def loss(self, cls_scores, bbox_preds, gt_bboxes, gt_labels, img_metas,
cfg):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_labels_list=gt_labels,
cls_out_channels=self.cls_out_channels,
sampling=False)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
losses_cls, losses_reg = multi_apply(self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_pos_samples=num_total_pos,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_reg=losses_reg)
def loss(self, rpn_cls_scores, rpn_bbox_preds, gt_bboxes, gt_bboxes_8_coo, img_shapes, coo_num, cfg):
featmap_sizes = [featmap.size()[-2:] for featmap in rpn_cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_shapes)
cls_reg_targets = anchor_target(
anchor_list, valid_flag_list, gt_bboxes, gt_bboxes_8_coo, img_shapes,
self.target_means, self.target_stds, coo_num, cfg)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
losses_cls, losses_reg = multi_apply(
self.loss_single,
rpn_cls_scores,
rpn_bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_pos + num_total_neg,
coo_num=coo_num,
cfg=cfg)
return dict(loss_rpn_cls=losses_cls, loss_rpn_reg=losses_reg)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=1,
sampling=False,
unmap_outputs=False)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_images = len(img_metas)
all_cls_scores = torch.cat([
s.permute(0, 2, 3, 1).reshape(
num_images, -1, self.cls_out_channels) for s in cls_scores
], 1)
all_labels = torch.cat(labels_list, -1).view(num_images, -1)
all_label_weights = torch.cat(label_weights_list,
-1).view(num_images, -1)
all_bbox_preds = torch.cat([
b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
for b in bbox_preds
], -2)
all_bbox_targets = torch.cat(bbox_targets_list,
-2).view(num_images, -1, 4)
all_bbox_weights = torch.cat(bbox_weights_list,
-2).view(num_images, -1, 4)
losses_cls, losses_bbox = multi_apply(self.loss_single,
all_cls_scores,
all_bbox_preds,
all_labels,
all_label_weights,
all_bbox_targets,
all_bbox_weights,
num_total_samples=num_total_pos,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
# anchor number of multi levels
num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
# concat all level anchors and flags to a single tensor
concat_anchor_list = []
for i in range(len(anchor_list)):
concat_anchor_list.append(torch.cat(anchor_list[i]))
all_anchor_list = images_to_levels(concat_anchor_list,
num_level_anchors)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
all_anchor_list,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(
self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None,
):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling,
)
if cls_reg_targets is None:
return None
(
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_pos,
num_total_neg,
) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg,
)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
features,
cls_scores,
bbox_preds,
teacher_features,
teacher_cls_scores,
teacher_bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
sampling = False if self.use_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos if self.use_focal_loss else
num_total_pos + num_total_neg)
losses_cls, losses_kd_cls, losses_reg = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
teacher_cls_scores,
teacher_bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
losses_at, = multi_apply(
attention_loss,
features,
teacher_features,
beta=cfg.teacher.beta)
return dict(loss_cls=losses_cls, loss_kd_cls=losses_kd_cls, loss_reg=losses_reg,
loss_at=losses_at)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None,
iteration=None):
# 3D images
featmap_sizes = [featmap.size()[-3:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
sampling = False if self.use_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling,
iteration=iteration)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg, anchors_list,
inside_flags) = cls_reg_targets
self.pos_indices = inside_flags
num_total_samples = (num_total_pos if self.use_focal_loss else
num_total_pos + num_total_neg)
level = {'level': 0}
losses_cls, losses_reg = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
anchors_list,
num_total_samples=num_total_samples,
cfg=cfg,
level=level,
gt_bboxes=gt_bboxes,
iteration=iteration)
return dict(loss_cls=losses_cls, loss_reg=losses_reg)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
# import time
# import datetime
# t1 = time.time()
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
# print('get_anchors in ', str(datetime.timedelta(seconds=time.time() - t1)))
# t2 = time.time()
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
# print('anchor_target in ', str(datetime.timedelta(seconds=time.time() - t2)))
# t3 = time.time()
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
# print('loss_single in ', str(datetime.timedelta(seconds=time.time() - t3)))
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
del featmap_sizes
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
del anchor_list, valid_flag_list, gt_bboxes, gt_bboxes_ignore, gt_labels, \
label_channels
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
if self.topk_func is not None and hasattr(
self, self.topk_func): # add topk_func
topk_func = getattr(self, self.topk_func)
self.ignore_topk = topk_func(self.init_ignore_topk, self.iter,
self.max_iters)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
img_metas,
num_total_samples=num_total_samples,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
sampling = False if self.use_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
# first get cls_reg_targets by anchor_target, there're the gt deltas.
# get target_deltas through gt_bboxes and anchor_list
# first sample the positive anchors by calculating the IoUs,
# then assign anchors to gt, then calculate the target deltas.
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos if self.use_focal_loss else
num_total_pos + num_total_neg)
# the predict values are the deltas,
# so just use the target to calculate the losses.
losses_cls, losses_reg = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_reg=losses_reg)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return dict(loss_cls=0, loss_reg=0)
# each tensor in the list is corresponding to a level.
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single_level,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
# each tensor in losses_cls and losses_reg is corresponding to a level.
return {'losses/loss_cls': losses_cls, 'losses/loss_bbox': losses_bbox}
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):#该函数对多个预测结果使用
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]#多少个结果，每个结果的大小
assert len(featmap_sizes) == len(self.anchor_generators)#保证特征层个数和anchor层数一致，保证anchor层和结果层一一对应
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas, device=device)#返回所有的anchor和有效的anchor索引
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)#使用GT来编码anchor
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (
num_total_pos + num_total_neg if self.sampling else num_total_pos)#所有的样本个数
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)#这里就是使用loss_single函数来计算两个损失
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def embed_loss(self,
embed_feats_list,
gt_bboxes,
gt_trackids,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in embed_feats_list]
assert len(featmap_sizes) == len(self.anchor_generators)
device = embed_feats_list[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cache_assigner = cfg.assigner
cfg.assigner = cfg.track_assigner
embed_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_trackids,
label_channels=label_channels,
sampling=self.sampling)
cfg.assigner = cache_assigner
if embed_targets is None:
return None
(track_targets, track_targets_weights_list, bbox_targets_list,
bbox_weights_list, num_total_pos, num_total_neg) = embed_targets
losses_triplet = list(
map(self.embed_loss_single, *(embed_feats_list, track_targets)))
return dict(loss_triplet=losses_triplet)
def loss(self,
cls_scores,
bbox_preds,
shape_preds,
loc_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.approx_generators)
# get loc targets
loc_targets, loc_weights, loc_avg_factor = ga_loc_target(
gt_bboxes,
featmap_sizes,
self.octave_base_scale,
self.anchor_strides,
center_ratio=cfg.center_ratio,
ignore_ratio=cfg.ignore_ratio)
# get sampled approxes
approxs_list, inside_flag_list = self.get_sampled_approxs(
featmap_sizes, img_metas, cfg)
# get squares and guided anchors
squares_list, guided_anchors_list, _ = self.get_anchors(
featmap_sizes, shape_preds, loc_preds, img_metas)
# get shape targets
sampling = False if not hasattr(cfg, 'ga_sampler') else True
shape_targets = ga_shape_target(approxs_list,
inside_flag_list,
squares_list,
gt_bboxes,
img_metas,
self.approxs_per_octave,
cfg,
sampling=sampling)
if shape_targets is None:
return None
(bbox_anchors_list, bbox_gts_list, anchor_weights_list, anchor_fg_num,
anchor_bg_num) = shape_targets
anchor_total_num = (anchor_fg_num if not sampling else anchor_fg_num +
anchor_bg_num)
# get anchor targets
sampling = False if self.cls_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(guided_anchors_list,
inside_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
# (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
# num_total_pos, num_total_neg) = cls_reg_targets
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg, level_anchor_list) = cls_reg_targets
num_total_samples = (num_total_pos if self.cls_focal_loss else
num_total_pos + num_total_neg)
# added by WSK
num_total_samples_cls = (num_total_pos_cls if self.cls_focal_loss else
num_total_pos_cls + num_total_neg_cls)
# get classification and bbox regression losses
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
level_anchor_list, # added by Shengkai Wu
bbox_targets_cls_list, # added by Shengkai Wu
bbox_weights_cls_list, # added by Shengkai Wu
num_total_samples_cls=num_total_samples_cls, # added by Shengkai Wu
num_total_samples=num_total_samples,
gt_bboxes=gt_bboxes, # added by Shengkai Wu
cfg=cfg)
# get anchor location loss
losses_loc = []
for i in range(len(loc_preds)):
loss_loc = self.loss_loc_single(loc_preds[i],
loc_targets[i],
loc_weights[i],
loc_avg_factor=loc_avg_factor,
cfg=cfg)
losses_loc.append(loss_loc)
# get anchor shape loss
losses_shape = []
for i in range(len(shape_preds)):
loss_shape = self.loss_shape_single(
shape_preds[i],
bbox_anchors_list[i],
bbox_gts_list[i],
anchor_weights_list[i],
anchor_total_num=anchor_total_num)
losses_shape.append(loss_shape)
return dict(loss_cls=losses_cls,
loss_bbox=losses_bbox,
loss_shape=losses_shape,
loss_loc=losses_loc)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
anchor_list0 = anchor_list.copy()
valid_flag_list0 = valid_flag_list.copy()
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cfg['sampler']['pos_fraction'] = 0.5
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
if 1:
lvls = len(anchor_list0)
bboxes_lst = [[] for k in range(lvls)]
for z in range(lvls):
for k in range(len(anchor_list0[z])):
bbox_pred = bbox_preds[k][z].permute(1, 2,
0).reshape(-1, 4)
a = anchor_list0[z][k]
# print(a.shape,bbox_pred.shape)
bboxes = delta2bbox(anchor_list0[z][k], bbox_pred,
self.target_means, self.target_stds)
bboxes_lst[z].append(bboxes.clone())
zz = 0
# label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cfg['sampler']['pos_fraction'] = 0.55 #0.55
cls_reg_targets = anchor_target(
bboxes_lst,
valid_flag_list0,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list,
bbox_weights_list, num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos + num_total_neg
if self.sampling else num_total_pos)
losses_cls2, losses_bbox2 = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
for k in range(lvls):
losses_cls[k] = losses_cls[k] / 2 + losses_cls2[k] / 2
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
cls_scores,
bbox_preds,
shape_preds,
loc_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.approx_generators)
device = cls_scores[0].device
# get loc targets
loc_targets, loc_weights, loc_avg_factor = ga_loc_target(
gt_bboxes,
featmap_sizes,
self.octave_base_scale,
self.anchor_strides,
center_ratio=cfg.center_ratio,
ignore_ratio=cfg.ignore_ratio)
# get sampled approxes
approxs_list, inside_flag_list = self.get_sampled_approxs(
featmap_sizes, img_metas, cfg, device=device)
# get squares and guided anchors
squares_list, guided_anchors_list, _ = self.get_anchors(
featmap_sizes, shape_preds, loc_preds, img_metas, device=device)
# get shape targets
sampling = False if not hasattr(cfg, 'ga_sampler') else True
shape_targets = ga_shape_target(
approxs_list,
inside_flag_list,
squares_list,
gt_bboxes,
img_metas,
self.approxs_per_octave,
cfg,
sampling=sampling)
if shape_targets is None:
return None
(bbox_anchors_list, bbox_gts_list, anchor_weights_list, anchor_fg_num,
anchor_bg_num) = shape_targets
anchor_total_num = (
anchor_fg_num if not sampling else anchor_fg_num + anchor_bg_num)
# get anchor targets
sampling = False if self.cls_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
# get classification and bbox regression losses
cfg['ga_sampler']['pos_fraction'] = 0.5 # 0.55
cfg['sampler']['pos_fraction'] = 0.5 # 0.55
cls_reg_targets = anchor_target(
guided_anchors_list,
inside_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (
num_total_pos if self.cls_focal_loss else num_total_pos +
num_total_neg)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
# get anchor location loss
losses_loc = []
for i in range(len(loc_preds)):
loss_loc = self.loss_loc_single(
loc_preds[i],
loc_targets[i],
loc_weights[i],
loc_avg_factor=loc_avg_factor,
cfg=cfg)
losses_loc.append(loss_loc)
# get anchor shape loss
losses_shape = []
for i in range(len(shape_preds)):
loss_shape = self.loss_shape_single(
shape_preds[i],
bbox_anchors_list[i],
bbox_gts_list[i],
anchor_weights_list[i],
anchor_total_num=anchor_total_num)
losses_shape.append(loss_shape)
if 1:
# get sampled approxes
approxs_list, inside_flag_list = self.get_sampled_approxs(
featmap_sizes, img_metas, cfg, device=device)
# get squares and guided anchors
squares_list, guided_anchors_list, _ = self.get_anchors(
featmap_sizes, shape_preds, loc_preds, img_metas, device=device)
# get shape targets
sampling = False if not hasattr(cfg, 'ga_sampler') else True
shape_targets = ga_shape_target(
approxs_list,
inside_flag_list,
squares_list,
gt_bboxes,
img_metas,
self.approxs_per_octave,
cfg,
sampling=sampling)
if shape_targets is None:
return None
(bbox_anchors_list, bbox_gts_list, anchor_weights_list, anchor_fg_num,
anchor_bg_num) = shape_targets
anchor_total_num = (
anchor_fg_num if not sampling else anchor_fg_num + anchor_bg_num)
# get anchor targets
sampling = False if self.cls_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
lvls = len(guided_anchors_list)
bboxes_lst = [[] for k in range(lvls)]
for z in range(lvls):
for k in range(len(guided_anchors_list[z])):
bbox_pred = bbox_preds[k][z].permute(1, 2, 0).reshape(-1, 4)
a = guided_anchors_list[z][k]
# print(a.shape,bbox_pred.shape)
bboxes = delta2bbox(guided_anchors_list[z][k], bbox_pred, self.target_means,
self.target_stds)
bboxes_lst[z].append(bboxes.clone())
zz = 0
# label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
# label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cfg['ga_sampler']['pos_fraction'] = 0.55 # 0.55
cfg['sampler']['pos_fraction'] = 0.6 # 0.55
cls_reg_targets = anchor_target(
guided_anchors_list,
inside_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (
num_total_pos if self.cls_focal_loss else num_total_pos +
num_total_neg)
# get classification and bbox regression losses
losses_cls2, losses_bbox2 = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
zz=0
for k in range(lvls):
losses_cls[k] = losses_cls[k] / 2 + losses_cls2[k] / 2
return dict(
loss_cls=losses_cls,
loss_bbox=losses_bbox,
loss_shape=losses_shape,
loss_loc=losses_loc)
def loss(self,
fam_cls_scores,
fam_bbox_preds,
refine_anchors,
odm_cls_scores,
odm_bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in odm_cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
# check for size zero boxes
for img_nr in range(len(gt_bboxes)):
zero_inds = gt_bboxes[img_nr][:, 2:4] == 0
gt_bboxes[img_nr][:, 2:4][zero_inds] = 1
device = odm_cls_scores[0].device
anchor_list, valid_flag_list = self.get_init_anchors(featmap_sizes,
img_metas,
device=device)
# anchor number of multi levels
num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
# concat all level anchors and flags to a single tensor
concat_anchor_list = []
for i in range(len(anchor_list)):
concat_anchor_list.append(torch.cat(anchor_list[i]))
all_anchor_list = images_to_levels(concat_anchor_list,
num_level_anchors)
# Feature Alignment Module
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg.fam_cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_fam_cls, losses_fam_bbox = multi_apply(
self.loss_fam_single,
fam_cls_scores,
fam_bbox_preds,
all_anchor_list,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg.fam_cfg)
# Oriented Detection Module targets
refine_anchors_list, valid_flag_list = self.get_refine_anchors(
featmap_sizes, refine_anchors, img_metas, device=device)
# anchor number of multi levels
num_level_anchors = [
anchors.size(0) for anchors in refine_anchors_list[0]
]
# concat all level anchors and flags to a single tensor
concat_anchor_list = []
for i in range(len(refine_anchors_list)):
concat_anchor_list.append(torch.cat(refine_anchors_list[i]))
all_anchor_list = images_to_levels(concat_anchor_list,
num_level_anchors)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(refine_anchors_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg.odm_cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_odm_cls, losses_odm_bbox = multi_apply(
self.loss_odm_single,
odm_cls_scores,
odm_bbox_preds,
all_anchor_list,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg.odm_cfg)
self.last_vals = dict(
gt_bboxes=gt_bboxes,
gt_labels=gt_labels,
img_metas=img_metas,
fam_cls_scores=fam_cls_scores,
fam_bbox_preds=fam_bbox_preds,
refine_anchors=refine_anchors,
odm_cls_scores=odm_cls_scores,
odm_bbox_preds=odm_bbox_preds,
)
if sum(losses_fam_cls) > 1E10 or \
sum(losses_fam_bbox) > 1E10 or \
sum(losses_odm_cls) > 1E10 or \
sum(losses_odm_bbox) > 1E10:
print("bad loss")
return dict(loss_fam_cls=losses_fam_cls,
loss_fam_bbox=losses_fam_bbox,
loss_odm_cls=losses_odm_cls,
loss_odm_bbox=losses_odm_bbox)
def loss(self,
cls_scores,
bbox_preds,
shape_preds,
loc_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.approx_generators)
loc_targets, loc_weights, loc_avg_factor = ga_loc_target(
gt_bboxes,
featmap_sizes,
self.octave_base_scale,
self.anchor_strides,
center_ratio=cfg.center_ratio,
ignore_ratio=cfg.ignore_ratio)
(approxs_list, valid_flag_list, base_approxs_list,
guided_anchors_list) = self.get_anchors(featmap_sizes, shape_preds,
img_metas)
sampling = False if not hasattr(cfg, 'ga_sampler') else True
shape_targets = ga_shape_target(approxs_list,
valid_flag_list,
base_approxs_list,
gt_bboxes,
img_metas,
self.approxs_per_octave,
cfg,
sampling=sampling)
if shape_targets is None:
return None
(bbox_anchors_list, bbox_gts_list, anchor_weights_list,
all_inside_flags, anchor_fg_num, anchor_ng_num) = shape_targets
anchor_total_num = (anchor_fg_num if not sampling else anchor_fg_num +
anchor_ng_num)
sampling = False if self.cls_focal_loss else True
label_channels = self.cls_out_channels if self.cls_sigmoid_loss else 1
cls_reg_targets = anchor_target(guided_anchors_list,
all_inside_flags,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos if self.cls_focal_loss else
num_total_pos + num_total_neg)
losses_cls, losses_reg = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
losses_loc, = multi_apply(self.loss_loc_single,
loc_preds,
loc_targets,
loc_weights,
loc_avg_factor=loc_avg_factor,
cfg=cfg)
losses_shape, = multi_apply(self.loss_shape_single,
shape_preds,
bbox_anchors_list,
bbox_gts_list,
anchor_weights_list,
anchor_total_num=anchor_total_num)
return dict(loss_cls=losses_cls,
loss_reg=losses_reg,
loss_shape=losses_shape,
loss_loc=losses_loc)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
# img_meta = dict(
# ori_shape=ori_shape,
# img_shape=img_shape,
# pad_shape=pad_shape,
# scale_factor=scale_factor,
# flip=flip)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
cls_reg_targets = anchor_target(
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=1,
sampling=False,
unmap_outputs=False)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_images = len(img_metas)
# cls_scores:[B(num_images), C, H, W] => [B, H, W, C] => [B, self.num_anchors x H x W (8732), self.cls_out_channels]
all_cls_scores = torch.cat([
s.permute(0, 2, 3, 1).reshape(
num_images, -1, self.cls_out_channels) for s in cls_scores
], 1)
# all_labels : [num_images, 8732]
all_labels = torch.cat(labels_list, -1).view(num_images, -1)
# all_label_weights : [num_images, 8732]
all_label_weights = torch.cat(label_weights_list, -1).view(
num_images, -1)
# bbox_preds: [B(num_images), C, H, W] => [B, H, W, C] => [B, self.num_anchors x H x W (8732), 4]
all_bbox_preds = torch.cat([
b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
for b in bbox_preds
], -2)
# all_bbox_targets:[num_images, 8732, 4]
all_bbox_targets = torch.cat(bbox_targets_list, -2).view(
num_images, -1, 4)
# all_bbox_weights:[num_images, 8732, 4]
all_bbox_weights = torch.cat(bbox_weights_list, -2).view(
num_images, -1, 4)
losses_cls, losses_reg = multi_apply(
self.loss_single,
all_cls_scores,
all_bbox_preds,
all_labels,
all_label_weights,
all_bbox_targets,
all_bbox_weights,
num_total_samples=num_total_pos,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_reg=losses_reg)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
"""
:param cls_scores: list of tensors. len(cls_scores) is the number of levels for the feature pyramid.
cls_scores[i].size() = [num_images, A*cls_out_channels, width_i, height_i]
:param bbox_preds: list of tensors. len(bbox_preds) is the number of levels for the feature pyramid.
bbox_preds[i].size() = [num_images, A*4, width_i, height_i] or [num_images, A*5, width_i, height_i]
:param gt_bboxes:
:param gt_labels:
:param img_metas:
:param cfg:
:param gt_bboxes_ignore:
:return:
"""
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=1,
sampling=False,
unmap_outputs=False)
if cls_reg_targets is None:
return None
# modifyied by Shengkai Wu
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg, level_anchor_list) = cls_reg_targets
num_images = len(img_metas)
all_cls_scores = torch.cat([
s.permute(0, 2, 3, 1).reshape(
num_images, -1, self.cls_out_channels) for s in cls_scores
], 1)
all_labels = torch.cat(labels_list, -1).view(num_images, -1)
all_label_weights = torch.cat(label_weights_list,
-1).view(num_images, -1)
all_bbox_preds = torch.cat([
b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
for b in bbox_preds
], -2)
all_bbox_targets = torch.cat(bbox_targets_list,
-2).view(num_images, -1, 4)
all_bbox_weights = torch.cat(bbox_weights_list,
-2).view(num_images, -1, 4)
# added by Shengkai Wu
all_anchors = torch.cat(level_anchor_list, -2).view(num_images, -1, 4)
# num_total_examples = num_total_pos + num_total_neg
# all_cls_scores.size() = [num_images,num_total_examples , cls_out_channels]
# all_bbox_preds.size() = [num_image, num_total_examples, 4]
# all_labels.size() = [num_images, num_total_examples]
# all_label_weights.size() = [num_images, num_total_examples]
# all_bbox_preds.size() = [num_images, num_total_examples, 4]
# all_bbox_targets.size() = [num_images, num_total_examples, 4]
# all_bbox_weights.size() = [num_images, num_total_examples, 4]
losses_cls, losses_bbox = multi_apply(
self.loss_single,
all_cls_scores,
all_bbox_preds,
all_labels,
all_label_weights,
all_bbox_targets,
all_bbox_weights,
all_anchors, #added by Shengkai Wu
num_total_samples=num_total_pos,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
fam_cls_scores,
fam_bbox_preds,
refine_anchors,
odm_cls_scores,
odm_bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in odm_cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = odm_cls_scores[0].device
anchors_list, valid_flag_list = self.get_init_anchors(featmap_sizes,
img_metas,
device=device)
# Feature Alignment Module
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchors_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg.fam_cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_fam_cls, losses_fam_bbox = multi_apply(
self.loss_fam_single,
fam_cls_scores,
fam_bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg.fam_cfg)
# Oriented Detection Module targets
refine_anchors_list, valid_flag_list = self.get_refine_anchors(
featmap_sizes, refine_anchors, img_metas, device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(refine_anchors_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg.odm_cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_odm_cls, losses_odm_bbox = multi_apply(
self.loss_odm_single,
odm_cls_scores,
odm_bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg.odm_cfg)
return dict(loss_fam_cls=losses_fam_cls,
loss_fam_bbox=losses_fam_bbox,
loss_odm_cls=losses_odm_cls,
loss_odm_bbox=losses_odm_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
# 生成anchor，将anchor与gt_bbox匹配，生成正负样本，计算gt_delta(anchor与其对应的gt_bbox的delta)
# 得到每个anchor的label/label_weights/delta/delta_weights，以及pos_inds/neg_inds
# 利用这些使用交叉熵和SmoothL1Loss，得到loss
# 其中，get_anchor完成了anchor生成
# anchor_target完成了anchor与gt_bbox匹配(assigner) + 正负样本生成（sampler）+ gt_delta计算
# + 得到每个anchor的label/label_weights/delta/delta_weights，以及pos_inds/neg_inds（返回值）
# loss_single根据anchor_target的返回值完成了loss计算
featmap_sizes = [featmap.size()[-2:]
for featmap in cls_scores] # 特征图的w/h
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
# anchor_list:每张图的anchor
# valid_flag_list:anchor是否有效的flag，是否有效是根据bbox是否超出图片来计算的
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
# anchor_target: 将anchor和gt_bbox匹配，得到正样本和负样本,并用sampler将这些结果进行封装，方便之后使用
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
# cls_reg_target包含：
# labels:每个anchor对应的label
# label_weights:每个anchor cls_loss的权重，负样本权重为1，正样本权重可为1也可为其他值
# bbox_targets：每个anchor与其对应的gt_bbox之前的delta，用于回归
# bbox_weights: 每个anchor bbox_reg的权重，正样本为1，负样本为0
# pos_inds：anchor中正样本的索引
# neg_inds: anchor中负样本的索引
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
del cls_reg_targets, labels_list, label_weights_list, bbox_targets_list, bbox_weights_list
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = cls_scores[0].device
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=1,
sampling=False,
unmap_outputs=False)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_images = len(img_metas)
all_cls_scores = torch.cat([
s.permute(0, 2, 3, 1).reshape(
num_images, -1, self.cls_out_channels) for s in cls_scores
], 1)
all_labels = torch.cat(labels_list, -1).view(num_images, -1)
all_label_weights = torch.cat(label_weights_list,
-1).view(num_images, -1)
all_bbox_preds = torch.cat([
b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
for b in bbox_preds
], -2)
all_bbox_targets = torch.cat(bbox_targets_list,
-2).view(num_images, -1, 4)
all_bbox_weights = torch.cat(bbox_weights_list,
-2).view(num_images, -1, 4)
# check NaN and Inf
assert torch.isfinite(all_cls_scores).all().item(), \
'classification scores become infinite or NaN!'
assert torch.isfinite(all_bbox_preds).all().item(), \
'bbox predications become infinite or NaN!'
losses_cls, losses_bbox = multi_apply(self.loss_single_level,
all_cls_scores,
all_bbox_preds,
all_labels,
all_label_weights,
all_bbox_targets,
all_bbox_weights,
num_total_samples=num_total_pos,
cfg=cfg)
return {
'rpn_losses/cls_loss': losses_cls,
'rpn_losses/bbox_loss': losses_bbox
}
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
#print(len(img_metas))
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
"""
if cfg.online_select:
cls_reg_targets = self.anchor_target(
cls_scores.copy(),
bbox_preds.copy(),
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
else:
"""
cls_reg_targets = anchor_target(
#cls_scores.copy(),
#bbox_preds.copy(),
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
#print(num_total_samples)
losses_cls, losses_bbox = multi_apply( # levels
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def generate_anchor(self):
anchor_target()
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
if self.use_forest and isinstance(cls_scores[0], list):
featmap_sizes = [featmap[-1].size()[-2:] for featmap in cls_scores]
device = cls_scores[0][-1].device
num_im = cls_scores[0][-1].shape[0]
anchor_cls_scores = [[] for i in range(num_im)]
for i in range(num_im):
for j in range(len(cls_scores)):
anchor_cls_scores[i].append(
cls_scores[j][-1][i, ...].permute(1, 2, 0).reshape(
-1, self.cls_out_channels))
else:
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
device = cls_scores[0].device
anchor_cls_scores = None
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling,
cls_scores=anchor_cls_scores)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos +
num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox, parent_loss, forest = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg)
# print(type(parent_loss), len(parent_loss), len(parent_loss[0]))
if forest:
all_loss = dict()
all_loss['loss_bbox'] = losses_bbox
all_loss['loss_fine_grained_cls'] = losses_cls
for tree_idx in range(len(self.all_classes_num) - 1):
parent_loss_cls = []
for i in range(len(parent_loss)):
parent_loss_cls.append(parent_loss[i][tree_idx])
all_loss['loss_tree{}_parent_cls'.format(tree_idx +
1)] = parent_loss_cls
return all_loss
else:
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss_first(self,
loc_preds,
bbox_preds,
gt_bboxes,
img_metas,
cfg,
gt_labels=None,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in bbox_preds]
assert len(featmap_sizes) == len(self.anchor_generators)
device = bbox_preds[0].device
# get loc targets
loc_targets, loc_weights, loc_avg_factor = ga_loc_target(
gt_bboxes,
featmap_sizes,
self.anchor_scales,
self.anchor_strides,
center_ratio=cfg.center_ratio,
ignore_ratio=cfg.ignore_ratio)
anchor_list, valid_flag_list = self.get_anchors(featmap_sizes,
img_metas,
device=device)
# self.num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
# get anchor targets
sampling = False if self.cls_focal_loss else True
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (num_total_pos if self.cls_focal_loss else
num_total_pos + num_total_neg)
losses_bbox_first = []
# get bbox regression losses on the first stage
for i in range(len(bbox_preds)):
loss_bbox = self.loss_first_single(
bbox_preds[i],
bbox_targets_list[i],
bbox_weights_list[i],
num_total_samples=num_total_samples,
cfg=cfg)
losses_bbox_first.append(loss_bbox)
# get anchor location loss
losses_loc = []
for i in range(len(loc_preds)):
loss_loc = self.loss_loc_single(loc_preds[i],
loc_targets[i],
loc_weights[i],
loc_avg_factor=loc_avg_factor,
cfg=cfg)
losses_loc.append(loss_loc)
return dict(loss_rpn_bbox_first=losses_bbox_first, loss_loc=losses_loc)
def loss(self,
fam_cls_scores,
fam_bbox_preds,
refine_anchors,
odm_cls_scores,
odm_bbox_preds,
bboxes,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
featmap_sizes = [featmap.size()[-2:] for featmap in odm_cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
device = odm_cls_scores[0].device
anchors_list, valid_flag_list = self.get_init_anchors(
featmap_sizes, img_metas, device=device)
refine_anchors_list, valid_flag_list = self.get_refine_anchors(
featmap_sizes, refine_anchors, img_metas, device=device)
# anchor number of multi levels
num_level_anchors = [anchors.size(0) for anchors in anchors_list[0]]
# Feature Alignment Module
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(
anchors_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg.fam_cfg,
refine_anchors_list,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling,
reg_decoded_bbox=self.reg_decoded_bbox,
use_vfl=False)
if cls_reg_targets is None:
return None
#labels_list 是一个list，长度为feature map levels数，一般为5，里面的每一个元素是一个tensor，shape为[bs,anchor_nums],其余的类似
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (
num_total_pos + num_total_neg if self.sampling else num_total_pos)
all_anchor_list = images_to_levels(anchors_list, num_level_anchors)
losses_fam_cls, losses_fam_bbox = multi_apply(
self.loss_fam_single,
fam_cls_scores,
fam_bbox_preds,
all_anchor_list,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg.fam_cfg)
# Oriented Detection Module targets
refine_anchors_list, valid_flag_list = self.get_refine_anchors(
featmap_sizes, refine_anchors, img_metas, device=device)
num_level_anchors = [anchors.size(0) for anchors in refine_anchors_list[0]]
output_bboxes, _ = self.get_refine_anchors(
featmap_sizes, bboxes, img_metas, device=device)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(
refine_anchors_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg.odm_cfg,
output_bboxes,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling,
reg_decoded_bbox=self.reg_decoded_bbox,
use_vfl=self.use_vfl)
# cls_reg_targets = anchor_target_atss(
# refine_anchors_list,
# valid_flag_list,
# gt_bboxes,
# img_metas,
# self.target_means,
# self.target_stds,
# cfg.odm_cfg,
# gt_bboxes_ignore_list=gt_bboxes_ignore,
# gt_labels_list=gt_labels,
# label_channels=label_channels,
# sampling=self.sampling,
# reg_decoded_bbox=self.reg_decoded_bbox)
# cls_reg_targets = anchor_target_rotated(
# refine_anchors_list,
# valid_flag_list,
# gt_bboxes,
# img_metas,
# self.target_means,
# self.target_stds,
# cfg.odm_cfg,
# output_bboxes,
# gt_bboxes_ignore_list=gt_bboxes_ignore,
# gt_labels_list=gt_labels,
# label_channels=label_channels,
# sampling=self.sampling,
# reg_decoded_bbox=self.reg_decoded_bbox,
# use_vfl=self.use_vfl)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (
num_total_pos + num_total_neg if self.sampling else num_total_pos)
all_anchor_list = images_to_levels(refine_anchors_list, num_level_anchors)
losses_odm_cls, losses_odm_bbox = multi_apply(
self.loss_odm_single,
odm_cls_scores,
odm_bbox_preds,
all_anchor_list,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg.odm_cfg)
return dict(loss_fam_cls=losses_fam_cls,
loss_fam_bbox=losses_fam_bbox,
loss_odm_cls=losses_odm_cls,
loss_odm_bbox=losses_odm_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
"""
:param cls_scores: list[Tensor]. len(cls_scores) equals to the number of feature map levels.
and cls_scores[i].size() is (batch, A*C, width_i, height_i). width_i and height_i is the size
of the i-th level feature map.
:param bbox_preds: list[Tensor]. len(bbox_preds) equals to the number of feature map levels.
and bbox_preds[i].size() is (batch, A*4, width_i, height_i). width_i and height_i is the size
of the i-th level feature map.
:param gt_bboxes: list[Tensor],Ground truth bboxes of each image. store the top-left and bottom-right corners
in the image coordinte;
:param gt_labels:
:param img_metas: list[dict], Meta info of each image.
:param cfg:
:param gt_bboxes_ignore:
:return:
"""
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
cls_reg_targets = anchor_target(
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list, num_total_pos, num_total_neg,
level_anchor_list) = cls_reg_targets
# added by WSK
# If sampling is adopted, num_total_samples = num_total_pos + num_total_neg;
# otherwise, num_total_samples = num_total_pos. For 'FocalLoss', 'GHMC', 'IOUbalancedSigmoidFocalLoss',
# sampling is not adopted.
num_total_samples = (
num_total_pos + num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
level_anchor_list, # added by Shengkai Wu
num_total_samples=num_total_samples,
gt_bboxes = gt_bboxes, # added by Shengkai Wu
cfg=cfg)
return dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
def loss(self,
cls_scores,
bbox_preds,
gt_bboxes,
gt_labels,
img_metas,
cfg,
gt_bboxes_ignore=None):
self.global_step = self.global_step + 1
featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
assert len(featmap_sizes) == len(self.anchor_generators)
anchor_list, valid_flag_list = self.get_anchors(
featmap_sizes, img_metas)
label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
original_anchors = [[k for k in m] for m in anchor_list]
original_anchors = list(map(list, zip(*original_anchors)))
original_anchors = [torch.cat(m) for m in original_anchors]
cls_reg_targets = anchor_target(
anchor_list,
valid_flag_list,
gt_bboxes,
img_metas,
self.target_means,
self.target_stds,
cfg,
gt_bboxes_ignore_list=gt_bboxes_ignore,
gt_labels_list=gt_labels,
label_channels=label_channels,
sampling=self.sampling)
if cls_reg_targets is None:
return None
(labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
num_total_pos, num_total_neg) = cls_reg_targets
num_total_samples = (
num_total_pos + num_total_neg if self.sampling else num_total_pos)
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list,
bbox_targets_list,
bbox_weights_list,
num_total_samples=num_total_samples,
cfg=cfg,
reduction_override="none")
cls_scores_flatten = [cls_score.permute(0, 2, 3, 1).reshape(-1, self.cls_out_channels) for cls_score in
cls_scores]
bbox_preds_flatten = [bbox_pred.permute(0, 2, 3, 1).reshape(-1, 4) for bbox_pred in bbox_preds]
labels_list_flatten = [label.reshape(-1, 1) for label in labels_list]
label_weights_list_flatten = [label_weights.reshape(-1, 1) for label_weights in label_weights_list]
bbox_targets_list_flatten = [bbox_targets.reshape(-1, 4) for bbox_targets in bbox_targets_list]
bbox_weights_list_flatten = [bbox_weights.reshape(-1, 4) for bbox_weights in bbox_weights_list]
original_anchors_flatten = [anchors.reshape(-1, 4) for anchors in original_anchors]
losses_cls_flatten = [loss_cls.reshape(-1, self.cls_out_channels) for loss_cls in losses_cls]
losses_bbox_flatten = [loss_bbox.reshape(-1, 4) for loss_bbox in losses_bbox]
cls_scores_flatten = torch.cat(cls_scores_flatten, dim=0)
bbox_preds_flatten = torch.cat(bbox_preds_flatten, dim=0)
labels_list_flatten = torch.cat(labels_list_flatten, dim=0)
label_weights_list_flatten = torch.cat(label_weights_list_flatten, dim=0)
bbox_targets_list_flatten = torch.cat(bbox_targets_list_flatten, dim=0)
bbox_weights_list_flatten = torch.cat(bbox_weights_list_flatten, dim=0)
original_anchors_flatten = torch.cat(original_anchors_flatten, dim=0)
losses_cls_flatten = torch.cat(losses_cls_flatten, dim=0)
losses_bbox_flatten = torch.cat(losses_bbox_flatten, dim=0)
split_point = [m.shape[0] * m.shape[1] for m in label_weights_list]
label_weights, bbox_weights, losses = self.predict_weights(
cls_score=cls_scores_flatten, bbox_pred=bbox_preds_flatten, labels=labels_list_flatten,
label_weights=label_weights_list_flatten, bbox_targets=bbox_targets_list_flatten,
bbox_weights=bbox_weights_list_flatten, anchors=original_anchors_flatten, loss_cls=losses_cls_flatten,
loss_bbox=losses_bbox_flatten)
label_weights_list_new = torch.split(label_weights, split_point)
bbox_weights_list_new = torch.split(bbox_weights, split_point)
label_weights_list_new = [m.reshape(2, -1) for m in label_weights_list_new]
bbox_weights_list_new = [m.reshape(2, -1, 4) for m in bbox_weights_list_new]
losses_cls, losses_bbox = multi_apply(
self.loss_single,
cls_scores,
bbox_preds,
labels_list,
label_weights_list_new,
bbox_targets_list,
bbox_weights_list_new,
num_total_samples=num_total_samples,
cfg=cfg,
reduction_override=None)
losses.update(
dict(loss_cls=losses_cls, loss_bbox=losses_bbox)
)
return losses