ResNet

ResNet backbone based on Deep Residual Learning for Image Recognition.

You can flexibly choose between basicblock and bottleneck as the building blocks for the ResNet architecture. And, you can also freely determine the number of stages and the repetition of blocks within the model. This flexibility supports the creation of various ResNet models, e.g. ResNet18, ResNet34, ResNet50, ResNet101, and ResNet152. Also this supports adjusting the number of stages and blocks for your specific requirements.

Compatibility matrix

Supporting necks	Supporting heads	torch.fx	NetsPresso
FPN YOLOPAFPN	FC ALLMLPDecoder AnchorDecoupledHead AnchorFreeDecoupledHead	Supported	Supported

Field list

Field	Description
name	(str) Name must be "resnet" to use ResNet backbone.
params.block_type	(str) Key value that determines which block to use, "basicblock" or "bottleneck".
params.norm_type	(str) Type of normalization layer. Supporting normalization layers are described in [here].
stage_params[n].channels	(int) The dimension of the first convolution layer in each block.
stage_params[n].num_blocks	(int) The number of blocks in the stage.
stage_params[n].replace_stride_with_dilation	(bool) Flag that determines whether to replace stride step with dilated convolution.

Model configuration examples

ResNet18

model:
  architecture:
    backbone:
      name: resnet
      params:
        block: basicblock
        norm_layer: batch_norm
      stage_params:
        - 
          channels: 64
          layers: 2
        - 
          channels: 128
          layers: 2
          replace_stride_with_dilation: False
        - 
          channels: 256
          layers: 2
          replace_stride_with_dilation: False
        - 
          plane: 512
          layers: 2
          replace_stride_with_dilation: False

ResNet34

model:
  architecture:
    backbone:
      name: resnet
      params:
        block: basicblock
        norm_layer: batch_norm
      stage_params:
        - 
          plane: 64
          layers: 3
        - 
          plane: 128
          layers: 4
          replace_stride_with_dilation: False
        - 
          plane: 256
          layers: 6
          replace_stride_with_dilation: False
        - 
          plane: 512
          layers: 3
          replace_stride_with_dilation: False

ResNet50

model:
  architecture:
    backbone:
      name: resnet
      params:
        block: bottleneck
        norm_layer: batch_norm
      stage_params:
        - 
          plane: 64
          layers: 3
        - 
          plane: 128
          layers: 4
          replace_stride_with_dilation: False
        - 
          plane: 256
          layers: 6
          replace_stride_with_dilation: False
        - 
          plane: 512
          layers: 3
          replace_stride_with_dilation: False

ResNet101

model:
  architecture:
    backbone:
      name: resnet
      params:
        block: bottleneck
        norm_layer: batch_norm
      stage_params:
        - 
          plane: 64
          layers: 3
        - 
          plane: 128
          layers: 4
          replace_stride_with_dilation: False
        - 
          plane: 256
          layers: 23
          replace_stride_with_dilation: False
        - 
          plane: 512
          layers: 3
          replace_stride_with_dilation: False

ResNet152

model:
  architecture:
    backbone:
      name: resnet
      params:
        block: bottleneck
        norm_layer: batch_norm
      stage_params:
        - 
          plane: 64
          layers: 3
        - 
          plane: 128
          layers: 8
          replace_stride_with_dilation: False
        - 
          plane: 256
          layers: 36
          replace_stride_with_dilation: False
        - 
          plane: 512
          layers: 3
          replace_stride_with_dilation: False

Related links

• pytorch/vision