def __init__(self, **kwargs):
super(CondInst, self).__init__()
self.cfg = {**default_cfg, **kwargs}
self.backbone = getattr(resnet, self.cfg['backbone'])(
pretrained=self.cfg['pretrained'],
reduction=self.cfg['reduction'],
norm_layer=self.cfg['norm_layer'])
c3, c4, c5 = self.backbone.inner_channels[-3:]
self.fpn = FPN(c3, c4, c5, out_channel=self.cfg['fpn_channels'])
self.grids = None
self.scales = nn.ModuleList(
[Scale(init_val=1.0) for _ in range(len(self.cfg['strides']))])
self.cate_control_branch = CateControllerBranch(
self.cfg['num_cls'],
self.cfg['fpn_channels'],
self.cfg['conv_inner_channels'],
norm_type='GN',
mask_out_channels=self.cfg['mask_out_channels'])
self.mask_branch = MaskBranch(
in_channels=self.cfg['fpn_channels'],
inner_channels=self.cfg['mask_branch_inner_channels'],
out_channels=self.cfg['mask_out_channels'])
self.dynamic_convs = DynamicConv(self.cate_control_branch.weight_nums,
self.cate_control_branch.bias_nums)
def __init__(self, **kwargs):
self.cfg = {**default_cfg, **kwargs}
super(RetinaNet, self).__init__()
self.backbones = getattr(resnet, self.cfg['backbone'])(pretrained=self.cfg['pretrained'])
c3, c4, c5 = self.backbones.inner_channels
self.neck = FPN(c3, c4, c5, self.cfg['fpn_channel'])
self.head = RetinaHead(in_channel=self.cfg['fpn_channel'],
inner_channel=self.cfg['fpn_channel'],
num_cls=self.cfg['num_cls'],
num_convs=self.cfg['head_conv_num'],
layer_num=5,
anchor_sizes=self.cfg['anchor_sizes'],
anchor_scales=self.cfg['anchor_scales'],
anchor_ratios=self.cfg['anchor_ratios'],
strides=self.cfg['strides'],
block_type=self.cfg['block_type'],
on_reg=self.cfg['on_reg'])
self.loss = PAALoss(
top_k=self.cfg['top_k'],
alpha=self.cfg['alpha'],
gamma=self.cfg['gamma'],
iou_thresh=self.cfg['iou_thresh'],
iou_loss_weight=self.cfg['iou_loss_weight'],
reg_loss_weight=self.cfg['reg_loss_weight'],
iou_type=self.cfg['iou_type'],
allow_low_quality_matches=self.cfg['allow_low_quality_matches']
)
def __init__(self,**kwargs):
self.cfg={**default_cfg,**kwargs}
super(GFocal, self).__init__()
self.backbones=getattr(resnet,self.cfg['backbone'])(pretrained=self.cfg['pretrained'])
c3,c4,c5=self.backbones.inner_channels
self.neck=FPN(c3,c4,c5,self.cfg['fpn_channel'])
self.head=GFocalHead(in_channel=self.cfg['fpn_channel'],
inner_channel=self.cfg['fpn_channel'],
num_cls=self.cfg['num_cls'],
num_convs=self.cfg['head_conv_num'],
layer_num=5,
anchor_sizes=self.cfg['anchor_sizes'],
anchor_scales=self.cfg['anchor_scales'],
anchor_ratios=self.cfg['anchor_ratios'],
strides=self.cfg['strides'],
block_type=self.cfg['block_type'],
reg_max=self.cfg['reg_max'],
subnet_dim=self.cfg['subnet_dim'],
m_top_k=self.cfg['m_top_k'],
add_mean=self.cfg['add_mean'])
self.loss = GFocalLoss(
strides=self.cfg['strides'],
top_k=self.cfg['top_k'],
beta=self.cfg['beta'],
iou_loss_weight=self.cfg['iou_loss_weight'],
reg_loss_weight=self.cfg['reg_loss_weight'],
iou_type=self.cfg['iou_type'],
anchor_num_per_loc=self.head.anchor_nums
)
def __init__(self, cfg=None):
super(FasterRCNN, self).__init__()
if cfg is None:
cfg = default_model_config
else:
cfg = {**default_model_config, **cfg}
self.backbones = switch_backbones(cfg['backbone'])
self.neck = FPN(self.backbones.inner_channels, cfg['fpn_out_channels'])
self.rpn = RPN(cfg)
self.roi_head = ROIHead(cfg, cfg['num_cls'] + 1)
def __init__(self,
anchor_sizes=None,
anchor_scales=None,
anchor_ratios=None,
strides=None,
num_cls=80,
backbone='resnet50'):
super(RetinaNet, self).__init__()
self.backbones = switch_backbones(backbone)
c3, c4, c5 = self.backbones.inner_channels
self.neck = FPN(c3, c4, c5, 256)
self.head = RetinaHead(256, 256, num_cls, 4, 5, anchor_sizes,
anchor_scales, anchor_ratios, strides)
def __init__(self,
layer_limits=None,
strides=None,
num_cls=80,
backbone='resnet50'):
super(FCOS, self).__init__()
self.backbones = switch_backbones(backbone)
c3, c4, c5 = self.backbones.inner_channels
self.neck = FPN(c3, c4, c5, 256)
self.head = FCOSHead(256,
256,
4,
num_cls,
strides,
layer_limits=layer_limits)