def __init__(self, resnet_type, num_classes=80, use_gn=False, planes=512):
super(YOLOF, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C5_inplanes = int(512 * expand_ratio[resnet_type])
self.dila_encoder = DilatedEncoder(C5_inplanes)
self.num_classes = num_classes
self.planes = planes
self.clsregcnt_head = FCOSClsRegCntHead(self.planes,
self.num_classes,
num_layers=4,
prior=0.01,
use_gn=use_gn,
cnt_on_reg=True)
self.strides = torch.tensor([16], dtype=torch.float)
self.positions = FCOSPositions(self.strides)
self.scales = nn.Parameter(torch.tensor([1.], dtype=torch.float32))
self.trans = nn.ConvTranspose2d(in_channels=C5_inplanes,
out_channels=int(C5_inplanes / 2),
kernel_size=4,
stride=2,
padding=1,
bias=False)
self.c4_out = nn.Conv2d(C5_inplanes, C5_inplanes, 1)
def __init__(self, resnet_type, num_classes=80, use_gn=False, planes=512):
super(YOLOF, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C5_inplanes = int(512 * expand_ratio[resnet_type])
self.dila_encoder = DilatedEncoder(C5_inplanes)
self.num_classes = num_classes
self.planes = planes
self.clsregcnt_head = FCOSClsRegCntHead(self.planes,
self.num_classes,
num_layers=4,
prior=0.01,
use_gn=use_gn,
cnt_on_reg=True)
self.strides = torch.tensor([32], dtype=torch.float)
self.positions = FCOSPositions(self.strides)
self.scales = nn.Parameter(torch.tensor([1.], dtype=torch.float32))
def __init__(self, resnet_type, num_classes=80, planes=256):
super(FCOS, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C3_inplanes, C4_inplanes, C5_inplanes = int(
128 * expand_ratio[resnet_type]), int(
256 * expand_ratio[resnet_type]), int(
512 * expand_ratio[resnet_type])
self.fpn = RetinaFPN(C3_inplanes,
C4_inplanes,
C5_inplanes,
planes,
use_p5=True)
self.num_classes = num_classes
self.planes = planes
self.cls_head = FCOSClsHead(self.planes,
self.num_classes,
num_layers=4,
prior=0.01)
self.regcenter_head = FCOSRegCenterHead(self.planes, num_layers=4)
self.strides = torch.tensor([8, 16, 32, 64, 128], dtype=torch.float)
self.positions = FCOSPositions(self.strides)
self.scales = nn.Parameter(torch.FloatTensor([1., 1., 1., 1., 1.]))
def __init__(self,
resnet_type,
num_classes=80,
use_TransConv=False,
use_gn=False,
fpn_bn=False,
planes=256):
super(FCOS, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C3_inplanes, C4_inplanes, C5_inplanes = int(
128 * expand_ratio[resnet_type]), int(
256 * expand_ratio[resnet_type]), int(
512 * expand_ratio[resnet_type])
if use_TransConv:
self.fpn = RetinaFPN_TransConv(C3_inplanes,
C4_inplanes,
C5_inplanes,
planes,
use_p5=True,
fpn_bn=fpn_bn)
else:
self.fpn = RetinaFPN(C3_inplanes,
C4_inplanes,
C5_inplanes,
planes,
use_p5=True,
fpn_bn=fpn_bn)
self.num_classes = num_classes
self.planes = planes
self.clsregcnt_head = FCOSClsRegCntHead(self.planes,
self.num_classes,
num_layers=4,
prior=0.01,
use_gn=use_gn,
cnt_on_reg=True)
self.strides = torch.tensor([8, 16, 32, 64, 128], dtype=torch.float)
self.positions = FCOSPositions(self.strides)
self.scales = nn.Parameter(
torch.tensor([1., 1., 1., 1., 1.], dtype=torch.float32))
def __init__(self, resnet_type, config):
super(YOLOF, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C5_inplanes = int(512 * expand_ratio[resnet_type])
self.multi_task = config.multi_task
self.num_classes = config.num_classes
self.dubble_run = config.dubble_run
self.clsregcnt_head_1 = nn.Sequential(
NormalNeck(C5_inplanes, config.yolof_encoder_channels),
FCOSClsRegCntHead(config.yolof_encoder_channels,
self.num_classes,
num_layers=4,
prior=config.prior,
use_gn=config.use_GN_head,
cnt_on_reg=config.cnt_on_reg))
if self.multi_task:
if config.use_ttf:
self.clsregcnt_head_2 = nn.Sequential(
TTFNeck(C5_inplanes, config.ttf_out_channels,
config.yolof_encoder_channels),
FCOSClsRegCntHead(config.yolof_encoder_channels,
self.num_classes,
num_layers=4,
prior=config.prior,
use_gn=config.use_GN_head,
cnt_on_reg=config.cnt_on_reg))
else:
self.clsregcnt_head_2 = nn.Sequential(
NormalNeck(C5_inplanes, config.yolof_encoder_channels),
FCOSClsRegCntHead(config.yolof_encoder_channels,
self.num_classes,
num_layers=4,
prior=config.prior,
use_gn=config.use_GN_head,
cnt_on_reg=config.cnt_on_reg))
self.strides = torch.tensor(config.strides, dtype=torch.float)
self.positions = FCOSPositions(self.strides)
self.scale = torch.tensor(1, dtype=torch.float, requires_grad=False)
def __init__(self, resnet_type, num_classes=80):
super(CenterNet, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C5_inplanes = int(512 * expand_ratio[resnet_type])
self.centernet_head = CenterNetHetRegWhHead(
C5_inplanes,
num_classes,
num_layers=3,
out_channels=[256, 128, 64])
def __init__(self, resnet_type, config):
super(YOLOF, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C5_inplanes = int(512 * expand_ratio[resnet_type])
self.inplanes = C5_inplanes
self.up_layers = []
self.shortcuts = []
self.num_layers = config.num_layers
self.ttf_out_channels = config.ttf_out_channels
for i in range(self.num_layers):
self.up_layers.append(
UP_layer(self.inplanes, self.ttf_out_channels[i]))
self.shortcuts.append(
ShortCut(int(C5_inplanes / (2**(i + 1))),
self.ttf_out_channels[i]))
self.inplanes = self.ttf_out_channels[i]
self.dila_encoder = DilatedEncoder(self.inplanes,
config.yolof_encoder_channels)
self.multi_task = config.multi_task
self.num_classes = config.num_classes
self.clsregcnt_head = FCOSClsRegCntHead(config.yolof_encoder_channels,
self.num_classes,
num_layers=4,
prior=config.prior,
use_gn=config.use_GN_head,
cnt_on_reg=config.cnt_on_reg)
self.strides = torch.tensor(config.strides, dtype=torch.float)
self.positions = FCOSPositions(self.strides)
self.scale = torch.tensor(1, dtype=torch.float)
def __init__(self, resnet_type, num_anchors=9, num_classes=80, planes=256):
super(RetinaNet, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C3_inplanes, C4_inplanes, C5_inplanes = int(
128 * expand_ratio[resnet_type]), int(
256 * expand_ratio[resnet_type]), int(
512 * expand_ratio[resnet_type])
self.fpn = RetinaFPN(C3_inplanes, C4_inplanes, C5_inplanes, planes)
self.pan = PAN(planes)
self.num_anchors = num_anchors
self.num_classes = num_classes
self.planes = planes
self.cls_head = RetinaClsHead(self.planes,
self.num_anchors,
self.num_classes,
num_layers=4,
prior=0.01)
self.reg_head = RetinaRegHead(self.planes,
self.num_anchors,
num_layers=4)
self.areas = torch.tensor([[32, 32], [64, 64], [128, 128], [256, 256],
[512, 512]])
self.ratios = torch.tensor([0.5, 1, 2])
self.scales = torch.tensor([2**0, 2**(1.0 / 3.0), 2**(2.0 / 3.0)])
self.strides = torch.tensor([8, 16, 32, 64, 128], dtype=torch.float)
self.anchors = RetinaAnchors(self.areas, self.ratios, self.scales,
self.strides)
def __init__(self, resnet_type, num_classes=80, planes=256):
super(FCOS, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C3_inplanes, C4_inplanes, C5_inplanes = int(
128 * expand_ratio[resnet_type]), int(
256 * expand_ratio[resnet_type]), int(
512 * expand_ratio[resnet_type])
self.fpn = RetinaFPN(C3_inplanes, C4_inplanes, C5_inplanes, planes)
self.num_classes = num_classes
self.planes = planes
self.cls_head = FCOSClsHead(self.planes,
self.num_classes,
num_layers=4,
prior=0.01)
self.regcenter_head = FCOSRegCenterHead(self.planes, num_layers=4)
def __init__(self,
resnet_type,
num_classes=80,
multi_head=False,
selayer=True,
use_ttf=False,
cls_mlp=False,
deploy=False):
super(CenterNet, self).__init__()
self.backbone = ResNetBackbone(resnet_type=resnet_type)
expand_ratio = {
"resnet18": 1,
"resnet34": 1,
"resnet50": 4,
"resnet101": 4,
"resnet152": 4
}
C5_inplanes = int(512 * expand_ratio[resnet_type])
self.multi_head = multi_head
self.use_ttf = use_ttf
self.deploy = deploy
self.centernet_head = CenterNetHetRegWhHead(
C5_inplanes,
num_classes[0],
num_layers=3,
out_channels=[256, 128, 64])
if multi_head:
if selayer:
if use_ttf:
self.centernet_head_2 = nn.Sequential(
TTFHead(C5_inplanes,
num_classes[1],
out_channels=[256, 128, 64],
selayer=selayer))
else:
self.centernet_head_2 = nn.Sequential(
SEBlock(C5_inplanes),
CenterNetHetRegWhHead(C5_inplanes,
num_classes[1],
num_layers=3,
out_channels=[256, 128, 64]))
else:
if use_ttf:
self.centernet_head_2 = TTFHead(
C5_inplanes,
num_classes[1],
out_channels=[256, 128, 64],
selayer=selayer)
else:
self.centernet_head_2 = CenterNetHetRegWhHead(
C5_inplanes,
num_classes[1],
num_layers=3,
out_channels=[256, 128, 64])
self.cls_mlp = cls_mlp
if cls_mlp:
self.cls = nn.Conv2d(num_classes[0], 1, 1)
self.offset = nn.Conv2d(2, 2, 1)
self.wh = nn.Conv2d(2, 2, 1)