File size: 16,115 Bytes
a166479 |
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 |
# Copyright (c) Facebook, Inc. and its affiliates.
# Modified by Bowen Cheng from https://github.com/facebookresearch/detr/blob/master/models/detr.py
"""
MaskFormer criterion.
"""
import torch
import numpy as np
import torch.nn.functional as F
import torch.distributed as dist
from torch import nn
import sys
import os
sys.path.append(os.path.dirname(__file__) + os.sep + '../')
from .point_features import point_sample, get_uncertain_point_coords_with_randomness
from .misc import is_dist_avail_and_initialized, nested_tensor_from_tensor_list, get_world_size
def dice_loss(
inputs: torch.Tensor,
targets: torch.Tensor,
num_masks: float,
):
"""
Compute the DICE loss, similar to generalized IOU for masks
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
"""
inputs = inputs.sigmoid()
inputs = inputs.flatten(1)
numerator = 2 * (inputs * targets).sum(-1)
denominator = inputs.sum(-1) + targets.sum(-1)
loss = 1 - (numerator + 1) / (denominator + 1)
return loss.sum() / num_masks
def sigmoid_ce_loss(
inputs: torch.Tensor,
targets: torch.Tensor,
num_masks: float,
):
"""
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
Returns:
Loss tensor
"""
loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
return loss.mean(1).sum() / num_masks
def sigmoid_focal_loss(inputs, targets, num_masks, alpha: float = 0.25, gamma: float = 2):
"""
Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002.
Args:
inputs: A float tensor of arbitrary shape.
The predictions for each example.
targets: A float tensor with the same shape as inputs. Stores the binary
classification label for each element in inputs
(0 for the negative class and 1 for the positive class).
alpha: (optional) Weighting factor in range (0,1) to balance
positive vs negative examples. Default = -1 (no weighting).
gamma: Exponent of the modulating factor (1 - p_t) to
balance easy vs hard examples.
Returns:
Loss tensor
"""
prob = inputs.sigmoid()
ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
p_t = prob * targets + (1 - prob) * (1 - targets)
loss = ce_loss * ((1 - p_t) ** gamma)
if alpha >= 0:
alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
loss = alpha_t * loss
return loss.mean(1).sum() / num_masks
def calculate_uncertainty(logits):
"""
We estimate uncerainty as L1 distance between 0.0 and the logit prediction in 'logits' for the
foreground class in `classes`.
Args:
logits (Tensor): A tensor of shape (R, 1, ...) for class-specific or
class-agnostic, where R is the total number of predicted masks in all images and C is
the number of foreground classes. The values are logits.
Returns:
scores (Tensor): A tensor of shape (R, 1, ...) that contains uncertainty scores with
the most uncertain locations having the highest uncertainty score.
"""
assert logits.shape[1] == 1
gt_class_logits = logits.clone()
return -(torch.abs(gt_class_logits))
class SetCriterion(nn.Module):
"""This class computes the loss for DETR.
The process happens in two steps:
1) we compute hungarian assignment between ground truth boxes and the outputs of the model
2) we supervise each pair of matched ground-truth / prediction (supervise class and box)
"""
def __init__(self, num_classes, matcher, weight_dict, eos_coef, losses,
num_points, oversample_ratio, importance_sample_ratio, device):
"""Create the criterion.
Parameters:
num_classes: number of object categories, omitting the special no-object category
matcher: module able to compute a matching between targets and proposals
weight_dict: dict containing as key the names of the losses and as values their relative weight.
eos_coef: relative classification weight applied to the no-object category
losses: list of all the losses to be applied. See get_loss for list of available losses.
"""
super().__init__()
self.num_classes = num_classes
self.matcher = matcher
self.weight_dict = weight_dict
self.eos_coef = eos_coef
self.losses = losses
self.device = device
empty_weight = torch.ones(self.num_classes + 1).to(device)
empty_weight[-1] = self.eos_coef
self.register_buffer("empty_weight", empty_weight)
# pointwise mask loss parameters
self.num_points = num_points
self.oversample_ratio = oversample_ratio
self.importance_sample_ratio = importance_sample_ratio
def loss_labels(self, outputs, targets, indices, num_masks):
"""Classification loss (NLL)
targets dicts must contain the key "labels" containing a tensor of dim [nb_target_boxes]
"""
assert "pred_logits" in outputs
src_logits = outputs["pred_logits"].float()
idx = self._get_src_permutation_idx(indices)
target_classes_o = torch.cat([t["labels"][J] for t, (_, J) in zip(targets, indices)]).to(self.device)
target_classes = torch.full(src_logits.shape[:2], 0, dtype=torch.int64, device=src_logits.device)
target_classes[idx] = target_classes_o
loss_ce = F.cross_entropy(src_logits.transpose(1, 2), target_classes, self.empty_weight)
losses = {"loss_ce": loss_ce}
return losses
def loss_masks(self, outputs, targets, indices, num_masks):
"""Compute the losses related to the masks: the focal loss and the dice loss.
targets dicts must contain the key "masks" containing a tensor of dim [nb_target_boxes, h, w]
"""
assert "pred_masks" in outputs
src_idx = self._get_src_permutation_idx(indices)
tgt_idx = self._get_tgt_permutation_idx(indices)
src_masks = outputs["pred_masks"] #
src_masks = src_masks[src_idx]
masks = [t["masks"] for t in targets]
# TODO use valid to mask invalid areas due to padding in loss
target_masks, valid = nested_tensor_from_tensor_list(masks).decompose()
target_masks = target_masks.to(src_masks)
target_masks = target_masks[tgt_idx]
# ===================================================================================
# No need to upsample predictions as we are using normalized coordinates :)
# N x 1 x H x W
# src_masks = src_masks[:, None]
# target_masks = target_masks[:, None]
# with torch.no_grad():
# # sample point_coords
# point_coords = get_uncertain_point_coords_with_randomness(
# src_masks,
# lambda logits: calculate_uncertainty(logits),
# self.num_points,
# self.oversample_ratio,
# self.importance_sample_ratio,
# )
# # get gt labels
# point_labels = point_sample(
# target_masks,
# point_coords,
# align_corners=False,
# ).squeeze(1)
# point_logits = point_sample(
# src_masks,
# point_coords,
# align_corners=False,
# ).squeeze(1)
# ===================================================================================
point_logits = src_masks.flatten(1)
point_labels = target_masks.flatten(1)
losses = {
"loss_mask": sigmoid_ce_loss(point_logits, point_labels, num_masks), # sigmoid_focal_loss(point_logits, point_labels, num_masks), #
"loss_dice": dice_loss(point_logits, point_labels, num_masks)
}
del src_masks
del target_masks
return losses
def _get_src_permutation_idx(self, indices):
# permute predictions following indices
batch_idx = torch.cat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)])
src_idx = torch.cat([src for (src, _) in indices])
return batch_idx, src_idx
def _get_tgt_permutation_idx(self, indices):
# permute targets following indices
batch_idx = torch.cat([torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)])
tgt_idx = torch.cat([tgt for (_, tgt) in indices])
return batch_idx, tgt_idx
def _get_binary_mask(self, target):
y, x = target.size()
target_onehot = torch.zeros(self.num_classes + 1, y, x).to(target.device)
target_onehot = target_onehot.scatter(dim=0, index=target.unsqueeze(0), value=1)
return target_onehot
def get_loss(self, loss, outputs, targets, indices, num_masks):
loss_map = {
'labels': self.loss_labels,
'masks': self.loss_masks,
}
assert loss in loss_map, f"do you really want to compute {loss} loss?"
return loss_map[loss](outputs, targets, indices, num_masks)
def forward(self, outputs, gt_masks):
"""This performs the loss computation.
Parameters:
outputs: dict of tensors, see the output specification of the model for the format
gt_masks: [bs, h_net_output, w_net_output]
"""
outputs_without_aux = {k: v for k, v in outputs.items() if k != "aux_outputs"}
targets = self._get_targets(gt_masks)
# Retrieve the matching between the outputs of the last layer and the targets
indices = self.matcher(outputs_without_aux, targets)
# Compute the average number of target boxes accross all nodes, for normalization purposes
num_masks = sum(len(t["labels"]) for t in targets)
num_masks = torch.as_tensor([num_masks], dtype=torch.float, device=next(iter(outputs.values())).device)
if is_dist_avail_and_initialized():
torch.distributed.all_reduce(num_masks)
num_masks = torch.clamp(num_masks / get_world_size(), min=1).item()
# Compute all the requested losses
losses = {}
for loss in self.losses:
losses.update(self.get_loss(loss, outputs, targets, indices, num_masks))
# In case of auxiliary losses, we repeat this process with the output of each intermediate layer.
if "aux_outputs" in outputs:
for i, aux_outputs in enumerate(outputs["aux_outputs"]):
indices = self.matcher(aux_outputs, targets)
for loss in self.losses:
l_dict = self.get_loss(loss, aux_outputs, targets, indices, num_masks)
l_dict = {k + f"_{i}": v for k, v in l_dict.items()}
losses.update(l_dict)
return losses
def _get_targets(self, gt_masks):
targets = []
for mask in gt_masks:
binary_masks = self._get_binary_mask(mask)
cls_label = torch.unique(mask)
labels = cls_label[1:]
binary_masks = binary_masks[labels]
targets.append({'masks': binary_masks, 'labels': labels})
return targets
def __repr__(self):
head = "Criterion " + self.__class__.__name__
body = [
"matcher: {}".format(self.matcher.__repr__(_repr_indent=8)),
"losses: {}".format(self.losses),
"weight_dict: {}".format(self.weight_dict),
"num_classes: {}".format(self.num_classes),
"eos_coef: {}".format(self.eos_coef),
"num_points: {}".format(self.num_points),
"oversample_ratio: {}".format(self.oversample_ratio),
"importance_sample_ratio: {}".format(self.importance_sample_ratio),
]
_repr_indent = 4
lines = [head] + [" " * _repr_indent + line for line in body]
return "\n".join(lines)
class Criterion(object):
def __init__(self, num_classes, alpha=0.5, gamma=2, weight=None, ignore_index=0):
self.num_classes = num_classes
self.alpha = alpha
self.gamma = gamma
self.weight = weight
self.ignore_index = ignore_index
self.smooth = 1e-5
self.ce_fn = nn.CrossEntropyLoss(weight=self.weight, ignore_index=self.ignore_index, reduction='none')
def get_loss(self, outputs, gt_masks):
"""This performs the loss computation.
Parameters:
outputs: dict of tensors, see the output specification of the model for the format
gt_masks: [bs, h_net_output, w_net_output]
"""
loss_labels = 0.0
loss_masks = 0.0
loss_dices = 0.0
num = gt_masks.shape[0]
pred_logits = [outputs["pred_logits"].float()] # [bs, num_query, num_classes + 1]
pred_masks = [outputs['pred_masks'].float()] # [bs, num_query, h, w]
targets = self._get_targets(gt_masks, pred_logits[0].shape[1], pred_logits[0].device)
for aux_output in outputs['aux_outputs']:
pred_logits.append(aux_output["pred_logits"].float())
pred_masks.append(aux_output["pred_masks"].float())
gt_label = targets['labels'] # [bs, num_query]
gt_mask_list = targets['masks']
for mask_cls, pred_mask in zip(pred_logits, pred_masks):
loss_labels += F.cross_entropy(mask_cls.transpose(1, 2), gt_label)
# loss_masks += self.focal_loss(pred_result, gt_masks.to(pred_result.device))
loss_dices += self.dice_loss(pred_mask, gt_mask_list)
return loss_labels/num, loss_dices/num
def binary_dice_loss(self, inputs, targets):
inputs = inputs.sigmoid()
inputs = inputs.flatten(1)
targets = targets.flatten(1)
numerator = 2 * torch.einsum("nc,mc->nm", inputs, targets)
denominator = inputs.sum(-1)[:, None] + targets.sum(-1)[None, :]
loss = 1 - (numerator + 1) / (denominator + 1)
return loss.mean()
def dice_loss(self, predict, targets):
bs = predict.shape[0]
total_loss = 0
for i in range(bs):
pred_mask = predict[i]
tgt_mask = targets[i].to(predict.device)
dice_loss_value = self.binary_dice_loss(pred_mask, tgt_mask)
total_loss += dice_loss_value
return total_loss/bs
def focal_loss(self, preds, labels):
"""
preds: [bs, num_class + 1, h, w]
labels: [bs, h, w]
"""
logpt = -self.ce_fn(preds, labels)
pt = torch.exp(logpt)
loss = -((1 - pt) ** self.gamma) * self.alpha * logpt
return loss.mean()
def _get_binary_mask(self, target):
y, x = target.size()
target_onehot = torch.zeros(self.num_classes + 1, y, x)
target_onehot = target_onehot.scatter(dim=0, index=target.unsqueeze(0), value=1)
return target_onehot
def _get_targets(self, gt_masks, num_query, device):
binary_masks = []
gt_labels = []
for mask in gt_masks:
mask_onehot = self._get_binary_mask(mask)
cls_label = torch.unique(mask)
labels = torch.full((num_query,), 0, dtype=torch.int64, device=gt_masks.device)
labels[:len(cls_label)] = cls_label
binary_masks.append(mask_onehot[cls_label])
gt_labels.append(labels)
return {"labels": torch.stack(gt_labels).to(device), "masks": binary_masks}
|