def append(self, d):
"""
Adds a single image to the blob
:param datom:
:return:
"""
i = len(self.imgs)
self.imgs.append(d['img'])
h, w, scale = d['img_size']
# all anchors
self.im_sizes.append((h, w, scale))
gt_boxes_ = d['gt_boxes'].astype(np.float32) * d['scale']
self.gt_boxes.append(gt_boxes_)
self.gt_classes.append(
np.column_stack((
i * np.ones(d['gt_classes'].shape[0], dtype=np.int64),
d['gt_classes'],
)))
# Add relationship info
if self.is_rel:
self.gt_rels.append(
np.column_stack(
(i * np.ones(d['gt_relations'].shape[0], dtype=np.int64),
d['gt_relations'])))
# Augment with anchor targets
if self.is_train:
train_anchors_, train_anchor_inds_, train_anchor_targets_, train_anchor_labels_ = \
anchor_target_layer(gt_boxes_, (h, w))
self.train_anchors.append(
np.hstack((train_anchors_, train_anchor_targets_)))
self.train_anchor_labels.append(
np.column_stack((
i * np.ones(train_anchor_inds_.shape[0], dtype=np.int64),
train_anchor_inds_,
train_anchor_labels_,
)))
if 'proposals' in d:
self.proposals.append(
np.column_stack(
(i * np.ones(d['proposals'].shape[0], dtype=np.float32),
d['scale'] * d['proposals'].astype(np.float32))))
# -- Add depth images to the blob
if self.use_depth:
self.depth_imgs.append(d['depth_img'])
train_load = cls(
dataset=train_data,
batch_size=batch_size * num_gpus,
shuffle=True,
num_workers=num_workers,
collate_fn=lambda x: coco_collate(
x, num_gpus=num_gpus, is_train=True),
drop_last=True,
# pin_memory=True,
**kwargs,
)
val_load = cls(
dataset=val_data,
batch_size=batch_size * num_gpus,
shuffle=False,
num_workers=num_workers,
collate_fn=lambda x: coco_collate(
x, num_gpus=num_gpus, is_train=False),
drop_last=True,
# pin_memory=True,
**kwargs,
)
return train_load, val_load
if __name__ == '__main__':
train, val = CocoDetection.splits()
gtbox = train[0]['gt_boxes']
img_size = train[0]['img_size']
anchor_strides, labels, bbox_targets = anchor_target_layer(gtbox, img_size)
def append(self, d):
"""
Adds a single image to the blob
:param datom:
:return:
"""
i = len(self.imgs)
self.imgs.append(d['img'])
h, w, scale = d['img_size']
# all anchors
self.im_sizes.append((h, w, scale))
gt_boxes_ = d['gt_boxes'].astype(np.float32) * d['scale']
self.gt_boxes.append(gt_boxes_)
gt_boxes_human_ = d['gt_boxes_human'].astype(np.float32) * d['scale']
self.gt_boxes_human.append(gt_boxes_human_)
### 여기서 image ind를 삽입
self.gt_classes.append(np.column_stack((
i * np.ones(d['gt_classes'].shape[0], dtype=np.int64),
d['gt_classes'],
)))
self.gt_human_classes.append(np.column_stack((
i * np.ones(d['gt_human_classes'].shape[0], dtype=np.int64),
d['gt_human_classes'],
)))
self.gt_hoi_classes.append(np.column_stack((
i * np.ones(d['gt_hoi_classes'].shape[0], dtype=np.int64),
d['gt_hoi_classes'],
)))
# # Add relationship info
# if self.is_rel:
# self.gt_rels.append(np.column_stack((
# i * np.ones(d['gt_relations'].shape[0], dtype=np.int64),
# d['gt_relations'])))
# Augment with anchor targets
if self.is_train:
train_anchors_, train_anchor_inds_, train_anchor_targets_, train_anchor_labels_ = \
anchor_target_layer(gt_boxes_, (h, w))
self.train_anchors.append(np.hstack((train_anchors_, train_anchor_targets_)))
self.train_anchor_labels.append(np.column_stack((
i * np.ones(train_anchor_inds_.shape[0], dtype=np.int64),
train_anchor_inds_,
train_anchor_labels_,
)))
train_anchors_human_, train_anchor_inds_human_, train_anchor_targets_human_, train_anchor_labels_human_ = \
anchor_target_layer(gt_boxes_human_, (h, w))
self.train_anchors_human.append(np.hstack((train_anchors_human_, train_anchor_targets_human_)))
self.train_anchor_labels_human.append(np.column_stack((
i * np.ones(train_anchor_inds_human_.shape[0], dtype=np.int64),
train_anchor_inds_human_,
train_anchor_labels_human_,
)))
def append(self, d):
"""
Adds a single image to the blob
:param datom:
:return:
"""
i = len(self.imgs)
self.imgs.append(d['img'])
h, w, scale = d['img_size']
if self.is_sal:
self.sal_maps.append(d['sal_map'])
if self.need_depth:
self.depth_maps.append(d['depth_map'])
# all anchors
self.im_sizes.append((h, w, scale))
gt_boxes_ = d['gt_boxes'].astype(np.float32) * d['scale']
self.gt_boxes.append(gt_boxes_)
self.gt_classes.append(np.column_stack((
i * np.ones(d['gt_classes'].shape[0], dtype=np.int64),
d['gt_classes'],
)))
if self.need_caption:
self.seq_labels.append(d['seq_labels'])
self.mask_labels.append(d['mask_labels'])
# Add relationship info
if self.is_rel:
self.gt_rels.append(np.column_stack((
i * np.ones(d['gt_relations'].shape[0], dtype=np.int64),
d['gt_relations'])))
if self.is_keyrel:
self.key_rels.append(np.column_stack((
i * np.ones(d['key_rels'].shape[0], dtype=np.int64),
d['key_rels']
)))
if self.need_cocoid:
self.coco_ids.append(d['coco_id'])
# Augment with anchor targets
if self.is_train:
train_anchors_, train_anchor_inds_, train_anchor_targets_, train_anchor_labels_ = \
anchor_target_layer(gt_boxes_, (h, w))
self.train_anchors.append(np.hstack((train_anchors_, train_anchor_targets_)))
self.train_anchor_labels.append(np.column_stack((
i * np.ones(train_anchor_inds_.shape[0], dtype=np.int64),
train_anchor_inds_,
train_anchor_labels_,
)))
if 'proposals' in d:
self.proposals.append(np.column_stack((i * np.ones(d['proposals'].shape[0], dtype=np.float32),
d['scale'] * d['proposals'].astype(np.float32))))