def __init__(self, cfg, input_shape: List[ShapeSpec]):
super().__init__(cfg, input_shape)
head_params = cfg.MODEL.META_ARCH
self.name = head_params.NAME
self.in_features = head_params.IN_FEATURES
self.num_points = head_params.NUM_POINTS
assert self.num_points == 2
self.center_sampling_radius = head_params.CENTER_SAMPLING_RADIUS
self.norm_reg_targets = head_params.NORM_REG_TARGETS
self.centerness_on_loc = head_params.CENTERNESS_ON_LOC
self.iou_loss_type = head_params.IOU_LOSS_TYPE
self.pre_nms_thresh = head_params.PRE_NMS_THRESH
self.pre_nms_top_n = head_params.PRE_NMS_TOP_N
self.slender_centerness = head_params.SLENDER_CENTERNESS
# init bbox pred
self.loc_init_conv = nn.Conv2d(self.feat_channels,
self.loc_feat_channels, 3, 1, 1)
self.loc_init_out = nn.Conv2d(self.loc_feat_channels,
self.num_points * 2, 1, 1, 0)
# make feature adaptive layer
self.cls_conv, self.loc_refine_conv = self.make_feature_adaptive_layers(
)
self._prepare_offset()
self.cls_out = nn.Conv2d(self.feat_channels, self.num_classes, 1, 1, 0)
self.ctn_out = nn.Conv2d(self.feat_channels, 1, 1, 1, 0)
self.loc_refine_out = nn.Conv2d(self.loc_feat_channels,
self.num_points * 2, 1, 1, 0)
self.scales_init = nn.ModuleList(
[Scale(init_value=1.0) for _ in range(5)])
self.scales_refine = nn.ModuleList(
[Scale(init_value=1.0) for _ in range(5)])
self._init_weights()
def __init__(self, cfg, input_shape: List[ShapeSpec]):
"""
Arguments:
in_channels (int): number of channels of the input feature
"""
super(FCOSHead, self).__init__()
# TODO: Implement the sigmoid version first.
# fmt: off
in_channels = input_shape[0].channels
num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
self.norm_reg_targets = cfg.MODEL.FCOS.NORM_REG_TARGETS
self.centerness_on_reg = cfg.MODEL.FCOS.CENTERNESS_ON_REG
self.use_dcn_in_tower = cfg.MODEL.FCOS.USE_DCN_IN_TOWER
self.use_dcn_v2 = cfg.MODEL.FCOS.USE_DCN_V2
# fmt: on
cls_tower = []
bbox_tower = []
for i in range(cfg.MODEL.FCOS.NUM_CONVS):
use_dcn = False
use_v2 = True
if self.use_dcn_in_tower and i == cfg.MODEL.FCOS.NUM_CONVS - 1:
conv_func = DFConv2d
bias = False
use_dcn = True
if not self.use_dcn_v2:
use_v2 = False
else:
conv_func = nn.Conv2d
bias = True
if use_dcn and not use_v2:
cls_tower.append(
conv_func(in_channels,
in_channels,
with_modulated_dcn=False,
kernel_size=3,
stride=1,
padding=1,
bias=bias))
else:
cls_tower.append(
conv_func(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1,
bias=bias))
cls_tower.append(nn.GroupNorm(32, in_channels))
cls_tower.append(nn.ReLU())
if use_dcn and not use_v2:
bbox_tower.append(
conv_func(in_channels,
in_channels,
with_modulated_dcn=False,
kernel_size=3,
stride=1,
padding=1,
bias=bias))
else:
bbox_tower.append(
conv_func(in_channels,
in_channels,
kernel_size=3,
stride=1,
padding=1,
bias=bias))
bbox_tower.append(nn.GroupNorm(32, in_channels))
bbox_tower.append(nn.ReLU())
self.add_module('cls_tower', nn.Sequential(*cls_tower))
self.add_module('bbox_tower', nn.Sequential(*bbox_tower))
self.cls_logits = nn.Conv2d(in_channels,
num_classes,
kernel_size=3,
stride=1,
padding=1)
self.bbox_pred = nn.Conv2d(in_channels,
4,
kernel_size=3,
stride=1,
padding=1)
self.centerness = nn.Conv2d(in_channels,
1,
kernel_size=3,
stride=1,
padding=1)
# initialization
for modules in [
self.cls_tower, self.bbox_tower, self.cls_logits,
self.bbox_pred, self.centerness
]:
for l in modules.modules():
if isinstance(l, nn.Conv2d):
torch.nn.init.normal_(l.weight, std=0.01)
torch.nn.init.constant_(l.bias, 0)
# initialize the bias for focal loss
prior_prob = cfg.MODEL.FCOS.PRIOR_PROB
bias_value = -math.log((1 - prior_prob) / prior_prob)
torch.nn.init.constant_(self.cls_logits.bias, bias_value)
self.scales = nn.ModuleList([Scale(init_value=1.0) for _ in range(5)])
def __init__(self, cfg, input_shape: List[ShapeSpec]):
"""
Arguments:
in_channels (int): number of channels of the input feature
"""
super(FCOSRepPointsHead, self).__init__()
# TODO: Implement the sigmoid version first.
# fmt: off
in_channels = input_shape[0].channels
num_classes = cfg.MODEL.FCOS.NUM_CLASSES
self.fpn_strides = cfg.MODEL.FCOS.FPN_STRIDES
self.norm_reg_targets = cfg.MODEL.FCOS.NORM_REG_TARGETS
self.centerness_on_reg = cfg.MODEL.FCOS.CENTERNESS_ON_REG
self.use_dcn_in_tower = cfg.MODEL.FCOS.USE_DCN_IN_TOWER
self.use_dcn_v2 = cfg.MODEL.FCOS.USE_DCN_V2
# fmt: on
cls_tower = []
bbox_tower = []
for i in range(cfg.MODEL.FCOS.NUM_CONVS):
use_dcn = False
use_v2 = True
if self.use_dcn_in_tower and i == cfg.MODEL.FCOS.NUM_CONVS - 1:
conv_func = DFConv2d
bias = False
use_dcn = True
if not self.use_dcn_v2:
use_v2 = False
else:
conv_func = nn.Conv2d
bias = True
if use_dcn and not use_v2:
cls_tower.append(
conv_func(
in_channels, in_channels,
with_modulated_dcn=False, kernel_size=3, stride=1, padding=1, bias=bias
)
)
else:
cls_tower.append(
conv_func(in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias)
)
cls_tower.append(nn.GroupNorm(32, in_channels))
cls_tower.append(nn.ReLU())
if use_dcn and not use_v2:
bbox_tower.append(
conv_func(
in_channels, in_channels,
with_modulated_dcn=False, kernel_size=3, stride=1, padding=1, bias=bias
)
)
else:
bbox_tower.append(
conv_func(in_channels, in_channels, kernel_size=3, stride=1, padding=1, bias=bias)
)
bbox_tower.append(nn.GroupNorm(32, in_channels))
bbox_tower.append(nn.ReLU())
self.add_module('cls_tower', nn.Sequential(*cls_tower))
self.add_module('bbox_tower', nn.Sequential(*bbox_tower))
# rep part
self.point_feat_channels = in_channels
self.num_points = 9
self.dcn_kernel = int(np.sqrt(self.num_points))
self.dcn_pad = int((self.dcn_kernel - 1) / 2)
self.cls_out_channels = num_classes
self.gradient_mul = 0.1
dcn_base = np.arange(-self.dcn_pad,
self.dcn_pad + 1).astype(np.float64)
dcn_base_y = np.repeat(dcn_base, self.dcn_kernel)
dcn_base_x = np.tile(dcn_base, self.dcn_kernel)
dcn_base_offset = np.stack([dcn_base_y, dcn_base_x], axis=1).reshape((-1))
dcn_base_offset = torch.tensor(dcn_base_offset, dtype=torch.float32).view(1, -1, 1, 1)
self.register_buffer("dcn_base_offset", dcn_base_offset)
self.deform_cls_conv = DeformConv(
self.point_feat_channels,
self.point_feat_channels,
self.dcn_kernel, 1, self.dcn_pad)
self.deform_reg_conv = DeformConv(
self.point_feat_channels,
self.point_feat_channels,
self.dcn_kernel, 1, self.dcn_pad)
points_out_dim = 2 * self.num_points
self.offsets_init = nn.Sequential(
nn.Conv2d(self.point_feat_channels,
self.point_feat_channels,
3, 1, 1),
nn.ReLU(inplace=True),
nn.Conv2d(self.point_feat_channels,
points_out_dim,
1, 1, 0))
self.offsets_refine = nn.Sequential(
nn.ReLU(),
nn.Conv2d(self.point_feat_channels,
points_out_dim,
1, 1, 0))
self.logits = nn.Sequential(
nn.ReLU(),
nn.Conv2d(self.point_feat_channels,
self.cls_out_channels,
1, 1, 0))
# self.cls_logits = nn.Conv2d(in_channels, num_classes, kernel_size=3, stride=1, padding=1)
# self.bbox_pred = nn.Conv2d(in_channels, 4, kernel_size=3, stride=1, padding=1)
self.centerness = nn.Conv2d(in_channels, 1, kernel_size=3, stride=1, padding=1)
self.ratio = nn.Conv2d(in_channels, 1, kernel_size=3, stride=1, padding=1)
# initialization
for modules in [self.cls_tower, self.bbox_tower,
# self.cls_logits, self.bbox_pred,
self.offsets_init,
self.offsets_refine,
self.deform_cls_conv,
self.deform_reg_conv,
self.centerness]:
for l in modules.modules():
if isinstance(l, nn.Conv2d):
torch.nn.init.normal_(l.weight, std=0.01)
torch.nn.init.constant_(l.bias, 0)
# initialize the bias for focal loss
prior_prob = cfg.MODEL.FCOS.PRIOR_PROB
bias_value = -math.log((1 - prior_prob) / prior_prob)
# torch.nn.init.constant_(self.cls_logits.bias, bias_value)
for module in self.logits.modules():
if hasattr(module, 'bias') and module.bias is not None:
torch.nn.init.constant_(module.bias, bias_value)
self.scales = nn.ModuleList([Scale(init_value=1.0) for _ in range(5)])