File size: 14,180 Bytes
b213d84 |
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 |
# Copyright (c) Facebook, Inc. and its affiliates.
from typing import Any, List
import torch
from torch.nn import functional as F
from detectron2.config import CfgNode
from detectron2.structures import Instances
from .mask_or_segm import MaskOrSegmentationLoss
from .registry import DENSEPOSE_LOSS_REGISTRY
from .utils import (
BilinearInterpolationHelper,
ChartBasedAnnotationsAccumulator,
LossDict,
extract_packed_annotations_from_matches,
)
@DENSEPOSE_LOSS_REGISTRY.register()
class DensePoseChartLoss:
"""
DensePose loss for chart-based training. A mesh is split into charts,
each chart is given a label (I) and parametrized by 2 coordinates referred to
as U and V. Ground truth consists of a number of points annotated with
I, U and V values and coarse segmentation S defined for all pixels of the
object bounding box. In some cases (see `COARSE_SEGM_TRAINED_BY_MASKS`),
semantic segmentation annotations can be used as ground truth inputs as well.
Estimated values are tensors:
* U coordinates, tensor of shape [N, C, S, S]
* V coordinates, tensor of shape [N, C, S, S]
* fine segmentation estimates, tensor of shape [N, C, S, S] with raw unnormalized
scores for each fine segmentation label at each location
* coarse segmentation estimates, tensor of shape [N, D, S, S] with raw unnormalized
scores for each coarse segmentation label at each location
where N is the number of detections, C is the number of fine segmentation
labels, S is the estimate size ( = width = height) and D is the number of
coarse segmentation channels.
The losses are:
* regression (smooth L1) loss for U and V coordinates
* cross entropy loss for fine (I) and coarse (S) segmentations
Each loss has an associated weight
"""
def __init__(self, cfg: CfgNode):
"""
Initialize chart-based loss from configuration options
Args:
cfg (CfgNode): configuration options
"""
# fmt: off
self.heatmap_size = cfg.MODEL.ROI_DENSEPOSE_HEAD.HEATMAP_SIZE
self.w_points = cfg.MODEL.ROI_DENSEPOSE_HEAD.POINT_REGRESSION_WEIGHTS
self.w_part = cfg.MODEL.ROI_DENSEPOSE_HEAD.PART_WEIGHTS
self.w_segm = cfg.MODEL.ROI_DENSEPOSE_HEAD.INDEX_WEIGHTS
self.n_segm_chan = cfg.MODEL.ROI_DENSEPOSE_HEAD.NUM_COARSE_SEGM_CHANNELS
# fmt: on
self.segm_trained_by_masks = cfg.MODEL.ROI_DENSEPOSE_HEAD.COARSE_SEGM_TRAINED_BY_MASKS
self.segm_loss = MaskOrSegmentationLoss(cfg)
def __call__(
self, proposals_with_gt: List[Instances], densepose_predictor_outputs: Any, **kwargs
) -> LossDict:
"""
Produce chart-based DensePose losses
Args:
proposals_with_gt (list of Instances): detections with associated ground truth data
densepose_predictor_outputs: an object of a dataclass that contains predictor outputs
with estimated values; assumed to have the following attributes:
* coarse_segm - coarse segmentation estimates, tensor of shape [N, D, S, S]
* fine_segm - fine segmentation estimates, tensor of shape [N, C, S, S]
* u - U coordinate estimates per fine labels, tensor of shape [N, C, S, S]
* v - V coordinate estimates per fine labels, tensor of shape [N, C, S, S]
where N is the number of detections, C is the number of fine segmentation
labels, S is the estimate size ( = width = height) and D is the number of
coarse segmentation channels.
Return:
dict: str -> tensor: dict of losses with the following entries:
* `loss_densepose_U`: smooth L1 loss for U coordinate estimates
* `loss_densepose_V`: smooth L1 loss for V coordinate estimates
* `loss_densepose_I`: cross entropy for raw unnormalized scores for fine
segmentation estimates given ground truth labels;
* `loss_densepose_S`: cross entropy for raw unnormalized scores for coarse
segmentation estimates given ground truth labels;
"""
# densepose outputs are computed for all images and all bounding boxes;
# i.e. if a batch has 4 images with (3, 1, 2, 1) proposals respectively,
# the outputs will have size(0) == 3+1+2+1 == 7
if not len(proposals_with_gt):
return self.produce_fake_densepose_losses(densepose_predictor_outputs)
accumulator = ChartBasedAnnotationsAccumulator()
packed_annotations = extract_packed_annotations_from_matches(proposals_with_gt, accumulator)
# NOTE: we need to keep the same computation graph on all the GPUs to
# perform reduction properly. Hence even if we have no data on one
# of the GPUs, we still need to generate the computation graph.
# Add fake (zero) loss in the form Tensor.sum() * 0
if packed_annotations is None:
return self.produce_fake_densepose_losses(densepose_predictor_outputs)
h, w = densepose_predictor_outputs.u.shape[2:]
interpolator = BilinearInterpolationHelper.from_matches(
packed_annotations,
(h, w),
)
j_valid_fg = interpolator.j_valid * ( # pyre-ignore[16]
packed_annotations.fine_segm_labels_gt > 0
)
# pyre-fixme[6]: For 1st param expected `Tensor` but got `int`.
if not torch.any(j_valid_fg):
return self.produce_fake_densepose_losses(densepose_predictor_outputs)
losses_uv = self.produce_densepose_losses_uv(
proposals_with_gt,
densepose_predictor_outputs,
packed_annotations,
interpolator,
j_valid_fg, # pyre-ignore[6]
)
losses_segm = self.produce_densepose_losses_segm(
proposals_with_gt,
densepose_predictor_outputs,
packed_annotations,
interpolator,
j_valid_fg, # pyre-ignore[6]
)
return {**losses_uv, **losses_segm}
def produce_fake_densepose_losses(self, densepose_predictor_outputs: Any) -> LossDict:
"""
Fake losses for fine segmentation and U/V coordinates. These are used when
no suitable ground truth data was found in a batch. The loss has a value 0
and is primarily used to construct the computation graph, so that
`DistributedDataParallel` has similar graphs on all GPUs and can perform
reduction properly.
Args:
densepose_predictor_outputs: DensePose predictor outputs, an object
of a dataclass that is assumed to have the following attributes:
* fine_segm - fine segmentation estimates, tensor of shape [N, C, S, S]
* u - U coordinate estimates per fine labels, tensor of shape [N, C, S, S]
* v - V coordinate estimates per fine labels, tensor of shape [N, C, S, S]
Return:
dict: str -> tensor: dict of losses with the following entries:
* `loss_densepose_U`: has value 0
* `loss_densepose_V`: has value 0
* `loss_densepose_I`: has value 0
* `loss_densepose_S`: has value 0
"""
losses_uv = self.produce_fake_densepose_losses_uv(densepose_predictor_outputs)
losses_segm = self.produce_fake_densepose_losses_segm(densepose_predictor_outputs)
return {**losses_uv, **losses_segm}
def produce_fake_densepose_losses_uv(self, densepose_predictor_outputs: Any) -> LossDict:
"""
Fake losses for U/V coordinates. These are used when no suitable ground
truth data was found in a batch. The loss has a value 0
and is primarily used to construct the computation graph, so that
`DistributedDataParallel` has similar graphs on all GPUs and can perform
reduction properly.
Args:
densepose_predictor_outputs: DensePose predictor outputs, an object
of a dataclass that is assumed to have the following attributes:
* u - U coordinate estimates per fine labels, tensor of shape [N, C, S, S]
* v - V coordinate estimates per fine labels, tensor of shape [N, C, S, S]
Return:
dict: str -> tensor: dict of losses with the following entries:
* `loss_densepose_U`: has value 0
* `loss_densepose_V`: has value 0
"""
return {
"loss_densepose_U": densepose_predictor_outputs.u.sum() * 0,
"loss_densepose_V": densepose_predictor_outputs.v.sum() * 0,
}
def produce_fake_densepose_losses_segm(self, densepose_predictor_outputs: Any) -> LossDict:
"""
Fake losses for fine / coarse segmentation. These are used when
no suitable ground truth data was found in a batch. The loss has a value 0
and is primarily used to construct the computation graph, so that
`DistributedDataParallel` has similar graphs on all GPUs and can perform
reduction properly.
Args:
densepose_predictor_outputs: DensePose predictor outputs, an object
of a dataclass that is assumed to have the following attributes:
* fine_segm - fine segmentation estimates, tensor of shape [N, C, S, S]
* coarse_segm - coarse segmentation estimates, tensor of shape [N, D, S, S]
Return:
dict: str -> tensor: dict of losses with the following entries:
* `loss_densepose_I`: has value 0
* `loss_densepose_S`: has value 0, added only if `segm_trained_by_masks` is False
"""
losses = {
"loss_densepose_I": densepose_predictor_outputs.fine_segm.sum() * 0,
"loss_densepose_S": self.segm_loss.fake_value(densepose_predictor_outputs),
}
return losses
def produce_densepose_losses_uv(
self,
proposals_with_gt: List[Instances],
densepose_predictor_outputs: Any,
packed_annotations: Any,
interpolator: BilinearInterpolationHelper,
j_valid_fg: torch.Tensor,
) -> LossDict:
"""
Compute losses for U/V coordinates: smooth L1 loss between
estimated coordinates and the ground truth.
Args:
proposals_with_gt (list of Instances): detections with associated ground truth data
densepose_predictor_outputs: DensePose predictor outputs, an object
of a dataclass that is assumed to have the following attributes:
* u - U coordinate estimates per fine labels, tensor of shape [N, C, S, S]
* v - V coordinate estimates per fine labels, tensor of shape [N, C, S, S]
Return:
dict: str -> tensor: dict of losses with the following entries:
* `loss_densepose_U`: smooth L1 loss for U coordinate estimates
* `loss_densepose_V`: smooth L1 loss for V coordinate estimates
"""
u_gt = packed_annotations.u_gt[j_valid_fg]
u_est = interpolator.extract_at_points(densepose_predictor_outputs.u)[j_valid_fg]
v_gt = packed_annotations.v_gt[j_valid_fg]
v_est = interpolator.extract_at_points(densepose_predictor_outputs.v)[j_valid_fg]
return {
"loss_densepose_U": F.smooth_l1_loss(u_est, u_gt, reduction="sum") * self.w_points,
"loss_densepose_V": F.smooth_l1_loss(v_est, v_gt, reduction="sum") * self.w_points,
}
def produce_densepose_losses_segm(
self,
proposals_with_gt: List[Instances],
densepose_predictor_outputs: Any,
packed_annotations: Any,
interpolator: BilinearInterpolationHelper,
j_valid_fg: torch.Tensor,
) -> LossDict:
"""
Losses for fine / coarse segmentation: cross-entropy
for segmentation unnormalized scores given ground truth labels at
annotated points for fine segmentation and dense mask annotations
for coarse segmentation.
Args:
proposals_with_gt (list of Instances): detections with associated ground truth data
densepose_predictor_outputs: DensePose predictor outputs, an object
of a dataclass that is assumed to have the following attributes:
* fine_segm - fine segmentation estimates, tensor of shape [N, C, S, S]
* coarse_segm - coarse segmentation estimates, tensor of shape [N, D, S, S]
Return:
dict: str -> tensor: dict of losses with the following entries:
* `loss_densepose_I`: cross entropy for raw unnormalized scores for fine
segmentation estimates given ground truth labels
* `loss_densepose_S`: cross entropy for raw unnormalized scores for coarse
segmentation estimates given ground truth labels;
may be included if coarse segmentation is only trained
using DensePose ground truth; if additional supervision through
instance segmentation data is performed (`segm_trained_by_masks` is True),
this loss is handled by `produce_mask_losses` instead
"""
fine_segm_gt = packed_annotations.fine_segm_labels_gt[
interpolator.j_valid # pyre-ignore[16]
]
fine_segm_est = interpolator.extract_at_points(
densepose_predictor_outputs.fine_segm,
slice_fine_segm=slice(None),
w_ylo_xlo=interpolator.w_ylo_xlo[:, None], # pyre-ignore[16]
w_ylo_xhi=interpolator.w_ylo_xhi[:, None], # pyre-ignore[16]
w_yhi_xlo=interpolator.w_yhi_xlo[:, None], # pyre-ignore[16]
w_yhi_xhi=interpolator.w_yhi_xhi[:, None], # pyre-ignore[16]
)[interpolator.j_valid, :]
return {
"loss_densepose_I": F.cross_entropy(fine_segm_est, fine_segm_gt.long()) * self.w_part,
"loss_densepose_S": self.segm_loss(
proposals_with_gt, densepose_predictor_outputs, packed_annotations
)
* self.w_segm,
}
|