def build_dynamic_backbone(cfg, input_shape: ShapeSpec):
"""
Create a Dynamic Backbone from config.
Args:
cfg: a dl_lib CfgNode
Returns:
backbone (Backbone): backbone module, must be a subclass of :class:`Backbone`.
"""
if input_shape is None:
input_shape = ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN))
backbone = DynamicNetwork(init_channel=cfg.MODEL.BACKBONE.INIT_CHANNEL,
input_shape=input_shape,
cell_num_list=cfg.MODEL.BACKBONE.CELL_NUM_LIST,
layer_num=cfg.MODEL.BACKBONE.LAYER_NUM,
norm=cfg.MODEL.BACKBONE.NORM,
cal_flops=cfg.MODEL.CAL_FLOPS,
cell_type=cfg.MODEL.BACKBONE.CELL_TYPE,
max_stride=cfg.MODEL.BACKBONE.MAX_STRIDE,
sep_stem=cfg.MODEL.BACKBONE.SEPT_STEM,
using_gate=cfg.MODEL.GATE.GATE_ON,
small_gate=cfg.MODEL.GATE.SMALL_GATE,
gate_bias=cfg.MODEL.GATE.GATE_INIT_BIAS,
drop_prob=cfg.MODEL.BACKBONE.DROP_PROB)
return backbone
def __init__(self, cfg):
super().__init__()
self.device = torch.device(cfg.MODEL.DEVICE)
self.cfg = cfg
# fmt: off
self.num_classes = cfg.MODEL.CENTERNET.NUM_CLASSES
# Loss parameters:
# Inference parameters:
self.max_detections_per_image = cfg.TEST.DETECTIONS_PER_IMAGE
# fmt: on
self.backbone = cfg.build_backbone(
cfg, input_shape=ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN))
)
self.upsample = cfg.build_upsample_layers(cfg)
self.head = cfg.build_head(cfg)
# self.cls_head = cfg.build_cls_head(cfg)
# self.wh_head = cfg.build_width_height_head(cfg)
# self.reg_head = cfg.build_center_reg_head(cfg)
# backbone_shape = self.backbone.output_shape()
# feature_shapes = [backbone_shape[f] for f in self.in_features]
self.mean, self.std = cfg.MODEL.PIXEL_MEAN, cfg.MODEL.PIXEL_STD
pixel_mean = torch.Tensor(self.mean).to(self.device).view(3, 1, 1)
pixel_std = torch.Tensor(self.std).to(self.device).view(3, 1, 1)
self.normalizer = lambda x: (x - pixel_mean) / pixel_std
self.to(self.device)
def output_shape(self):
return {
name: ShapeSpec(channels=self._out_feature_channels[name],
height=self._out_feature_resolution[name][0],
width=self._out_feature_resolution[name][0],
stride=self._out_feature_strides[name])
for name in self._out_features
}
def output_shape(self):
"""
Returns:
dict[str->ShapeSpec]
"""
# this is a backward-compatible default
return {
name: ShapeSpec(channels=self._out_feature_channels[name],
stride=self._out_feature_strides[name])
for name in self._out_features
}
def build_backbone(cfg, input_shape=None):
"""
Build a backbone.
Returns:
an instance of :class:`Backbone`
"""
if input_shape is None:
input_shape = ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN))
backbone = ResnetBackbone(cfg, input_shape)
assert isinstance(backbone, Backbone)
return backbone
def build_backbone(cfg, input_shape=None):
"""
Build a backbone from `cfg.MODEL.BACKBONE.NAME`.
Returns:
an instance of :class:`Backbone`
"""
if input_shape is None:
input_shape = ShapeSpec(channels=len(cfg.MODEL.PIXEL_MEAN),
height=cfg.INPUT.FIX_SIZE_FOR_FLOPS[0],
width=cfg.INPUT.FIX_SIZE_FOR_FLOPS[1])
backbone = build_dynamic_backbone(cfg, input_shape)
assert isinstance(backbone, Backbone)
return backbone
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
}