def FBNetV2FpnBackbone(cfg, _):
backbone = FBNetV2FPN(
bottom_up=build_fbnet_backbone(cfg),
in_features=cfg.MODEL.FPN.IN_FEATURES,
out_channels=cfg.MODEL.FPN.OUT_CHANNELS,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
)
return backbone
def build_efficientnet_fpn_backbone(cfg, input_shape: ShapeSpec):
bottom_up = build_efficientnet_backbone(cfg, input_shape)
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return backbone
def __init__(self, cfg, input_shape):
bottom_up = build_resnet_backbone(cfg, input_shape)
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
super().__init__(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
input_shapes = bottom_up.output_shape()
self.quantizer = build_quantizer(cfg, input_shapes)
def build_mnetv2_fpn_backbone(cfg, input_shape: ShapeSpec):
bottom_up = build_mnetv2_backbone(cfg, input_shape)
# ['res5', ...]
in_features = cfg.MODEL.FPN.IN_FEATURES
# scalar
out_features = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return backbone
def build_mobilenetv2_flgc_fpn_backbone(cfg, input_shape: ShapeSpec):
out_features = cfg.MODEL.MOBILENET.OUT_FEATURES
bottom_up = MobileNetV2(cfg, groups_in_1x1=8)
bottom_up._out_features = out_features
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return backbone
def fPNBackBone(model_name="resnet34", pretrained=False, freeze_at=2):
# ["res2", "res3", "res4", "res5"]
in_features = ["res2", "res3", "res4", "res5"]
out_channels = 256
norm = "" # "GN"
fuse_type = "sum" # "sum" or "avg"
backbone = FPN(
bottom_up=ResnetBone(model_name, pretrained, freeze_at),
in_features=in_features,
out_channels=out_channels,
norm=norm,
top_block=LastLevelMaxPool(),
fuse_type=fuse_type,
)
return backbone
def build_efficientnetv2_bifpn_backbone(cfg, input_shape: ShapeSpec):
"""
Args:
cfg: a detectron2 CfgNode
Returns:
modeling (Backbone): modeling module, must be a subclass of :class:`Backbone`.
"""
bottom_up = build_efficientnetv2_backbone(cfg, input_shape)
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = BiFPN(
bottom_up=bottom_up,
out_channels=out_channels,
top_block=LastLevelMaxPool(),
)
return backbone
def build_efficientnet_fpn_backbone(cfg, input_shape: ShapeSpec):
"""
Args:
cfg: a detectron2 CfgNode
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`.
"""
bottom_up = build_efficientnet_backbone(cfg, input_shape)
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE)
return backbone
def build_fpn(cfg, input_shape: ShapeSpec):
"""Build feature pyramid network.
Args:
cfg: a detectron2 CfgNode
bottom_up(Backbone): bottom_up backbone
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`.
"""
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
fpn = FPN(
input_shapes=input_shape,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return fpn
def build_vovnet_fpn_backbone(cfg, input_shape: ShapeSpec):
"""Create a VoVNet + FPN instance from config. The type of network (mobilenet-like or deeper model) is defined by the CONV_BODY config entry.
Args:
cfg: a detectron2 CfgNode
input_shape (ShapeSpec): this argument is needed by Detectron2 althought not used here
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`.
"""
bottom_up = build_vovnet_backbone(cfg, input_shape)
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return backbone
def build_mobilenetv2_quantization_fpn_backbone(cfg, input_shape: ShapeSpec):
"""
Args:
cfg: a detectron2 CfgNode
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`.
"""
_out_features = cfg.MODEL.MOBILENET.OUT_FEATURES
bottom_up = mobilenet_v2(cfg)
bottom_up._out_features = _out_features
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return backbone
def build_hrnet_fpn_backbone(cfg, input_shape: ShapeSpec):
"""
Args:
cfg: a detectron2 CfgNode
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`.
"""
bottom_up = HRNet(cfg)
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = HRFPN(
cfg=cfg,
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
# import ipdb;ipdb.set_trace()
return backbone
def build_mobilenet_fpn_backbone(cfg, input_shape: ShapeSpec):
"""Create a MobileNetV2 + FPN instance from config.
Args:
cfg (cfgNode): Detectron2's cfg object
input_shape (ShapeSpec): this argument is needed by Detectron2 althought not used here
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`
"""
bottom_up = build_mobilenet_backbone(cfg, input_shape)
in_features = cfg.MODEL.FPN.IN_FEATURES
out_channels = cfg.MODEL.FPN.OUT_CHANNELS
backbone = FPN(
bottom_up=bottom_up,
in_features=in_features,
out_channels=out_channels,
norm=cfg.MODEL.FPN.NORM,
top_block=LastLevelMaxPool(),
fuse_type=cfg.MODEL.FPN.FUSE_TYPE,
)
return backbone
p4_out = self.p4_out(
self.fuse_p4_out([p4, p4_tr, self.down_p3(p3_out)]))
p5_out = self.p5_out(self.fuse_p5_out([p5, self.down_p4(p4_out)]))
return {
"p2": p2_out,
"p3": p3_out,
"p4": p4_out,
"p5": p5_out,
"p6": self.top_block(p5_out)[0]
}
def output_shape(self):
return {
name: ShapeSpec(channels=self._out_feature_channels[name],
stride=self._out_feature_strides[name])
for name in self._out_features
}
if __name__ == "__main__":
m = timm.create_model('spnasnet_100',
pretrained=True,
features_only=True,
out_indices=(1, 2, 3, 4))
x = torch.rand(1, 3, 224, 224)
m2 = BiFPN(bottom_up=m, out_channels=112, top_block=LastLevelMaxPool())
# torch.jit.trace(m2, x)
m2 = torch.jit.script(m2)
print(m2(x))