def match_anchors(anns,
mlvl_priors,
locations,
ignore_thresh=None,
get_label=lambda x: x['category_id'],
debug=False):
loc_targets = []
cls_targets = []
ignores = []
num_levels, priors_per_level = mlvl_priors.size()[:2]
for (lx, ly), priors in zip(locations, mlvl_priors):
loc_targets.append(priors.new_zeros((lx, ly, priors_per_level, 4)))
cls_targets.append(
priors.new_zeros((lx, ly, priors_per_level), dtype=torch.long))
ignores.append(
priors.new_zeros((lx, ly, priors_per_level), dtype=torch.uint8))
for ann in anns:
l, t, w, h = ann['bbox']
x = l + w / 2
y = t + h / 2
size = torch.tensor([w, h])
ious = iou_1m_with_size(size, mlvl_priors)
max_iou, max_ind = ious.view(-1).max(dim=0)
level, i = divmod(max_ind.item(), priors_per_level)
if debug:
print("[%d,%d]: %.4f" % (level, i, max_iou.item()))
lx, ly = locations[level]
pw, ph = mlvl_priors[level, i]
cx, offset_x = divmod(x * lx, 1)
cx = int(cx)
cy, offset_y = divmod(y * ly, 1)
cy = int(cy)
tx = inverse_sigmoid(offset_x)
ty = inverse_sigmoid(offset_y)
tw = log(w / pw)
th = log(h / ph)
loc_targets[level][cx, cy, i] = torch.tensor([tx, ty, tw, th])
cls_targets[level][cx, cy, i] = get_label(ann)
ignore = ious > ignore_thresh
for level, i in torch.nonzero(ignore):
lx, ly = locations[level]
cx, offset_x = divmod(x * lx, 1)
cx = int(cx)
cy, offset_y = divmod(y * ly, 1)
cy = int(cy)
ignores[level][cx, cy, i] = 1
loc_t = torch.cat([t.view(-1, 4) for t in loc_targets], dim=0)
cls_t = torch.cat([t.view(-1) for t in cls_targets], dim=0)
ignore = torch.cat([t.view(-1) for t in ignores], dim=0)
return loc_t, cls_t, ignore
def __init__(self, num_anchors, num_classes, f_channels=256, num_layers=4, lite=False, concat=True):
super().__init__()
self.num_classes = num_classes
self.concat = concat
self.loc_head = _make_head(
f_channels, num_layers, num_anchors * 4, lite=lite)
self.cls_head = _make_head(
f_channels, num_layers, num_anchors * num_classes, lite=lite)
bias_init_constant(self.cls_head[-1], inverse_sigmoid(0.01))
def __init__(self, num_anchors, in_channels, f_channels=256, lite=False):
super().__init__()
kernel_size = 5 if lite else 3
self.conv = Conv2d(
in_channels, f_channels, kernel_size=kernel_size,
norm_layer='default', activation='default', depthwise_separable=lite)
self.loc_conv = Conv2d(f_channels, num_anchors * 4, kernel_size=1)
self.cls_conv = Conv2d(f_channels, num_anchors * 2, kernel_size=1)
bias_init_constant(self.cls_conv, inverse_sigmoid(0.01))
def __init__(self, num_anchors, num_classes=2, in_channels=245, f_channels=256):
super().__init__()
self.num_classes = num_classes
self.conv = DWConv2d(
in_channels, f_channels, kernel_size=5,
mid_norm_layer='default', norm_layer='default',
activation='default')
self.loc_conv = Conv2d(
f_channels, num_anchors * 4, kernel_size=1)
self.cls_conv = Conv2d(
f_channels, num_anchors * num_classes, kernel_size=1)
bias_init_constant(self.cls_conv, inverse_sigmoid(0.01))
def __init__(self, num_anchors, num_classes, in_channels, lite=False, concat=True):
super().__init__()
self.num_classes = num_classes
self.concat = concat
num_anchors = _tuple(num_anchors, len(in_channels))
self.preds = nn.ModuleList([
nn.Sequential(
get_norm_layer('default', c),
Conv2d(c, n * (num_classes + 4), kernel_size=3, depthwise_separable=lite, mid_norm_layer='default')
)
for c, n in zip(in_channels, num_anchors)
])
for p in self.preds:
get_last_conv(p).bias.data[4:].fill_(inverse_sigmoid(0.01))
def __init__(self, in_channels, num_anchors=3, num_classes=80, lite=False):
super().__init__()
self.num_classes = num_classes
out_channels = num_anchors * (5 + num_classes)
channels = in_channels
self.conv51 = nn.Sequential(
BasicBlock(channels[-1], channels[-1], lite=lite),
BasicBlock(channels[-1], channels[-1], lite=lite),
Conv2d(channels[-1],
channels[-1] // 2,
kernel_size=1,
norm_layer='default',
activation='default'),
)
self.conv52 = Conv2d(channels[-1] // 2,
channels[-1],
kernel_size=3,
norm_layer='default',
activation='default',
depthwise_separable=lite)
self.pred5 = Conv2d(channels[-1], out_channels, kernel_size=1)
self.lat5 = Conv2d(channels[-1] // 2,
channels[-1] // 4,
kernel_size=1,
norm_layer='default')
self.conv41 = nn.Sequential(
BasicBlock(channels[-2] + channels[-1] // 4,
channels[-2],
lite=lite),
BasicBlock(channels[-2], channels[-2], lite=lite),
Conv2d(channels[-2],
channels[-2] // 2,
kernel_size=1,
norm_layer='default',
activation='default'),
)
self.conv42 = Conv2d(channels[-2] // 2,
channels[-2],
kernel_size=3,
norm_layer='default',
activation='default',
depthwise_separable=lite)
self.pred4 = Conv2d(channels[-2], out_channels, kernel_size=1)
self.lat4 = Conv2d(channels[-2] // 2,
channels[-2] // 4,
kernel_size=1,
norm_layer='default')
self.conv31 = nn.Sequential(
BasicBlock(channels[-3] + channels[-2] // 4,
channels[-3],
lite=lite),
BasicBlock(channels[-3], channels[-3], lite=lite),
Conv2d(channels[-3],
channels[-3] // 2,
kernel_size=1,
norm_layer='default',
activation='default'),
)
self.conv32 = Conv2d(channels[-3] // 2,
channels[-3],
kernel_size=3,
norm_layer='default',
activation='default',
depthwise_separable=lite)
self.pred3 = Conv2d(channels[-3], out_channels, kernel_size=1)
get_last_conv(self.pred3).bias.data[4].fill_(inverse_sigmoid(0.01))
get_last_conv(self.pred3).bias.data[4].fill_(inverse_sigmoid(0.01))
get_last_conv(self.pred3).bias.data[4].fill_(inverse_sigmoid(0.01))