1] == 3, \
"gt_polygons_image should have 3 channels for interior, edges and vertices"
gt_edges = gt_polygons_image[:, 1, ...]
gt_vertices = gt_polygons_image[:, 2, ...]
gt_edges_minus_vertices = gt_edges - gt_vertices
gt_edges_minus_vertices = gt_edges_minus_vertices.clamp(0, 1)
align90_loss = frame_field_utils.framefield_align_error(c0, c2, z_90deg, complex_dim=1)
avg_align90_loss = torch.mean(align90_loss * gt_edges_minus_vertices)
return avg_align90_loss
class CrossfieldSmoothLoss(Loss):
def __init__(self, name):
super(CrossfieldSmoothLoss, self).__init__(name)
self.laplacian_penalty = frame_field_utils.LaplacianPenalty(channels=4)
def compute(self, pred_batch, gt_batch):
c0c2 = pred_batch["crossfield"]
gt_polygons_image = gt_batch["gt_polygons_image"]
gt_edges_inv = 1 - gt_polygons_image[:, 1, ...]
penalty = self.laplacian_penalty(c0c2)
avg_penalty = torch.mean(penalty * gt_edges_inv[:, None, ...])
return avg_penalty
class SegCrossfieldLoss(Loss):
def __init__(self, name, pred_channel):
super(SegCrossfieldLoss, self).__init__(name)
self.pred_channel = pred_channel
def compute(self, pred_batch, gt_batch):
# TODO: don't apply on corners: corner_map = gt_batch["gt_polygons_image"][:, 2, :, :]
# TODO: apply on all seg at once? Like seg is now?
c0 = pred_batch["crossfield"][:, :2]
c2 = pred_batch["crossfield"][:, 2:]
seg_slice_grads_normed = pred_batch["seg_grads_normed"][:, self.pred_channel, ...]
seg_slice_grad_norm = pred_batch["seg_grad_norm"][:, self.pred_channel, ...]
align_loss = frame_field_utils.framefield_align_error(c0, c2, seg_slice_grads_normed, complex_dim=1)
# normed_align_loss = align_loss * seg_slice_grad_norm
# avg_align_loss = torch.sum(normed_align_loss) / (torch.sum(seg_slice_grad_norm) + 1e-6)
avg_align_loss = torch.mean(align_loss * seg_slice_grad_norm.detach())
# (prev line) Don't back-propagate to seg_slice_grad_norm so that seg smoothness is not encouraged
# Save extra info for viz:
self.extra_info["seg_slice_grads"] = pred_batch["seg_grads"][:, self.pred_channel, ...]
return avg_align_loss
class SegEdgeInteriorLoss(Loss):
"""
Enforce seg edge to be equal to interior grad norm except inside buildings
"""
def __init__(self, name):
super(SegEdgeInteriorLoss, self).__init__(name)
def compute(self, pred_batch, batch):
seg_interior = pred_batch["seg"][:, 0, ...]
seg_edge = pred_batch["seg"][:, 1, ...]
seg_interior_grad_norm = pred_batch["seg_grad_norm"][:, 0, ...]
raw_loss = torch.abs(seg_edge - seg_interior_grad_norm)
# Apply the loss only on interior boundaries and outside of objects
outside_mask = (torch.cos(np.pi * seg_interior) + 1) / 2
boundary_mask = (1 - torch.cos(np.pi * seg_interior_grad_norm)) / 2
mask = torch.max(outside_mask, boundary_mask).float()
avg_loss = torch.mean(raw_loss * mask)
return avg_loss
|