Skip to content

Docs for NetsPresso Trainer

MobileNetV3

nota-netspresso/netspresso-trainer

MobileNetV3¶

MobileNetV3 backbone based on Searching for MobileNetV3.

We provide a configuration that allows users to define each inverted residual block in MobileNetV3 individually. You can specify in a list format the number and form of each inverted residual block for every stage. Using this, we provide both MobileNetV3-small and MobileNetV3-large.

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 "mobilenetv3" to use `MobileNetV3` backbone.
`stage_params[n].in_channels`	(list[int]) Input dimensions for the inverted residual blocks in the stage.
`stage_params[n].kernel_sizes`	(list[int]) Convolution kernel sizes for the inverted residual blocks in the stage.
`stage_params[n].expanded_channels`	(list[int]) Expanded dimensions for the inverted residual blocks in the stage.
`stage_params[n].out_channels`	(list[int]) Output dimensions for the inverted residual blocks in the stage.
`stage_params[n].use_se`	(list[bool]) Flags that determine whether to use squeeze-and-excitation blocks for the inverted residual blocks in the stage.
`stage_params[n].activation`	(list[str]) Type of activation functions for the inverted residual blocks in the stage. Supporting activation functions are described in [here]
`stage_params[n].stride`	(list[int]) Stride values for the inverted residual blocks included in the stage.
`stage_params[n].dilation`	(list[int]) Dilation values for the inverted residual blocks in the stage.

Model configuration examples¶

MobileNetV3-small

model:
  architecture:
    backbone:
      name: mobilenetv3
      params: ~
      stage_params:
        -
          in_channels: [16]
          kernel_sizes: [3]
          expanded_channels: [16]
          out_channels: [16]
          use_se: [True]
          act_type: ["relu"]
          stride: [2]
        -
          in_channels: [16, 24]
          kernel_sizes: [3, 3]
          expanded_channels: [72, 88]
          out_channels: [24, 24]
          use_se: [False, False]
          act_type: ["relu", "relu"]
          stride: [2, 1]
        -
          in_channels: [24, 40, 40, 40, 48]
          kernel_sizes: [5, 5, 5, 5, 5]
          expanded_channels: [96, 240, 240, 120, 144]
          out_channels: [40, 40, 40, 48, 48]
          use_se: [True, True, True, True, True]
          act_type: ["hard_swish", "hard_swish", "hard_swish", "hard_swish", "hard_swish"]
          stride: [2, 1, 1, 1, 1]
        -
          in_channels: [48, 96, 96]
          kernel_sizes: [5, 5, 5]
          expanded_channels: [288, 576, 576]
          out_channels: [96, 96, 96]
          use_se: [True, True, True]
          act_type: ["hard_swish", "hard_swish", "hard_swish"]
          stride: [2, 1, 1]

MobileNetV3-large

model:
  architecture:
    backbone:
      name: mobilenetv3
      params: ~
      stage_params:
        -
          in_channels: [16, 16, 24]
          kernel_sizes: [3, 3, 3]
          expanded_channels: [16, 64, 72]
          out_channels: [16, 24, 24]
          use_se: [False, False, False]
          act_type: ["relu", "relu", "relu"]
          stride: [1, 2, 1]
        - 
          in_channels: [24, 40, 40]
          kernel_sizes: [5, 5, 5]
          expanded_channels: [72, 120, 120]
          out_channels: [40, 40, 40]
          use_se: [True, True, True]
          act_type: ["relu", "relu", "relu"]
          stride: [2, 1, 1]
        -
          in_channels: [40, 80, 80, 80, 80, 112]
          kernel_sizes: [3, 3, 3, 3, 3, 3]
          expanded_channels: [240, 200, 184, 184, 480, 672]
          out_channels: [80, 80, 80, 80, 112, 112]
          use_se: [False, False, False, False, True, True]
          act_type: ["hard_swish", "hard_swish", "hard_swish", "hard_swish", "hard_swish", "hard_swish"]
          stride: [2, 1, 1, 1, 1, 1]
        -
          in_channels: [112, 160, 160]
          kernel_sizes: [5, 5, 5]
          expanded_channels: [672, 960, 960]
          out_channels: [160, 160, 160]
          use_se: [True, True, True]
          act_type: ["hard_swish", "hard_swish", "hard_swish"]
          stride: [2, 1, 1]

pytorch/vision