def encode(target, anchor_boxes, params):
n = anchor_boxes.size(0)
m = target.bbox.size(0)
if m == 0:
return struct(location=target.bbox.new_zeros(n, 4),
classification=target.bbox.new_zeros(n,
dtype=torch.long))
ious = box.iou_matrix(box.point_form(anchor_boxes), target.bbox)
if params.top_anchors > 0:
top_ious, inds = ious.topk(params.top_anchors, dim=0)
ious = ious.scatter(0, inds, top_ious * 2)
max_ious, max_ids = ious.max(1)
class_target = encode_classes(target.label,
max_ious,
max_ids,
match_thresholds=params.match_thresholds)
location = target.bbox[max_ids]
if params.location_loss == "l1":
location = encode_boxes(location, anchor_boxes)
return struct(location=location, classification=class_target)
def random_target():
num_boxes = random.randint(1, 50)
boxes = torch.Tensor(
[random_box(dim, num_classes) for b in range(0, num_boxes)])
boxes = box.point_form(boxes)
label = torch.LongTensor(num_boxes).random_(0, num_classes)
return (boxes, label)
def decode(prediction, anchor_boxes):
'''Decode (encoded) prediction and anchor boxes to give detected boxes.
Args:
preditction: (tensor) box prediction in encoded form, sized [n, 4].
anchor_boxes: (tensor) bounding boxes in extents form, sized [m, 4].
Returns:
boxes: (tensor) detected boxes in point form, sized [k, 4].
label: (tensor) detected class label [k].
'''
assert prediction.shape == anchor_boxes.shape
loc_pos, loc_size = box.split(prediction)
anchor_pos, anchor_size = box.split(anchor_boxes)
pos = loc_pos * anchor_size + anchor_pos
sizes = loc_size.exp() * anchor_size
return box.point_form(torch.cat([pos, sizes], pos.dim() - 1))
def add_image_noise(image):
nonlocal totals
n = image.target._size
centre, size = box.split(box.extents_form(image.target.bbox))
centre.add_(offset * size)
if image.category == 'train':
centre.add_(torch.randn(n, 2) * noise * size)
size.mul_(torch.randn(n, 2) * noise + 1)
noisy = box.point_form(torch.cat([centre, size], 1))
if image.category == 'train':
totals += struct(iou=box.iou_matrix_matched(
noisy, image.target.bbox).sum(),
n=n)
return image._extend(target=image.target._extend(bbox=noisy))
def crop_anchors(boxes, image_dim):
return box.extents_form(clamp(box.point_form(boxes), (0, 0), image_dim))