def output_shape(self):
stage1_shape = ShapeSpec(channels=48, stride=4)
stage2_shape = ShapeSpec(channels=64, stride=8)
stage3_shape = ShapeSpec(channels=80, stride=16)
stage4_shape = ShapeSpec(channels=96, stride=32)
stage5_shape = ShapeSpec(channels=112, stride=64)
return {
'conv_stage1_out': stage1_shape,
'conv_stage2_out': stage2_shape,
'conv_stage3_out': stage3_shape,
'conv_stage4_out': stage4_shape,
'conv_stage5_out': stage5_shape
}
def output_shape(self):
out_shape = {
self.names[i]: ShapeSpec(channels=768, stride=2**(i + 2))
for i in range(len(self.names))
}
return out_shape
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
}
def __init__(self, config: Config, *args, **kwargs):
super().__init__()
self.config = config
pretrained = config.get("pretrained", False)
pretrained_path = config.get("pretrained_path", None)
self.resnet = build_resnet_backbone(config, ShapeSpec(channels=3))
if pretrained:
state_dict = torch.hub.load_state_dict_from_url(pretrained_path,
progress=False)
new_state_dict = OrderedDict()
replace_layer = {"backbone.": ""}
for key, value in state_dict["model"].items():
new_key = re.sub(r"(backbone\.)",
lambda x: replace_layer[x.groups()[0]], key)
new_state_dict[new_key] = value
self.resnet.load_state_dict(new_state_dict, strict=False)
self.out_dim = 2048
def output_shape(self):
"""
Returns:
dict[str->ShapeSpec]
"""
return {"2D-Feature": ShapeSpec(channels=1024, stride=32)}
def output_shape(self):
return {"conv1": ShapeSpec(channels=64, stride=16)}
def output_shape(self):
return {
name: ShapeSpec(channels=c, stride=s)
for name, c, s in zip(self._out_features, self._out_channels,
self._out_strides)
}
def build_backbone(config: omegaconf.DictConfig):
"""
Difference between this and the build_backbone provided
by detectron2 is as follows:
- Different stem, include caffe_maxpool
- Number of blocks is different, unconfigurable in detectron
- Freeze-at operates differently in detectron
"""
input_shape = ShapeSpec(channels=len(config.MODEL.PIXEL_MEAN))
norm = config.MODEL.RESNETS.NORM
stem = BasicStem(
in_channels=input_shape.channels,
out_channels=config.MODEL.RESNETS.STEM_OUT_CHANNELS,
norm=norm,
caffe_maxpool=config.MODEL.MAX_POOL,
)
freeze_at = config.MODEL.BACKBONE.FREEZE_AT
if freeze_at >= 1:
for p in stem.parameters():
p.requires_grad = False
out_features = config.MODEL.RESNETS.OUT_FEATURES
depth = config.MODEL.RESNETS.DEPTH
num_groups = config.MODEL.RESNETS.NUM_GROUPS
width_per_group = config.MODEL.RESNETS.WIDTH_PER_GROUP
bottleneck_channels = num_groups * width_per_group
in_channels = config.MODEL.RESNETS.STEM_OUT_CHANNELS
out_channels = config.MODEL.RESNETS.RES2_OUT_CHANNELS
stride_in_1x1 = config.MODEL.RESNETS.STRIDE_IN_1X1
res5_dilation = config.MODEL.RESNETS.RES5_DILATION
assert res5_dilation in {1, 2}, f"res5_dilation cannot be {res5_dilation}."
num_blocks_per_stage = {
50: [3, 4, 6, 3],
101: [3, 4, 23, 3],
152: [3, 8, 36, 3]
}[depth]
stages = []
out_stage_idx = [{
"res2": 2,
"res3": 3,
"res4": 4,
"res5": 5
}[f] for f in out_features]
max_stage_idx = max(out_stage_idx)
for idx, stage_idx in enumerate(range(2, max_stage_idx + 1)):
dilation = res5_dilation if stage_idx == 5 else 1
first_stride = 1 if idx == 0 or (stage_idx == 5
and dilation == 2) else 2
stage_kargs = {
"num_blocks":
num_blocks_per_stage[idx],
"stride_per_block":
[first_stride] + [1] * (num_blocks_per_stage[idx] - 1),
"in_channels":
in_channels,
"bottleneck_channels":
bottleneck_channels,
"out_channels":
out_channels,
"num_groups":
num_groups,
"norm":
norm,
"stride_in_1x1":
stride_in_1x1,
"dilation":
dilation,
}
stage_kargs["block_class"] = BottleneckBlock
blocks = ResNet.make_stage(**stage_kargs)
in_channels = out_channels
out_channels *= 2
bottleneck_channels *= 2
stages.append(blocks)
return ResNet(stem, stages, out_features=out_features, freeze_at=-1)
def output_shape(self):
return {self.name: ShapeSpec(stride=8, channels=96, height=16, width=16)}
