def extract_feat(self, img, gt_bboxes):
if self.visual_modality:
x = super().extract_feat(img)[-1]
feats = self.maxpool(self.extractor([x], bbox2roi(gt_bboxes)))
return feats.view(feats.size(0), -1)
return None
def test_sabl_bbox_head_loss():
"""Tests bbox head loss when truth is empty and non-empty."""
self = SABLHead(
num_classes=4,
cls_in_channels=3,
reg_in_channels=3,
cls_out_channels=3,
reg_offset_out_channels=3,
reg_cls_out_channels=3,
roi_feat_size=7)
# Dummy proposals
proposal_list = [
torch.Tensor([[23.6667, 23.8757, 228.6326, 153.8874]]),
]
target_cfg = mmcv.Config(dict(pos_weight=1))
# Test bbox loss when truth is empty
gt_bboxes = [torch.empty((0, 4))]
gt_labels = [torch.LongTensor([])]
sampling_results = _dummy_bbox_sampling(proposal_list, gt_bboxes,
gt_labels)
bbox_targets = self.get_targets(sampling_results, gt_bboxes, gt_labels,
target_cfg)
labels, label_weights, bbox_targets, bbox_weights = bbox_targets
# Create dummy features "extracted" for each sampled bbox
num_sampled = sum(len(res.bboxes) for res in sampling_results)
rois = bbox2roi([res.bboxes for res in sampling_results])
dummy_feats = torch.rand(num_sampled, 3, 7, 7)
cls_scores, bbox_preds = self.forward(dummy_feats)
losses = self.loss(cls_scores, bbox_preds, rois, labels, label_weights,
bbox_targets, bbox_weights)
assert losses.get('loss_cls', 0) > 0, 'cls-loss should be non-zero'
assert losses.get('loss_bbox_cls',
0) == 0, 'empty gt bbox-cls-loss should be zero'
assert losses.get('loss_bbox_reg',
0) == 0, 'empty gt bbox-reg-loss should be zero'
# Test bbox loss when truth is non-empty
gt_bboxes = [
torch.Tensor([[23.6667, 23.8757, 238.6326, 151.8874]]),
]
gt_labels = [torch.LongTensor([2])]
sampling_results = _dummy_bbox_sampling(proposal_list, gt_bboxes,
gt_labels)
rois = bbox2roi([res.bboxes for res in sampling_results])
bbox_targets = self.get_targets(sampling_results, gt_bboxes, gt_labels,
target_cfg)
labels, label_weights, bbox_targets, bbox_weights = bbox_targets
# Create dummy features "extracted" for each sampled bbox
num_sampled = sum(len(res.bboxes) for res in sampling_results)
dummy_feats = torch.rand(num_sampled, 3, 7, 7)
cls_scores, bbox_preds = self.forward(dummy_feats)
losses = self.loss(cls_scores, bbox_preds, rois, labels, label_weights,
bbox_targets, bbox_weights)
assert losses.get('loss_bbox_cls',
0) > 0, 'empty gt bbox-cls-loss should be zero'
assert losses.get('loss_bbox_reg',
0) > 0, 'empty gt bbox-reg-loss should be zero'
def aug_test(self, features, proposal_list, img_metas, rescale=False):
"""Test with augmentations.
If rescale is False, then returned bboxes and masks will fit the scale
of imgs[0].
"""
rcnn_test_cfg = self.test_cfg
aug_bboxes = []
aug_scores = []
for x, img_meta in zip(features, img_metas):
# only one image in the batch
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
flip_direction = img_meta[0]['flip_direction']
proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
scale_factor, flip, flip_direction)
# "ms" in variable names means multi-stage
ms_scores = []
rois = bbox2roi([proposals])
for i in range(self.num_stages):
bbox_results = self._bbox_forward(i, x, rois)
ms_scores.append(bbox_results['cls_score'])
if i < self.num_stages - 1:
bbox_label = bbox_results['cls_score'][:, :-1].argmax(
dim=1)
rois = self.bbox_head[i].regress_by_class(
rois, bbox_label, bbox_results['bbox_pred'],
img_meta[0])
cls_score = sum(ms_scores) / float(len(ms_scores))
bboxes, scores = self.bbox_head[-1].get_bboxes(
rois,
cls_score,
bbox_results['bbox_pred'],
img_shape,
scale_factor,
rescale=False,
cfg=None)
aug_bboxes.append(bboxes)
aug_scores.append(scores)
# after merging, bboxes will be rescaled to the original image size
merged_bboxes, merged_scores = merge_aug_bboxes(
aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
rcnn_test_cfg.score_thr,
rcnn_test_cfg.nms,
rcnn_test_cfg.max_per_img)
bbox_result = bbox2result(det_bboxes, det_labels,
self.bbox_head[-1].num_classes)
if self.with_mask:
if det_bboxes.shape[0] == 0:
segm_result = [[[]
for _ in range(self.mask_head[-1].num_classes)]
]
else:
aug_masks = []
aug_img_metas = []
for x, img_meta in zip(features, img_metas):
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
flip_direction = img_meta[0]['flip_direction']
_bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
scale_factor, flip, flip_direction)
mask_rois = bbox2roi([_bboxes])
for i in range(self.num_stages):
mask_results = self._mask_forward(i, x, mask_rois)
aug_masks.append(
mask_results['mask_pred'].sigmoid().cpu().numpy())
aug_img_metas.append(img_meta)
merged_masks = merge_aug_masks(aug_masks, aug_img_metas,
self.test_cfg)
ori_shape = img_metas[0][0]['ori_shape']
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks,
det_bboxes,
det_labels,
rcnn_test_cfg,
ori_shape,
scale_factor=1.0,
rescale=False)
return [(bbox_result, segm_result)]
else:
return [bbox_result]
def aug_test(self, imgs, img_metas, proposals=None, rescale=False):
"""Test with augmentations.
If rescale is False, then returned bboxes and masks will fit the scale
of imgs[0].
"""
# recompute feats to save memory
proposal_list = self.aug_test_rpn(self.extract_feats(imgs), img_metas,
self.test_cfg.rpn)
rcnn_test_cfg = self.test_cfg.rcnn
aug_bboxes = []
aug_scores = []
for x, img_meta in zip(self.extract_feats(imgs), img_metas):
# only one image in the batch
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
scale_factor, flip)
# "ms" in variable names means multi-stage
ms_scores = []
rois = bbox2roi([proposals])
for i in range(self.num_stages):
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
bbox_feats = bbox_roi_extractor(
x[:len(bbox_roi_extractor.featmap_strides)], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
ms_scores.append(cls_score)
if i < self.num_stages - 1:
bbox_label = cls_score.argmax(dim=1)
rois = bbox_head.regress_by_class(rois, bbox_label,
bbox_pred, img_meta[0])
cls_score = sum(ms_scores) / float(len(ms_scores))
iou_pred = None
bboxes, scores = self.bbox_head[-1].get_det_bboxes(rois,
cls_score,
bbox_pred,
iou_pred,
img_shape,
scale_factor,
rescale=False,
cfg=None)
aug_bboxes.append(bboxes)
aug_scores.append(scores)
# after merging, bboxes will be rescaled to the original image size
merged_bboxes, merged_scores = merge_aug_bboxes(
aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
rcnn_test_cfg.score_thr,
rcnn_test_cfg.nms,
rcnn_test_cfg.max_per_img)
bbox_result = bbox2result(det_bboxes, det_labels,
self.bbox_head[-1].num_classes)
if self.with_mask:
if det_bboxes.shape[0] == 0:
segm_result = [[]
for _ in range(self.mask_head[-1].num_classes -
1)]
else:
aug_masks = []
aug_img_metas = []
for x, img_meta in zip(self.extract_feats(imgs), img_metas):
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
_bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
scale_factor, flip)
mask_rois = bbox2roi([_bboxes])
for i in range(self.num_stages):
mask_feats = self.mask_roi_extractor[i](
x[:len(self.mask_roi_extractor[i].featmap_strides
)], mask_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
mask_pred = self.mask_head[i](mask_feats)
aug_masks.append(mask_pred.sigmoid().cpu().numpy())
aug_img_metas.append(img_meta)
merged_masks = merge_aug_masks(aug_masks, aug_img_metas,
self.test_cfg.rcnn)
ori_shape = img_metas[0][0]['ori_shape']
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks,
det_bboxes,
det_labels,
rcnn_test_cfg,
ori_shape,
scale_factor=1.0,
rescale=False)
return bbox_result, segm_result
else:
return bbox_result
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_bboxes_ignore,
gt_labels,
ref_img, # images of reference frame
ref_bboxes, # gt bbox of reference frame
gt_pids, # gt ids of current frame bbox mapped to reference frame
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
ref_x = self.extract_feat(ref_img)
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs)
losses.update(rpn_losses)
proposal_inputs = rpn_outs + (img_meta, self.test_cfg.rpn)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(
self.train_cfg.rcnn.sampler, context=self)
num_imgs = img.size(0)
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(
proposal_list[i], gt_bboxes[i], gt_bboxes_ignore[i],
gt_labels[i], gt_pids[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
gt_pids[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
bbox_img_n = [res.bboxes.size(0) for res in sampling_results]
ref_rois = bbox2roi(ref_bboxes)
ref_bbox_img_n = [x.size(0) for x in ref_bboxes]
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
ref_bbox_feats = self.bbox_roi_extractor(
ref_x[:self.bbox_roi_extractor.num_inputs], ref_rois)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
# fetch bbox and object_id targets
bbox_targets, (ids, id_weights) = self.bbox_head.get_target(
sampling_results, gt_bboxes, gt_labels, self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
match_score = self.track_head(bbox_feats, ref_bbox_feats,
bbox_img_n, ref_bbox_img_n)
loss_match = self.track_head.loss(match_score,
ids, id_weights)
losses.update(loss_match)
# mask head forward and loss
if self.with_mask:
pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(
sampling_results, gt_masks, self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def simple_test(self, img, img_meta, proposals=None, rescale=False):
"""Run inference on a single image.
Args:
img (Tensor): must be in shape (N, C, H, W)
img_meta (list[dict]): a list with one dictionary element.
See `mmdet/datasets/pipelines/formatting.py:Collect` for
details of meta dicts.
proposals : if specified overrides rpn proposals
rescale (bool): if True returns boxes in original image space
Returns:
dict: results
"""
x = self.extract_feat(img)
proposal_list = self.simple_test_rpn(
x, img_meta, self.test_cfg.rpn) if proposals is None else proposals
img_shape = img_meta[0]['img_shape']
ori_shape = img_meta[0]['ori_shape']
scale_factor = img_meta[0]['scale_factor']
# "ms" in variable names means multi-stage
ms_bbox_result = {}
ms_segm_result = {}
ms_scores = []
rcnn_test_cfg = self.test_cfg.rcnn
rois = bbox2roi(proposal_list)
for i in range(self.num_stages):
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
bbox_feats = bbox_roi_extractor(
x[:len(bbox_roi_extractor.featmap_strides)], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
ms_scores.append(cls_score)
if i < self.num_stages - 1:
bbox_label = cls_score.argmax(dim=1)
rois = bbox_head.regress_by_class(rois, bbox_label, bbox_pred,
img_meta[0])
# average 3 stage cls score as final cls score, but the output box come from final stage
cls_score = sum(ms_scores) / self.num_stages
det_bboxes, det_labels = self.bbox_head[
-1].get_det_bboxes( # bbox_pred may be tuple
rois,
cls_score,
bbox_pred,
img_shape,
scale_factor,
rescale=rescale,
cfg=rcnn_test_cfg)
bbox_result = bbox2result(det_bboxes, det_labels,
self.bbox_head[-1].num_classes)
ms_bbox_result['ensemble'] = bbox_result
if self.with_mask:
if det_bboxes.shape[0] == 0:
mask_classes = self.mask_head[-1].num_classes - 1
segm_result = [[] for _ in range(mask_classes)]
else:
if isinstance(scale_factor, float): # aspect ratio fixed
_bboxes = (det_bboxes[:, :4] *
scale_factor if rescale else det_bboxes)
else:
_bboxes = (
det_bboxes[:, :4] *
torch.from_numpy(scale_factor).to(det_bboxes.device)
if rescale else det_bboxes)
mask_rois = bbox2roi([_bboxes])
aug_masks = []
for i in range(self.num_stages):
mask_roi_extractor = self.mask_roi_extractor[i]
mask_feats = mask_roi_extractor(
x[:len(mask_roi_extractor.featmap_strides)], mask_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
mask_pred = self.mask_head[i](mask_feats)
aug_masks.append(mask_pred.sigmoid().cpu().numpy())
merged_masks = merge_aug_masks(aug_masks,
[img_meta] * self.num_stages,
self.test_cfg.rcnn)
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks, _bboxes, det_labels, rcnn_test_cfg,
ori_shape, scale_factor, rescale)
ms_segm_result['ensemble'] = segm_result
if self.with_mask:
results = (ms_bbox_result['ensemble'], ms_segm_result['ensemble'])
else:
results = ms_bbox_result['ensemble']
return results
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
"""
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_meta (list[dict]): list of image info dict where each dict has:
'img_shape', 'scale_factor', 'flip', and my also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
gt_bboxes (list[Tensor]): each item are the truth boxes for each
image in [tl_x, tl_y, br_x, br_y] format.
gt_labels (list[Tensor]): class indices corresponding to each box
gt_bboxes_ignore (None | list[Tensor]): specify which bounding
boxes can be ignored when computing the loss.
gt_masks (None | Tensor) : true segmentation masks for each box
used if the architecture supports a segmentation task.
proposals : override rpn proposals with custom proposals. Use when
`with_rpn` is False.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
x = self.extract_feat(img) # x得到的是一个元组，对于resnet，可包含多个conv层的结果
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x) # rpn_outs = rpn_cls_score + rpn_bbox_pred
# 注意:rpn_bbox_pred预测的是delta，是(tx,ty,tw,th)
# 计算loss时使用的是train的配置
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
# loss:生成anchor，将anchor与gt_bbox匹配，生成正负样本，计算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
rpn_losses = self.rpn_head.loss(
*rpn_loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
# rpn_losses 字典类型，包含loss_rpn_cls,loss_rpn_bbox
losses.update(rpn_losses) # 将rpn_loss加入到总体的loss中
# 得到proposal时使用的是train的配置
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
# get_bboxes:得到nms之后的proposals
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals
'''
with_bbox和with_mask是训练时的选项
eg.在faster_rcnn_r50_caffe_c4_1x.py中
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
'''
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(
self.train_cfg.rcnn.sampler, context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
# 在RPN部分完成了anchor的匹配和正负样本生成，用于计算loss
# 并根据bbox_pred和NMS得到proposal
# 在RCNN部分，将proposal和gt_bbox匹配并生成正负样本，用于计算这部分的loss
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# 所有图的roi，（img_ind, x1, y1, x2, y2)
# TODO: a more flexible way to decide which feature maps to use
# todo:看SingleRoIExtractor
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
# bbox_feats尺寸(roi_num, roi_out_channels, roi_w(7), roi_h(7))
if self.with_shared_head:
# 将ROI再经过一段特征提取层，这里用的是Resnet的conv5
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(
res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(
res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(sampling_results,
gt_masks,
self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def simple_test_mask_(self,
x,
img_metas,
det_bboxes,
det_labels,
semantic_logits,
rescale=False):
ori_shape = img_metas[0]['ori_shape']
scale_factor = img_metas[0]['scale_factor']
ref_size = (np.int(np.round(ori_shape[0] * scale_factor)),
np.int(np.round(ori_shape[1] * scale_factor)))
sem_pred = torch.argmax(semantic_logits, dim=1)[0]
panoptic_mask = torch.zeros_like(sem_pred, dtype=torch.long)
cat = [255]
if det_bboxes.shape[0] == 0:
intermediate_logits = semantic_logits[0, :self.num_stuff]
else:
# if det_bboxes is rescaled to the original image size, we need to
# rescale it back to the testing scale to obtain RoIs.
if rescale and not isinstance(scale_factor, float):
scale_factor = torch.from_numpy(scale_factor).to(
det_bboxes.device)
_bboxes = (det_bboxes[:, :4] *
scale_factor if rescale else det_bboxes)
mask_rois = bbox2roi([_bboxes])
mask_feats = self.mask_roi_extractor(
x[:len(self.mask_roi_extractor.featmap_strides)], mask_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
mask_pred = self.mask_head(mask_feats)
confidence = det_bboxes[:, 4]
idx = torch.argsort(confidence, descending=True)
bbx_inv = invert_roi_bbx(det_bboxes[:, :4],
tuple(mask_pred.shape[2:]), ref_size)
bbx_idx = torch.arange(0,
det_bboxes.size(0),
dtype=torch.long,
device=det_bboxes.device)
mask_pred = roi_sampling(mask_pred,
bbx_inv,
bbx_idx,
ref_size,
padding="zero")
ML_A = mask_pred.new_zeros(mask_pred.shape[0], mask_pred.shape[-2],
mask_pred.shape[-1])
ML_B = ML_A.clone()
occupied = torch.zeros_like(sem_pred, dtype=torch.bool)
i = 0
for id_i in idx:
label_i = det_labels[id_i]
mask_pred_i = mask_pred[id_i, label_i + 1, :, :]
mask_i = (mask_pred_i.sigmoid() >
self.test_cfg.rcnn.mask_thr_binary)
mask_i = mask_i.type(torch.bool)
intersection = occupied & mask_i
if intersection.float().sum() / mask_i.float().sum(
) > self.test_cfg.panoptic.overlap_thr:
continue
mask_i = mask_i ^ intersection
occupied += mask_i
y0 = max(int(det_bboxes[id_i, 1] + 1), 0)
y1 = min(int((det_bboxes[id_i, 3] - 1).round() + 1),
ref_size[0])
x0 = max(int(det_bboxes[id_i, 0] + 1), 0)
x1 = min(int((det_bboxes[id_i, 2] - 1).round() + 1),
ref_size[1])
ML_A[i] = 4 * mask_pred_i
ML_B[i, y0:y1,
x0:x1] = semantic_logits[0, label_i + self.num_stuff,
y0:y1, x0:x1]
cat.append(label_i.item() + self.num_stuff)
i = i + 1
ML_A = ML_A[:i]
ML_B = ML_B[:i]
FL = (ML_A.sigmoid() + ML_B.sigmoid()) * (ML_A + ML_B)
intermediate_logits = torch.cat(
[semantic_logits[0, :self.num_stuff], FL], dim=0)
cat = torch.tensor(cat, dtype=torch.long)
intermediate_mask = torch.argmax(F.softmax(intermediate_logits, dim=0),
dim=0) + 1
intermediate_mask = intermediate_mask - self.num_stuff
intermediate_mask[intermediate_mask <= 0] = 0
unique = torch.unique(intermediate_mask)
ignore_val = intermediate_mask.max().item() + 1
ignore_arr = torch.ones(
(ignore_val, ), dtype=unique.dtype,
device=unique.device) * ignore_val
total_unique = unique.shape[0]
ignore_arr[unique] = torch.arange(total_unique).cuda(ignore_arr.device)
panoptic_mask = ignore_arr[intermediate_mask]
panoptic_mask[intermediate_mask == ignore_val] = 0
cat_ = cat[unique].long()
sem_pred[panoptic_mask > 0] = self.num_stuff
sem_pred[sem_pred >= self.num_stuff] = self.num_stuff
cls_stuff, area = torch.unique(sem_pred, return_counts=True)
cls_stuff[
area < self.test_cfg.panoptic.min_stuff_area] = self.num_stuff
cls_stuff = cls_stuff[cls_stuff != self.num_stuff]
tmp = torch.ones((self.num_stuff + 1, ),
dtype=cls_stuff.dtype,
device=cls_stuff.device) * self.num_stuff
tmp[cls_stuff] = torch.arange(cls_stuff.shape[0]).cuda(tmp.device)
new_sem_pred = tmp[sem_pred]
cat_ = torch.cat((cat_, cls_stuff.cpu().long()), -1)
bool_mask = new_sem_pred != self.num_stuff
panoptic_mask[bool_mask] = new_sem_pred[bool_mask] + total_unique
return panoptic_mask.cpu(), cat_.cpu()
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None,
epoch=None,
**kwargs):
"""
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_meta (list[dict]): list of image info dict where each dict has:
'img_shape', 'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
gt_bboxes (list[Tensor]): each item are the truth boxes for each
image in [tl_x, tl_y, br_x, br_y] format.
gt_labels (list[Tensor]): class indices corresponding to each box
gt_bboxes_ignore (None | list[Tensor]): specify which bounding
boxes can be ignored when computing the loss.
gt_masks (None | Tensor) : true segmentation masks for each box
used if the architecture supports a segmentation task.
proposals : override rpn proposals with custom proposals. Use when
`with_rpn` is False.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
# x = self.extract_feat(img)
losses = dict()
# keep the original usage
if self.dfn_balance is None:
x = self.extract_feat(img)
else:
# count the number of defects in each image
defect_nums = [0] * img.shape[0]
for i, gt_label in enumerate(gt_labels):
defect_cnt = 0
for label in gt_label:
if self.dfn_balance.background_id != int(label):
defect_cnt += 1
defect_nums[i] = defect_cnt
# split the images into defect(1) image and normal(0) image
dfn_labels = [0 if i == 0 else 1 for i in defect_nums]
dfn_labels = torch.Tensor(dfn_labels).long().cuda()
# calculate dfn loss
x, dfn_loss = self.extract_defect_feat(img, targets=dfn_labels)
# balance different network loss
dfn_weight = self.get_dfn_weight(epoch)
loss_dfn = dfn_loss['loss']
loss_dfn = loss_dfn * dfn_weight
losses.update(loss_dfn=loss_dfn)
# delete the ignored annotations
if self.ignore_ids is not None:
for ignore_id in self.ignore_ids:
for i, gt_label in enumerate(gt_labels):
keep_ind = gt_label != ignore_id
gt_labels[i] = gt_labels[i][keep_ind]
gt_bboxes[i] = gt_bboxes[i][keep_ind]
keep_ind = [True] * len(gt_labels)
for i in range(len(gt_labels) - 1, -1, -1):
if gt_labels[i].shape[0] == 0:
img_meta.pop(i)
gt_bboxes.pop(i)
gt_labels.pop(i)
keep_ind[i] = False
img = img[keep_ind]
x = list(x)
for i in range(len(x)):
x[i] = x[i][keep_ind]
x = tuple(x)
# if have no any images, then set all loss to be 0 except dfn loss
if len(gt_labels) == 0 or len(
gt_bboxes) == 0 or img.shape[0] == 0:
loss_zeros = [
torch.Tensor([0.]).float().cuda()
for i in range(len(x))
]
loss_ones = [
torch.Tensor([1.]).float().cuda()
for i in range(len(x))
]
losses.update(loss_rpn_cls=loss_zeros,
loss_rpn_bbox=loss_zeros)
loss_zero = torch.Tensor([0.]).float().cuda()
loss_one = torch.Tensor([1.]).float().cuda()
losses['loss_cls'] = loss_zero
losses['loss_bbox'] = loss_zero
losses['acc'] = loss_one
return losses
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
if mask_feats.shape[0] > 0:
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(
sampling_results, gt_masks, self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def forward_train(self, # 核心双阶段检测器的流程
img,
img_metas,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
"""
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_metas (list[dict]): list of image info dict where each dict
has: 'img_shape', 'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
gt_bboxes (list[Tensor]): each item are the truth boxes for each
image in [tl_x, tl_y, br_x, br_y] format(left top point, right bottom point).
gt_labels (list[Tensor]): class indices corresponding to each box
gt_bboxes_ignore (None | list[Tensor]): specify which bounding
boxes can be ignored when computing the loss.
gt_masks (None | Tensor) : true segmentation masks for each box
used if the architecture supports a segmentation task.
proposals : override rpn proposals with custom proposals. Use when
`with_rpn` is False.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
# 前向 backbone 和 neck
x = self.extract_feat(img)
losses = dict()
# RPN forward and loss
if self.with_rpn:
# rpn_head在上面__init__函数里面build了,返回了rpn_head对象，
# 但是因为都继承了nn.Module(实现了__call__)，可以直接用实例名字调用里面的forward函数，从而进行了前向传播
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_metas,
self.train_cfg.rpn)
# 这里loss是rpn_head里面的实现了，因为rpn_head继承了anchor_head，
# 所以用了父类anchor_head实现的loss，其实就是anchor的那一套，返回是一个字典
rpn_losses = self.rpn_head.loss(
*rpn_loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_metas, proposal_cfg)
# 得到proposal,把ａｎｃｈｏｒ转化为对应的框的信息，然后ＮＭＳ再取top-N个候选框
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
'''
============
proposal_list
2
torch.Size([2000, 5])
******
torch.Size([2000, 5])
============
============
'''
# assign gts and sample proposals
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(
self.train_cfg.rcnn.sampler, context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
# bbox_targets： 长度为4的tuple,元素是tensor
'''
bbox_targets:长度为4
四个元素分别的维度
labels： torch.Size([1024])
label_weights： torch.Size([1024])
bbox_targets torch.Size([1024, 4])
bbox_weights torch.Size([1024, 4])
'''
# tensor维度为1维与tensor长度和cls_score,bbox_pred一致,值为0或1,用来表示预测是否命中真实框
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(
res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(
res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
if mask_feats.shape[0] > 0:
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(
sampling_results, gt_masks, self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_bboxes_ignore,
gt_labels,
gt_masks=None,
proposals=None):
losses = dict()
x = self.extract_feat(img)
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs)
losses.update(rpn_losses)
proposal_inputs = rpn_outs + (img_meta, self.test_cfg.rpn)
proposal_list = self.rpn_head.get_proposals(*proposal_inputs)
else:
proposal_list = proposals
if self.with_bbox:
(pos_proposals, neg_proposals, pos_assigned_gt_inds, pos_gt_bboxes,
pos_gt_labels) = multi_apply(
sample_bboxes,
proposal_list,
gt_bboxes,
gt_bboxes_ignore,
gt_labels,
cfg=self.train_cfg.rcnn)
(labels, label_weights, bbox_targets,
bbox_weights) = self.bbox_head.get_bbox_target(
pos_proposals, neg_proposals, pos_gt_bboxes, pos_gt_labels,
self.train_cfg.rcnn)
rois = bbox2roi([
torch.cat([pos, neg], dim=0)
for pos, neg in zip(pos_proposals, neg_proposals)
])
# TODO: a more flexible way to configurate feat maps
roi_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
cls_score, bbox_pred = self.bbox_head(roi_feats)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred, labels,
label_weights, bbox_targets,
bbox_weights)
losses.update(loss_bbox)
if self.with_mask:
mask_targets = self.mask_head.get_mask_target(
pos_proposals, pos_assigned_gt_inds, gt_masks,
self.train_cfg.rcnn)
pos_rois = bbox2roi(pos_proposals)
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
mask_pred = self.mask_head(mask_feats)
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
torch.cat(pos_gt_labels))
losses.update(loss_mask)
return losses
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
losses = dict()
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(
*rpn_loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_inputs = rpn_outs + (img_meta, self.test_cfg.rpn)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
for i in range(self.num_stages):
rcnn_train_cfg = self.train_cfg.rcnn[i]
lw = self.train_cfg.stage_loss_weights[i]
# assign gts and sample proposals
assign_results, sampling_results = multi_apply(
assign_and_sample,
proposal_list,
gt_bboxes,
gt_bboxes_ignore,
gt_labels,
cfg=rcnn_train_cfg)
# bbox head forward and loss
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
rois = bbox2roi([res.bboxes for res in sampling_results])
bbox_feats = bbox_roi_extractor(x[:bbox_roi_extractor.num_inputs],
rois)
cls_score, bbox_pred = bbox_head(bbox_feats)
bbox_targets = bbox_head.get_target(sampling_results, gt_bboxes,
gt_labels, rcnn_train_cfg)
loss_bbox = bbox_head.loss(cls_score, bbox_pred, *bbox_targets)
for name, value in loss_bbox.items():
losses['s{}.{}'.format(i, name)] = (value * lw if
'loss' in name else value)
# mask head forward and loss
if self.with_mask:
mask_roi_extractor = self.mask_roi_extractor[i]
mask_head = self.mask_head[i]
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = mask_roi_extractor(
x[:mask_roi_extractor.num_inputs], pos_rois)
mask_pred = mask_head(mask_feats)
mask_targets = mask_head.get_target(sampling_results, gt_masks,
rcnn_train_cfg)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = mask_head.loss(mask_pred, mask_targets, pos_labels)
for name, value in loss_mask.items():
losses['s{}.{}'.format(i, name)] = (value * lw
if 'loss' in name else
value)
# refine bboxes
if i < self.num_stages - 1:
pos_is_gts = [res.pos_is_gt for res in sampling_results]
roi_labels = bbox_targets[0] # bbox_targets is a tuple
with torch.no_grad():
proposal_list = bbox_head.refine_bboxes(
rois, roi_labels, bbox_pred, pos_is_gts, img_meta)
return losses
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
losses = dict()
# trans gt_masks to gt_obbs
gt_obbs = gt_mask_bp_obbs_list(gt_masks)
gt_obbs = [choose_best_Rroi_batch(gt) for gt in gt_obbs]
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, gt_labels, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals (hbb assign)
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn[0].assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn[0].sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
## rbbox
gt_obbs_torch = [
torch.from_numpy(gt).float().to(rois.device) for gt in gt_obbs
]
rbbox_targets = self.bbox_head.get_target(sampling_results,
gt_obbs_torch, gt_labels,
self.train_cfg.rcnn[0])
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*rbbox_targets)
# losses.update(loss_bbox)
for name, value in loss_bbox.items():
losses['s{}.{}'.format(0, name)] = (value)
return losses
def aug_test(self, imgs, img_metas, proposals=None, rescale=None):
# raise NotImplementedError
# import pdb; pdb.set_trace()
proposal_list = self.aug_test_rpn_rotate(
self.extract_feats(imgs), img_metas, self.test_cfg.rpn)
rcnn_test_cfg = self.test_cfg.rcnn
aug_rbboxes = []
aug_rscores = []
for x, img_meta in zip(self.extract_feats(imgs), img_metas):
# only one image in the batch
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
scale_factor, flip)
angle = img_meta[0]['angle']
# print('img shape: ', img_shape)
if angle != 0:
try:
proposals = bbox_rotate_mapping(proposal_list[0][:, :4], img_shape,
angle)
except:
import pdb; pdb.set_trace()
rois = bbox2roi([proposals])
# recompute feature maps to save GPU memory
roi_feats = self.bbox_roi_extractor(
x[:len(self.bbox_roi_extractor.featmap_strides)], rois)
if self.with_shared_head:
roi_feats = self.shared_head(roi_feats)
cls_score, bbox_pred = self.bbox_head(roi_feats)
bbox_label = cls_score.argmax(dim=1)
rrois = self.bbox_head.regress_by_class_rbbox(roi2droi(rois), bbox_label,
bbox_pred,
img_meta[0])
rrois_enlarge = copy.deepcopy(rrois)
rrois_enlarge[:, 3] = rrois_enlarge[:, 3] * self.rbbox_roi_extractor.w_enlarge
rrois_enlarge[:, 4] = rrois_enlarge[:, 4] * self.rbbox_roi_extractor.h_enlarge
rbbox_feats = self.rbbox_roi_extractor(
x[:len(self.rbbox_roi_extractor.featmap_strides)], rrois_enlarge)
if self.with_shared_head_rbbox:
rbbox_feats = self.shared_head_rbbox(rbbox_feats)
rcls_score, rbbox_pred = self.rbbox_head(rbbox_feats)
rbboxes, rscores = self.rbbox_head.get_det_rbboxes(
rrois,
rcls_score,
rbbox_pred,
img_shape,
scale_factor,
rescale=rescale,
cfg=None)
aug_rbboxes.append(rbboxes)
aug_rscores.append(rscores)
merged_rbboxes, merged_rscores = merge_rotate_aug_bboxes(
aug_rbboxes, aug_rscores, img_metas, rcnn_test_cfg
)
det_rbboxes, det_rlabels = multiclass_nms_rbbox(
merged_rbboxes, merged_rscores, rcnn_test_cfg.score_thr,
rcnn_test_cfg.nms, rcnn_test_cfg.max_per_img)
if rescale:
_det_rbboxes = det_rbboxes
else:
_det_rbboxes = det_rbboxes.clone()
_det_rbboxes[:, :4] *= img_metas[0][0]['scale_factor']
rbbox_results = dbbox2result(_det_rbboxes, det_rlabels,
self.rbbox_head.num_classes)
return rbbox_results
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_polygons,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
if self.with_attention:
x = self.add_attention(x)
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_inputs = rpn_outs + (img_meta, self.test_cfg.rpn)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
if self.with_deform_adjust:
x = self.deform_adjust(x, rpn_outs[1])
# assign gts and sample proposals
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_polygons[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_polygons,
gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# mask head forward and loss
if self.with_mask:
pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(sampling_results,
gt_masks,
self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def simple_test(self, x, proposal_list, img_metas, rescale=False):
if self.with_semantic:
_, semantic_feat = self.semantic_head(x)
else:
semantic_feat = None
img_shape = img_metas[0]['img_shape']
ori_shape = img_metas[0]['ori_shape']
scale_factor = img_metas[0]['scale_factor']
# "ms" in variable names means multi-stage
ms_bbox_result = {}
ms_segm_result = {}
ms_scores = []
rcnn_test_cfg = self.test_cfg
rois = bbox2roi(proposal_list)
for i in range(self.num_stages):
bbox_head = self.bbox_head[i]
bbox_results = self._bbox_forward(i,
x,
rois,
semantic_feat=semantic_feat)
ms_scores.append(bbox_results['cls_score'])
if i < self.num_stages - 1:
bbox_label = bbox_results['cls_score'].argmax(dim=1)
rois = bbox_head.regress_by_class(rois, bbox_label,
bbox_results['bbox_pred'],
img_metas[0])
cls_score = sum(ms_scores) / float(len(ms_scores))
det_bboxes, det_labels = self.bbox_head[-1].get_bboxes(
rois,
cls_score,
bbox_results['bbox_pred'],
img_shape,
scale_factor,
rescale=rescale,
cfg=rcnn_test_cfg)
bbox_result = bbox2result(det_bboxes, det_labels,
self.bbox_head[-1].num_classes)
ms_bbox_result['ensemble'] = bbox_result
if self.with_mask:
if det_bboxes.shape[0] == 0:
mask_classes = self.mask_head[-1].num_classes
segm_result = [[] for _ in range(mask_classes)]
else:
_bboxes = (det_bboxes[:, :4] *
det_bboxes.new_tensor(scale_factor)
if rescale else det_bboxes)
mask_rois = bbox2roi([_bboxes])
aug_masks = []
mask_roi_extractor = self.mask_roi_extractor[-1]
mask_feats = mask_roi_extractor(
x[:len(mask_roi_extractor.featmap_strides)], mask_rois)
if self.with_semantic and 'mask' in self.semantic_fusion:
mask_semantic_feat = self.semantic_roi_extractor(
[semantic_feat], mask_rois)
mask_feats += mask_semantic_feat
last_feat = None
for i in range(self.num_stages):
mask_head = self.mask_head[i]
if self.mask_info_flow:
mask_pred, last_feat = mask_head(mask_feats, last_feat)
else:
mask_pred = mask_head(mask_feats)
aug_masks.append(mask_pred.sigmoid().cpu().numpy())
merged_masks = merge_aug_masks(aug_masks,
[img_metas] * self.num_stages,
self.test_cfg)
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks, _bboxes, det_labels, rcnn_test_cfg,
ori_shape, scale_factor, rescale)
ms_segm_result['ensemble'] = segm_result
if self.with_mask:
results = (ms_bbox_result['ensemble'], ms_segm_result['ensemble'])
else:
results = ms_bbox_result['ensemble']
return results
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
"""
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_meta (list[dict]): list of image info dict where each dict has:
'img_shape', 'scale_factor', 'flip', and my also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
gt_bboxes (list[Tensor]): each item are the truth boxes for each
image in [tl_x, tl_y, br_x, br_y] format.
gt_labels (list[Tensor]): class indices corresponding to each box
gt_bboxes_ignore (None | list[Tensor]): specify which bounding
boxes can be ignored when computing the loss.
gt_masks (None | Tensor) : true segmentation masks for each box
used if the architecture supports a segmentation task.
proposals : override rpn proposals with custom proposals. Use when
`with_rpn` is False.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
x = self.extract_feat(img)
# pdb.set_trace()
losses = dict()
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
for i in range(self.num_stages):
self.current_stage = i
# pdb.set_trace()
rcnn_train_cfg = self.train_cfg.rcnn[i]
lw = self.train_cfg.stage_loss_weights[i]
# assign gts and sample proposals
sampling_results = []
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(rcnn_train_cfg.assigner)
bbox_sampler = build_sampler(rcnn_train_cfg.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
for j in range(num_imgs):
assign_result = bbox_assigner.assign(
proposal_list[j], gt_bboxes[j], gt_bboxes_ignore[j],
gt_labels[j])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[j],
gt_bboxes[j],
gt_labels[j],
feats=[lvl_feat[j][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
rois = bbox2roi([res.bboxes for res in sampling_results])
if len(rois) == 0:
# If there are no predicted and/or truth boxes, then we cannot
# compute head / mask losses
continue
bbox_feats = bbox_roi_extractor(x[:bbox_roi_extractor.num_inputs],
rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
bbox_targets = bbox_head.get_target(sampling_results, gt_bboxes,
gt_labels, rcnn_train_cfg,
img_meta)
loss_bbox = bbox_head.loss(cls_score, bbox_pred, *bbox_targets)
for name, value in loss_bbox.items():
losses['s{}.{}'.format(
i, name)] = (value * lw if 'loss' in name else value)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
mask_roi_extractor = self.mask_roi_extractor[i]
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = mask_roi_extractor(
x[:mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
# reuse positive bbox feats
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
mask_head = self.mask_head[i]
mask_pred = mask_head(mask_feats)
mask_targets = mask_head.get_target(sampling_results, gt_masks,
rcnn_train_cfg)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = mask_head.loss(mask_pred, mask_targets, pos_labels)
for name, value in loss_mask.items():
losses['s{}.{}'.format(
i, name)] = (value * lw if 'loss' in name else value)
# refine bboxes
if i < self.num_stages - 1:
pos_is_gts = [res.pos_is_gt for res in sampling_results]
roi_labels = bbox_targets[0] # bbox_targets is a tuple
with torch.no_grad():
proposal_list = bbox_head.refine_bboxes(
rois, roi_labels, bbox_pred, pos_is_gts, img_meta)
return losses
def aug_test(self, img_feats, img_metas, proposals=None, rescale=False):
"""Test with augmentations.
If rescale is False, then returned bboxes and masks will fit the scale
of imgs[0].
"""
if self.with_semantic:
semantic_feats = [
self.semantic_head(feat)[1] for feat in img_feats
]
else:
semantic_feats = [None] * len(img_metas)
# recompute feats to save memory
proposal_list = self.aug_test_rpn(img_feats, img_metas,
self.test_cfg.rpn)
rcnn_test_cfg = self.test_cfg
aug_bboxes = []
aug_scores = []
for x, img_meta, semantic in zip(img_feats, img_metas, semantic_feats):
# only one image in the batch
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
flip_direction = img_meta[0]['flip_direction']
proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
scale_factor, flip, flip_direction)
# "ms" in variable names means multi-stage
ms_scores = []
rois = bbox2roi([proposals])
for i in range(self.num_stages):
bbox_head = self.bbox_head[i]
bbox_results = self._bbox_forward(i,
x,
rois,
semantic_feat=semantic)
ms_scores.append(bbox_results['cls_score'])
if i < self.num_stages - 1:
bbox_label = bbox_results['cls_score'].argmax(dim=1)
rois = bbox_head.regress_by_class(
rois, bbox_label, bbox_results['bbox_pred'],
img_meta[0])
cls_score = sum(ms_scores) / float(len(ms_scores))
bboxes, scores = self.bbox_head[-1].get_bboxes(
rois,
cls_score,
bbox_results['bbox_pred'],
img_shape,
scale_factor,
rescale=False,
cfg=None)
aug_bboxes.append(bboxes)
aug_scores.append(scores)
# after merging, bboxes will be rescaled to the original image size
merged_bboxes, merged_scores = merge_aug_bboxes(
aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
det_bboxes, det_labels = multiclass_nms(merged_bboxes, merged_scores,
rcnn_test_cfg.score_thr,
rcnn_test_cfg.nms,
rcnn_test_cfg.max_per_img)
bbox_result = bbox2result(det_bboxes, det_labels,
self.bbox_head[-1].num_classes)
if self.with_mask:
if det_bboxes.shape[0] == 0:
segm_result = [[]
for _ in range(self.mask_head[-1].num_classes -
1)]
else:
aug_masks = []
aug_img_metas = []
for x, img_meta, semantic in zip(img_feats, img_metas,
semantic_feats):
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
flip_direction = img_meta[0]['flip_direction']
_bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
scale_factor, flip, flip_direction)
mask_rois = bbox2roi([_bboxes])
mask_feats = self.mask_roi_extractor[-1](
x[:len(self.mask_roi_extractor[-1].featmap_strides)],
mask_rois)
if self.with_semantic:
semantic_feat = semantic
mask_semantic_feat = self.semantic_roi_extractor(
[semantic_feat], mask_rois)
if mask_semantic_feat.shape[-2:] != mask_feats.shape[
-2:]:
mask_semantic_feat = F.adaptive_avg_pool2d(
mask_semantic_feat, mask_feats.shape[-2:])
mask_feats += mask_semantic_feat
last_feat = None
for i in range(self.num_stages):
mask_head = self.mask_head[i]
if self.mask_info_flow:
mask_pred, last_feat = mask_head(
mask_feats, last_feat)
else:
mask_pred = mask_head(mask_feats)
aug_masks.append(mask_pred.sigmoid().cpu().numpy())
aug_img_metas.append(img_meta)
merged_masks = merge_aug_masks(aug_masks, aug_img_metas,
self.test_cfg)
ori_shape = img_metas[0][0]['ori_shape']
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks,
det_bboxes,
det_labels,
rcnn_test_cfg,
ori_shape,
scale_factor=1.0,
rescale=False)
return bbox_result, segm_result
else:
return bbox_result
def aug_test(self, imgs, img_metas, proposals=None, rescale=False):
"""Test with augmentations.
If rescale is False, then returned bboxes and masks will fit the scale
of imgs[0].
"""
# recompute feats to save memory
base_feats = list(self.extract_feats(imgs))
proposal_list = self.aug_test_rpn(base_feats, img_metas,
self.test_cfg.rpn)
rcnn_test_cfg = self.test_cfg.rcnn
aug_bboxes = []
aug_scores = []
aug_variance = []
for x, img_meta in zip(base_feats, img_metas):
# only one image in the batch
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
flip_direction = img_meta[0].get('flip_direction', 'horizontal')
proposals = bbox_mapping(proposal_list[0][:, :4], img_shape,
scale_factor, flip, flip_direction)
# "ms" in variable names means multi-stage
ms_scores = []
rois = bbox2roi([proposals])
for i in range(self.num_stages):
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
bbox_feats = bbox_roi_extractor(
x[:len(bbox_roi_extractor.featmap_strides)], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
ms_scores.append(cls_score)
if i < self.num_stages - 1:
bbox_label = cls_score.argmax(dim=1)
rois = bbox_head.regress_by_class(rois, bbox_label,
bbox_pred, img_meta[0])
cls_score = sum(ms_scores) / float(len(ms_scores))
bboxes, scores = self.bbox_head[-1].get_det_bboxes(rois,
cls_score,
bbox_pred,
img_shape,
scale_factor,
rescale=False,
cfg=None)
# aug_bboxes.append(bboxes) # notice: bboxes may be tuple
if isinstance(bboxes, tuple):
aug_bboxes.append(bboxes[0])
aug_variance.append(bboxes[1])
else:
aug_bboxes.append(bboxes)
aug_scores.append(scores)
return_variance = rcnn_test_cfg.get('return_variance', False)
if len(aug_variance
) == 0 or rcnn_test_cfg.nms.type != 'soft_nms_variance_voting':
merged_bboxes, merged_scores = merge_aug_bboxes(
aug_bboxes, aug_scores, img_metas, rcnn_test_cfg)
det_bboxes, det_labels = multiclass_nms(merged_bboxes,
merged_scores,
rcnn_test_cfg.score_thr,
rcnn_test_cfg.nms,
rcnn_test_cfg.max_per_img)
else:
merged_bboxes, merged_scores, merged_variance = merge_aug_bboxes_variance(
aug_bboxes, aug_scores, aug_variance, img_metas, rcnn_test_cfg)
det_bboxes, det_labels = multiclass_soft_nms_variance_voting(
merged_bboxes,
merged_variance,
merged_scores,
rcnn_test_cfg.score_thr,
rcnn_test_cfg.nms,
rcnn_test_cfg.max_per_img,
return_variance=return_variance)
bbox_result = bbox2result(det_bboxes,
det_labels,
self.bbox_head[-1].num_classes,
with_variance=return_variance)
if self.with_mask:
if det_bboxes.shape[0] == 0:
segm_result = [[]
for _ in range(self.mask_head[-1].num_classes -
1)]
else:
aug_masks = []
aug_img_metas = []
for x, img_meta in zip(self.extract_feats(imgs), img_metas):
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
flip = img_meta[0]['flip']
flip_direction = img_meta[0].get('flip_direction',
'horizontal')
_bboxes = bbox_mapping(det_bboxes[:, :4], img_shape,
scale_factor, flip, flip_direction)
mask_rois = bbox2roi([_bboxes])
for i in range(self.num_stages):
mask_feats = self.mask_roi_extractor[i](
x[:len(self.mask_roi_extractor[i].featmap_strides
)], mask_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
mask_pred = self.mask_head[i](mask_feats)
aug_masks.append(mask_pred.sigmoid().cpu().numpy())
aug_img_metas.append(img_meta)
merged_masks = merge_aug_masks(aug_masks, aug_img_metas,
self.test_cfg.rcnn)
ori_shape = img_metas[0][0]['ori_shape']
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks,
det_bboxes,
det_labels,
rcnn_test_cfg,
ori_shape,
scale_factor=1.0,
rescale=False)
return bbox_result, segm_result
else:
return bbox_result
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
"""
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_meta (list[dict]): list of image info dict where each dict has:
'img_shape', 'scale_factor', 'flip', and my also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
gt_bboxes (list[Tensor]): each item are the truth boxes for each
image in [tl_x, tl_y, br_x, br_y] format.
gt_labels (list[Tensor]): class indices corresponding to each box
gt_bboxes_ignore (None | list[Tensor]): specify which bounding
boxes can be ignored when computing the loss.
gt_masks (None | Tensor) : true segmentation masks for each box
used if the architecture supports a segmentation task.
proposals : override rpn proposals with custom proposals. Use when
`with_rpn` is False.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
x = self.extract_feat(img) # x得到的是一个元组，对于resnet，可包含多个conv层的结果
conv4 = (x[0], )
conv5 = x[1]
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(conv4) # rpn使用的是resnet的conv4
# rpn_outs = rpn_cls_score + rpn_bbox_pred
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
large_seperate_conv_output = (self.large_seperate_conv(conv5), )
# assign gts and sample proposals
if self.with_bbox:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# rois是基于特征图尺寸的，不是基于原图尺寸的
# roi:(roi_batch_inds,x1,y1,x2,y2)
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
large_seperate_conv_output[:self.bbox_roi_extractor.
num_inputs], rois)
# bbox_feats是基于特征图尺寸的，如果需要基于原图尺寸，需要传入scale_factor参数
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
return losses
def simple_test(self, img, img_meta, proposals=None, rescale=False):
x = self.extract_feat(img)
proposal_list = self.simple_test_rpn(
x, img_meta, self.test_cfg.rpn) if proposals is None else proposals
img_shape = img_meta[0]['img_shape']
ori_shape = img_meta[0]['ori_shape']
scale_factor = img_meta[0]['scale_factor']
# "ms" in variable names means multi-stage
ms_bbox_result = {}
ms_segm_result = {}
ms_scores = []
rcnn_test_cfg = self.test_cfg.rcnn
rois = bbox2roi(proposal_list)
for i in range(self.num_stages):
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
bbox_feats = bbox_roi_extractor(
x[:len(bbox_roi_extractor.featmap_strides)], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
ms_scores.append(cls_score)
if self.test_cfg.keep_all_stages:
det_bboxes, det_labels = bbox_head.get_det_bboxes(
rois,
cls_score,
bbox_pred,
img_shape,
scale_factor,
rescale=rescale,
cfg=rcnn_test_cfg)
bbox_result = bbox2result(det_bboxes, det_labels,
bbox_head.num_classes)
ms_bbox_result['stage{}'.format(i)] = bbox_result
if self.with_mask:
mask_roi_extractor = self.mask_roi_extractor[i]
mask_head = self.mask_head[i]
if det_bboxes.shape[0] == 0:
mask_classes = mask_head.num_classes - 1
segm_result = [[] for _ in range(mask_classes)]
else:
_bboxes = (det_bboxes[:, :4] *
scale_factor if rescale else det_bboxes)
mask_rois = bbox2roi([_bboxes])
mask_feats = mask_roi_extractor(
x[:len(mask_roi_extractor.featmap_strides)],
mask_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats, i)
mask_pred = mask_head(mask_feats)
segm_result = mask_head.get_seg_masks(
mask_pred, _bboxes, det_labels, rcnn_test_cfg,
ori_shape, scale_factor, rescale)
ms_segm_result['stage{}'.format(i)] = segm_result
if i < self.num_stages - 1:
bbox_label = cls_score.argmax(dim=1)
rois = bbox_head.regress_by_class(rois, bbox_label, bbox_pred,
img_meta[0])
cls_score = sum(ms_scores) / self.num_stages
# if not self.with_IoU:
iou_pred = None
det_bboxes, det_labels = self.bbox_head[-1].get_det_bboxes(
rois,
cls_score,
bbox_pred,
iou_pred,
img_shape,
scale_factor,
rescale=rescale,
cfg=rcnn_test_cfg)
bbox_result = bbox2result(det_bboxes, det_labels,
self.bbox_head[-1].num_classes)
ms_bbox_result['ensemble'] = bbox_result
if self.with_mask:
if det_bboxes.shape[0] == 0:
mask_classes = self.mask_head[-1].num_classes - 1
segm_result = [[] for _ in range(mask_classes)]
else:
if isinstance(scale_factor, float): # aspect ratio fixed
_bboxes = (det_bboxes[:, :4] *
scale_factor if rescale else det_bboxes)
else:
_bboxes = (
det_bboxes[:, :4] *
torch.from_numpy(scale_factor).to(det_bboxes.device)
if rescale else det_bboxes)
mask_rois = bbox2roi([_bboxes])
aug_masks = []
for i in range(self.num_stages):
mask_roi_extractor = self.mask_roi_extractor[i]
mask_feats = mask_roi_extractor(
x[:len(mask_roi_extractor.featmap_strides)], mask_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
mask_pred = self.mask_head[i](mask_feats)
aug_masks.append(mask_pred.sigmoid().cpu().numpy())
merged_masks = merge_aug_masks(aug_masks,
[img_meta] * self.num_stages,
self.test_cfg.rcnn)
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks, _bboxes, det_labels, rcnn_test_cfg,
ori_shape, scale_factor, rescale)
ms_segm_result['ensemble'] = segm_result
if not self.test_cfg.keep_all_stages:
if self.with_mask:
results = (ms_bbox_result['ensemble'],
ms_segm_result['ensemble'])
else:
results = ms_bbox_result['ensemble']
else:
if self.with_mask:
results = {
stage: (ms_bbox_result[stage], ms_segm_result[stage])
for stage in ms_bbox_result
}
else:
results = ms_bbox_result
return results
def simple_test_seq(self,
x,
img_metas,
det_bboxes,
det_labels,
rescale=False):
"""Simple test for seq head without augmentation."""
# image shapes of images in the batch
ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
num_imgs = len(det_bboxes)
if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
rec_results = [[[] for _ in range(self.seq_head.num_classes)]
for _ in range(num_imgs)]
else:
# if det_bboxes is rescaled to the original image size, we need to
# rescale it back to the testing scale to obtain RoIs.
if rescale and not isinstance(scale_factors[0], float):
scale_factors = [
torch.from_numpy(scale_factor).to(det_bboxes[0].device)
for scale_factor in scale_factors
]
if torch.onnx.is_in_onnx_export():
# avoid seq_pred.split with static number of prediction
seq_preds = []
_bboxes = []
for i, boxes in enumerate(det_bboxes):
boxes = boxes[:, :4]
if rescale:
boxes *= scale_factors[i]
_bboxes.append(boxes)
img_inds = boxes[:, :1].clone() * 0 + i
seq_rois = torch.cat([img_inds, boxes], dim=-1)
seq_result = self._seq_forward(x, seq_rois)
seq_preds.append(seq_result['seq_pred'])
else:
_bboxes = [
det_bboxes[i][:, :4] *
scale_factors[i] if rescale else det_bboxes[i][:, :4]
for i in range(len(det_bboxes))
]
seq_rois = bbox2roi(_bboxes)
seq_results = self._seq_forward(x, seq_rois)
seq_pred = seq_results['seq_pred']
# split batch seq prediction back to each image
num_seq_roi_per_img = [
det_bbox.shape[0] for det_bbox in det_bboxes
]
seq_preds = seq_pred.split(num_seq_roi_per_img, 0)
# apply seq post-processing to each image individually
rec_results = []
for i in range(num_imgs):
if det_bboxes[i].shape[0] == 0:
rec_results.append(
[[] for _ in range(self.seq_head.num_classes)])
else:
rec_result = self.seq_head.get_rec_seqs(
seq_preds[i], _bboxes[i], det_labels[i],
self.test_cfg, ori_shapes[i], scale_factors[i],
rescale)
rec_results.append(rec_result)
return rec_results
def simple_test_bboxes(self,
x,
img_meta,
proposals,
rcnn_test_cfg,
rescale=False):
"""Test only det bboxes without augmentation."""
rois = bbox2roi(proposals)
roi_feats = self.bbox_roi_extractor(
x[:len(self.bbox_roi_extractor.featmap_strides)], rois)
cls_score, bbox_pred = self.bbox_head(roi_feats)
img_shape = img_meta[0]['img_shape']
scale_factor = img_meta[0]['scale_factor']
is_first = img_meta[0]['is_first']
det_bboxes, det_labels = self.bbox_head.get_det_bboxes(
rois,
cls_score,
bbox_pred,
img_shape,
scale_factor,
rescale=rescale,
cfg=rcnn_test_cfg)
if det_bboxes.nelement()==0:
det_obj_ids=np.array([], dtype=np.int64)
if is_first:
self.prev_bboxes = None
self.prev_roi_feats = None
self.prev_det_labels = None
return det_bboxes, det_labels, det_obj_ids
res_det_bboxes = det_bboxes.clone()
if rescale:
res_det_bboxes[:, :4] *= scale_factor
det_rois = bbox2roi([res_det_bboxes])
det_roi_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], det_rois)
# recompute bbox match feature
if is_first or (not is_first and self.prev_bboxes is None):
det_obj_ids = np.arange(det_bboxes.size(0))
# save bbox and features for later matching
self.prev_bboxes = det_bboxes
self.prev_roi_feats = det_roi_feats
self.prev_det_labels = det_labels
else:
assert self.prev_roi_feats is not None
# only support one image at a time
bbox_img_n = [det_bboxes.size(0)]
prev_bbox_img_n = [self.prev_roi_feats.size(0)]
match_score = self.track_head(det_roi_feats, self.prev_roi_feats,
bbox_img_n, prev_bbox_img_n)[0]
match_logprob = torch.nn.functional.log_softmax(match_score, dim=1)
label_delta = (self.prev_det_labels == det_labels.view(-1,1)).float()
bbox_ious = bbox_overlaps(det_bboxes[:,:4], self.prev_bboxes[:,:4])
# compute comprehensive score
comp_scores = self.track_head.compute_comp_scores(match_logprob,
det_bboxes[:,4].view(-1, 1),
bbox_ious,
label_delta,
add_bbox_dummy=True)
match_likelihood, match_ids = torch.max(comp_scores, dim =1)
# translate match_ids to det_obj_ids, assign new id to new objects
# update tracking features/bboxes of exisiting object,
# add tracking features/bboxes of new object
match_ids = match_ids.cpu().numpy().astype(np.int32)
det_obj_ids = np.ones((match_ids.shape[0]), dtype=np.int32) * (-1)
best_match_scores = np.ones((self.prev_bboxes.size(0))) * (-100)
for idx, match_id in enumerate(match_ids):
if match_id == 0:
# add new object
det_obj_ids[idx] = self.prev_roi_feats.size(0)
self.prev_roi_feats = torch.cat((self.prev_roi_feats, det_roi_feats[idx][None]), dim=0)
self.prev_bboxes = torch.cat((self.prev_bboxes, det_bboxes[idx][None]), dim=0)
self.prev_det_labels = torch.cat((self.prev_det_labels, det_labels[idx][None]), dim=0)
else:
# multiple candidate might match with previous object, here we choose the one with
# largest comprehensive score
obj_id = match_id - 1
match_score = comp_scores[idx, match_id]
if match_score > best_match_scores[obj_id]:
det_obj_ids[idx] = obj_id
best_match_scores[obj_id] = match_score
# udpate feature
self.prev_roi_feats[obj_id] = det_roi_feats[idx]
self.prev_bboxes[obj_id] = det_bboxes[idx]
return det_bboxes, det_labels, det_obj_ids
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
losses = dict()
# trans gt_masks to gt_obbs
# gt_obbs = gt_mask_bp_obbs_list(gt_masks)
# gt_obbs = gt_mask_bp_obbs_list(mask2obb_mask(gt_masks))
gt_obbs = gt_mask_bp_obbs_list(
tran2mix_mask(gt_bboxes, gt_masks, gt_labels))
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals (hbb assign)
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn[0].assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn[0].sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
## rbbox
# rbbox_targets = self.bbox_head.get_target(
# sampling_results, gt_masks, gt_labels, self.train_cfg.rcnn[0])
# rbbox_targets = self.bbox_head.get_target(
# sampling_results, mask2obb_mask(gt_masks), gt_labels, self.train_cfg.rcnn[0])
rbbox_targets = self.bbox_head.get_target(
sampling_results, tran2mix_mask(gt_bboxes, gt_masks,
gt_labels), gt_labels,
self.train_cfg.rcnn[0])
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*rbbox_targets)
# losses.update(loss_bbox)
for name, value in loss_bbox.items():
losses['s{}.{}'.format(0, name)] = (value)
pos_is_gts = [res.pos_is_gt for res in sampling_results]
roi_labels = rbbox_targets[0]
with torch.no_grad():
# import pdb
# pdb.set_trace()
rotated_proposal_list = self.bbox_head.refine_rbboxes(
roi2droi(rois), roi_labels, bbox_pred, pos_is_gts, img_meta)
# import pdb
# pdb.set_trace()
# assign gts and sample proposals (rbb assign)
if self.with_rbbox:
bbox_assigner = build_assigner(self.train_cfg.rcnn[1].assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn[1].sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
gt_obbs_best_roi = choose_best_Rroi_batch(gt_obbs[i])
assign_result = bbox_assigner.assign(rotated_proposal_list[i],
gt_obbs_best_roi,
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
rotated_proposal_list[i],
torch.from_numpy(gt_obbs_best_roi).float().to(
rotated_proposal_list[i].device),
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
if self.with_rbbox:
# (batch_ind, x_ctr, y_ctr, w, h, angle)
rrois = dbbox2roi([res.bboxes for res in sampling_results])
# feat enlarge
# rrois[:, 3] = rrois[:, 3] * 1.2
# rrois[:, 4] = rrois[:, 4] * 1.4
rrois[:, 3] = rrois[:, 3] * self.rbbox_roi_extractor.w_enlarge
rrois[:, 4] = rrois[:, 4] * self.rbbox_roi_extractor.h_enlarge
rbbox_feats = self.rbbox_roi_extractor(
x[:self.rbbox_roi_extractor.num_inputs], rrois)
if self.with_shared_head_rbbox:
rbbox_feats = self.shared_head_rbbox(rbbox_feats)
cls_score, rbbox_pred = self.rbbox_head(rbbox_feats)
rbbox_targets = self.rbbox_head.get_target_rbbox(
sampling_results, gt_obbs, gt_labels, self.train_cfg.rcnn[1])
loss_rbbox = self.rbbox_head.loss(cls_score, rbbox_pred,
*rbbox_targets)
for name, value in loss_rbbox.items():
losses['s{}.{}'.format(1, name)] = (value)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
pos_rrois = dbbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rrois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_rotate_adaptive_target(
sampling_results, gt_masks, self.train_cfg.rcnn[1],
self.mask_roi_extractor.ratio_max)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(sampling_results,
gt_masks,
self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def forward_train_pair(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
# img: b, 2*c = 6, h, w
b, c, h, w = img.shape
img = img.reshape(-1, c // 2, h, w)
# img: 2*b, c, h, w [0,1] , [2,3] , ......
x = self.extract_feat(img)
# x : tuple,5 layer
losses = dict()
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_outs_half = []
for outs_0 in rpn_outs:
tmp = []
for outs_1 in outs_0:
tmp.append(outs_1[::2, :, :, :])
rpn_outs_half.append(tmp)
rpn_loss_inputs = tuple(rpn_outs_half) + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(
*rpn_loss_inputs, gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
# copy train img_meta to pair. 1,2,3->1,1,2,2,3,3
img_meta = [img_meta[i // 2] for i in range(2 * len(img_meta))]
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# generate blank gt for normal img
gt_bboxes_ = []
gt_labels_ = []
for i in range(len(gt_bboxes)):
gt_bboxes_.append(gt_bboxes[i])
gt_bboxes_.append(torch.Tensor([[1, 1, 1, 1]]).to(gt_bboxes[i].device))
gt_labels_.append(gt_labels[i])
gt_labels_.append(torch.Tensor([[0]]).to(gt_labels[i].device))
for i in range(self.num_stages):
self.current_stage = i
rcnn_train_cfg = self.train_cfg.rcnn[i]
lw = self.train_cfg.stage_loss_weights[i]
# assign gts and sample proposals
sampling_results = []
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(rcnn_train_cfg.assigner)
bbox_sampler = build_sampler(
rcnn_train_cfg.sampler, context=self)
assert img.size(0) % 2 == 0
num_pairs = img.size(0) // 2
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for j in range(num_pairs):
i_train = 2 * j
i_normal = i_train + 1
assign_result_train = bbox_assigner.assign(proposal_list[i_train],
gt_bboxes_[i_train],
gt_bboxes_ignore[i_train],
gt_labels_[i_train])
assign_result_normal = bbox_assigner.assign(proposal_list[i_normal],
gt_bboxes_[i_normal],
gt_bboxes_ignore[i_normal],
gt_labels_[i_normal])
sampling_result_train, sampling_results_normal = bbox_sampler.pair_sample(
assign_result_train,
assign_result_normal,
proposal_list[i_train],
proposal_list[i_normal],
gt_bboxes_[i_train],
gt_labels_[i_train],
feats_train=[lvl_feat[i_train][None] for lvl_feat in x],
feats_normal=[lvl_feat[i_normal][None] for lvl_feat in x])
sampling_results.append(sampling_result_train)
sampling_results.append(sampling_results_normal)
# bbox head forward and loss
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
rois = bbox2roi([res.bboxes for res in sampling_results])
bbox_feats = bbox_roi_extractor(x[:bbox_roi_extractor.num_inputs],
rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
bbox_targets = bbox_head.get_target(sampling_results, gt_bboxes,
gt_labels, rcnn_train_cfg)
loss_bbox = bbox_head.loss(cls_score, bbox_pred, *bbox_targets)
for name, value in loss_bbox.items():
losses['s{}.{}'.format(i, name)] = (
value * lw if 'loss' in name else value)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
mask_roi_extractor = self.mask_roi_extractor[i]
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = mask_roi_extractor(
x[:mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
# reuse positive bbox feats
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(
res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(
res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
mask_head = self.mask_head[i]
mask_pred = mask_head(mask_feats)
mask_targets = mask_head.get_target(sampling_results, gt_masks,
rcnn_train_cfg)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = mask_head.loss(mask_pred, mask_targets, pos_labels)
for name, value in loss_mask.items():
losses['s{}.{}'.format(i, name)] = (
value * lw if 'loss' in name else value)
# refine bboxes
if i < self.num_stages - 1:
pos_is_gts = [res.pos_is_gt for res in sampling_results]
roi_labels = bbox_targets[0] # bbox_targets is a tuple
# print('stage:{}'.format(i))
# ic(proposal_list[1])
with torch.no_grad():
proposal_list = bbox_head.refine_bboxes(
rois, roi_labels, bbox_pred, pos_is_gts, img_meta)
# ic(proposal_list[1])
# exit(0)
return losses
def simple_test(self, x, proposal_list, img_metas, rescale=False):
"""Test without augmentation."""
assert self.with_bbox, 'Bbox head must be implemented.'
num_imgs = len(proposal_list)
img_shapes = tuple(meta['img_shape'] for meta in img_metas)
ori_shapes = tuple(meta['ori_shape'] for meta in img_metas)
scale_factors = tuple(meta['scale_factor'] for meta in img_metas)
# "ms" in variable names means multi-stage
ms_bbox_result = {}
ms_segm_result = {}
ms_scores = []
rcnn_test_cfg = self.test_cfg
rois = bbox2roi(proposal_list)
for i in range(self.num_stages):
bbox_results = self._bbox_forward(i, x, rois)
# split batch bbox prediction back to each image
cls_score = bbox_results['cls_score']
bbox_pred = bbox_results['bbox_pred']
num_proposals_per_img = tuple(
len(proposals) for proposals in proposal_list)
rois = rois.split(num_proposals_per_img, 0)
cls_score = cls_score.split(num_proposals_per_img, 0)
bbox_pred = bbox_pred.split(num_proposals_per_img, 0)
ms_scores.append(cls_score)
if i < self.num_stages - 1:
bbox_label = [s[:, :-1].argmax(dim=1) for s in cls_score]
rois = torch.cat([
self.bbox_head[i].regress_by_class(rois[j], bbox_label[j],
bbox_pred[j],
img_metas[j])
for j in range(num_imgs)
])
# average scores of each image by stages
cls_score = [
sum([score[i] for score in ms_scores]) / float(len(ms_scores))
for i in range(num_imgs)
]
# apply bbox post-processing to each image individually
det_bboxes = []
det_labels = []
for i in range(num_imgs):
det_bbox, det_label = self.bbox_head[-1].get_bboxes(
rois[i],
cls_score[i],
bbox_pred[i],
img_shapes[i],
scale_factors[i],
rescale=rescale,
cfg=rcnn_test_cfg)
det_bboxes.append(det_bbox)
det_labels.append(det_label)
bbox_results = [
bbox2result(det_bboxes[i], det_labels[i],
self.bbox_head[-1].num_classes)
for i in range(num_imgs)
]
ms_bbox_result['ensemble'] = bbox_results
if self.with_mask:
if all(det_bbox.shape[0] == 0 for det_bbox in det_bboxes):
mask_classes = self.mask_head[-1].num_classes
segm_results = [[[] for _ in range(mask_classes)]
for _ in range(num_imgs)]
else:
if rescale and not isinstance(scale_factors[0], float):
scale_factors = [
torch.from_numpy(scale_factor).to(det_bboxes[0].device)
for scale_factor in scale_factors
]
_bboxes = [
det_bboxes[i][:, :4] *
scale_factors[i] if rescale else det_bboxes[i][:, :4]
for i in range(len(det_bboxes))
]
mask_rois = bbox2roi(_bboxes)
num_mask_rois_per_img = tuple(
_bbox.size(0) for _bbox in _bboxes)
aug_masks = []
for i in range(self.num_stages):
mask_results = self._mask_forward(i, x, mask_rois)
mask_pred = mask_results['mask_pred']
# split batch mask prediction back to each image
mask_pred = mask_pred.split(num_mask_rois_per_img, 0)
aug_masks.append(
[m.sigmoid().cpu().numpy() for m in mask_pred])
# apply mask post-processing to each image individually
segm_results = []
for i in range(num_imgs):
if det_bboxes[i].shape[0] == 0:
segm_results.append(
[[]
for _ in range(self.mask_head[-1].num_classes)])
else:
aug_mask = [mask[i] for mask in aug_masks]
merged_masks = merge_aug_masks(
aug_mask, [[img_metas[i]]] * self.num_stages,
rcnn_test_cfg)
segm_result = self.mask_head[-1].get_seg_masks(
merged_masks, _bboxes[i], det_labels[i],
rcnn_test_cfg, ori_shapes[i], scale_factors[i],
rescale)
segm_results.append(segm_result)
ms_segm_result['ensemble'] = segm_results
if self.with_mask:
results = list(
zip(ms_bbox_result['ensemble'], ms_segm_result['ensemble']))
else:
results = ms_bbox_result['ensemble']
return results
def forward_train(self,
img,
img_metas,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
"""
Args:
img (Tensor): of shape (N, C, H, W) encoding input images.
Typically these should be mean centered and std scaled.
img_metas (list[dict]): list of image info dict where each dict
has: 'img_shape', 'scale_factor', 'flip', and may also contain
'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
For details on the values of these keys see
`mmdet/datasets/pipelines/formatting.py:Collect`.
gt_bboxes (list[Tensor]): each item are the truth boxes for each
image in [tl_x, tl_y, br_x, br_y] format.
gt_labels (list[Tensor]): class indices corresponding to each box
gt_bboxes_ignore (None | list[Tensor]): specify which bounding
boxes can be ignored when computing the loss.
gt_masks (None | Tensor) : true segmentation masks for each box
used if the architecture supports a segmentation task.
proposals : override rpn proposals with custom proposals. Use when
`with_rpn` is False.
Returns:
dict[str, Tensor]: a dictionary of loss components
"""
x = self.extract_feat(img)
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_metas,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_metas, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
# assign gts and sample proposals
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
if self.with_bbox:
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = self.mask_roi_extractor(
x[:self.mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
if mask_feats.shape[0] > 0:
mask_pred = self.mask_head(mask_feats)
mask_targets = self.mask_head.get_target(
sampling_results, gt_masks, self.train_cfg.rcnn)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = self.mask_head.loss(mask_pred, mask_targets,
pos_labels)
losses.update(loss_mask)
return losses
def forward_train(self,
img,
img_meta,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
losses = dict()
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_meta,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_meta, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
for i in range(self.num_stages):
self.current_stage = i
rcnn_train_cfg = self.train_cfg.rcnn[i]
lw = self.train_cfg.stage_loss_weights[i]
# assign gts and sample proposals
sampling_results = []
if self.with_bbox or self.with_mask:
bbox_assigner = build_assigner(rcnn_train_cfg.assigner)
bbox_sampler = build_sampler(rcnn_train_cfg.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
for j in range(num_imgs):
assign_result = bbox_assigner.assign(
proposal_list[j], gt_bboxes[j], gt_bboxes_ignore[j],
gt_labels[j])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[j],
gt_bboxes[j],
gt_labels[j],
feats=[lvl_feat[j][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
bbox_roi_extractor = self.bbox_roi_extractor[i]
bbox_head = self.bbox_head[i]
rois = bbox2roi([res.bboxes for res in sampling_results])
bbox_feats = bbox_roi_extractor(x[:bbox_roi_extractor.num_inputs],
rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = bbox_head(bbox_feats)
bbox_targets = bbox_head.get_target(sampling_results, gt_bboxes,
gt_labels, rcnn_train_cfg)
loss_bbox = bbox_head.loss(cls_score, bbox_pred, *bbox_targets)
for name, value in loss_bbox.items():
losses['s{}.{}'.format(
i, name)] = (value * lw if 'loss' in name else value)
# mask head forward and loss
if self.with_mask:
if not self.share_roi_extractor:
mask_roi_extractor = self.mask_roi_extractor[i]
pos_rois = bbox2roi(
[res.pos_bboxes for res in sampling_results])
mask_feats = mask_roi_extractor(
x[:mask_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
mask_feats = self.shared_head(mask_feats)
else:
# reuse positive bbox feats
pos_inds = []
device = bbox_feats.device
for res in sampling_results:
pos_inds.append(
torch.ones(res.pos_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds.append(
torch.zeros(res.neg_bboxes.shape[0],
device=device,
dtype=torch.uint8))
pos_inds = torch.cat(pos_inds)
mask_feats = bbox_feats[pos_inds]
mask_head = self.mask_head[i]
mask_pred = mask_head(mask_feats)
mask_targets = mask_head.get_target(sampling_results, gt_masks,
rcnn_train_cfg)
pos_labels = torch.cat(
[res.pos_gt_labels for res in sampling_results])
loss_mask = mask_head.loss(mask_pred, mask_targets, pos_labels)
for name, value in loss_mask.items():
losses['s{}.{}'.format(
i, name)] = (value * lw if 'loss' in name else value)
# refine bboxes
if i < self.num_stages - 1:
pos_is_gts = [res.pos_is_gt for res in sampling_results]
roi_labels = bbox_targets[0] # bbox_targets is a tuple
with torch.no_grad():
proposal_list = bbox_head.refine_bboxes(
rois, roi_labels, bbox_pred, pos_is_gts, img_meta)
return losses
def forward_train(self,
img,
img_metas,
gt_bboxes,
gt_labels,
gt_bboxes_ignore=None,
gt_masks=None,
proposals=None):
x = self.extract_feat(img)
losses = dict()
# RPN forward and loss
if self.with_rpn:
rpn_outs = self.rpn_head(x)
rpn_loss_inputs = rpn_outs + (gt_bboxes, img_metas,
self.train_cfg.rpn)
rpn_losses = self.rpn_head.loss(*rpn_loss_inputs,
gt_bboxes_ignore=gt_bboxes_ignore)
losses.update(rpn_losses)
proposal_cfg = self.train_cfg.get('rpn_proposal',
self.test_cfg.rpn)
proposal_inputs = rpn_outs + (img_metas, proposal_cfg)
proposal_list = self.rpn_head.get_bboxes(*proposal_inputs)
else:
proposal_list = proposals
if self.with_bbox:
# assign gts and sample proposals
bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
bbox_sampler = build_sampler(self.train_cfg.rcnn.sampler,
context=self)
num_imgs = img.size(0)
if gt_bboxes_ignore is None:
gt_bboxes_ignore = [None for _ in range(num_imgs)]
sampling_results = []
for i in range(num_imgs):
assign_result = bbox_assigner.assign(proposal_list[i],
gt_bboxes[i],
gt_bboxes_ignore[i],
gt_labels[i])
sampling_result = bbox_sampler.sample(
assign_result,
proposal_list[i],
gt_bboxes[i],
gt_labels[i],
feats=[lvl_feat[i][None] for lvl_feat in x])
sampling_results.append(sampling_result)
# bbox head forward and loss
rois = bbox2roi([res.bboxes for res in sampling_results])
# TODO: a more flexible way to decide which feature maps to use
bbox_feats = self.bbox_roi_extractor(
x[:self.bbox_roi_extractor.num_inputs], rois)
if self.with_shared_head:
bbox_feats = self.shared_head(bbox_feats)
cls_score, bbox_pred = self.bbox_head(bbox_feats)
bbox_targets = self.bbox_head.get_target(sampling_results,
gt_bboxes, gt_labels,
self.train_cfg.rcnn)
loss_bbox = self.bbox_head.loss(cls_score, bbox_pred,
*bbox_targets)
losses.update(loss_bbox)
# Grid head forward and loss
sampling_results = self._random_jitter(sampling_results, img_metas)
pos_rois = bbox2roi([res.pos_bboxes for res in sampling_results])
grid_feats = self.grid_roi_extractor(
x[:self.grid_roi_extractor.num_inputs], pos_rois)
if self.with_shared_head:
grid_feats = self.shared_head(grid_feats)
# Accelerate training
max_sample_num_grid = self.train_cfg.rcnn.get('max_num_grid', 192)
sample_idx = torch.randperm(
grid_feats.shape[0])[:min(grid_feats.
shape[0], max_sample_num_grid)]
grid_feats = grid_feats[sample_idx]
grid_pred = self.grid_head(grid_feats)
grid_targets = self.grid_head.get_target(sampling_results,
self.train_cfg.rcnn)
grid_targets = grid_targets[sample_idx]
loss_grid = self.grid_head.loss(grid_pred, grid_targets)
losses.update(loss_grid)
return losses
