grouped (aka depthwise-separable) convolutions for int8

The latest TensorRT version(4.0.1.6) features support for the group (aka depthwise-separable) convolutions, which makes it possible to convert MobileNet-V2 into TRT execution plan without using plugin layers.

I found group convolutions that can’t be int8. before and after the group convolution, the model switched from int8(fp32) to fp32(int8).How can I improve speed of my model?

Adding reformat layer: conv2_2/linear reformatted input 0 (conv2_2/dwise/bn) from Float(1,56,3136,301056) to Int8(1,56,3136:4,75264)
Adding reformat layer: conv3_1/linear + block_3_1 reformatted input 0 (conv3_1/dwise/bn) from Float(1,56,3136,451584) to Int8(1,56,3136:4,112896)
Adding reformat layer: conv4_3/linear reformatted input 0 (conv4_3/dwise/bn) from Float(1,28,784,150528) to Int8(1,28,784:4,37632)
Adding reformat layer: conv4_4/linear + block_4_4 reformatted input 0 (conv4_4/dwise/bn) from Float(1,28,784,301056) to Int8(1,28,784:4,75264)
Adding reformat layer: conv4_5/linear + block_4_5 reformatted input 0 (conv4_5/dwise/bn) from Float(1,28,784,301056) to Int8(1,28,784:4,75264)
Adding reformat layer: conv4_6/linear + block_4_6 reformatted input 0 (conv4_6/dwise/bn) from Float(1,28,784,301056) to Int8(1,28,784:4,75264)
Adding reformat layer: conv4_7/linear reformatted input 0 (conv4_7/dwise/bn) from Float(1,14,196,75264) to Int8(1,14,196:4,18816)
Adding reformat layer: conv5_1/linear + block_5_1 reformatted input 0 (conv5_1/dwise/bn) from Float(1,14,196,112896) to Int8(1,14,196:4,28224)
Adding reformat layer: conv5_2/linear + block_5_2 reformatted input 0 (conv5_2/dwise/bn) from Float(1,14,196,112896) to Int8(1,14,196:4,28224)
Adding reformat layer: conv5_3/linear reformatted input 0 (conv5_3/dwise/bn) from Float(1,7,49,28224) to Int8(1,7,49:4,7056)
Adding reformat layer: conv6_1/dwise + relu6_1/dwise reformatted input 0 (conv6_1/expand/bn) from Int8(1,7,49:4,11760) to Float(1,7,49,47040)
Adding reformat layer: conv6_1/linear + block_6_1 reformatted input 0 (conv6_1/dwise/bn) from Float(1,7,49,47040) to Int8(1,7,49:4,11760)
Adding reformat layer: conv6_2/dwise + relu6_2/dwise reformatted input 0 (conv6_2/expand/bn) from Int8(1,7,49:4,11760) to Float(1,7,49,47040)
Adding reformat layer: conv6_2/linear + block_6_2 reformatted input 0 (conv6_2/dwise/bn) from Float(1,7,49,47040) to Int8(1,7,49:4,11760)
Adding reformat layer: conv6_3/dwise + relu6_3/dwise reformatted input 0 (conv6_3/expand/bn) from Int8(1,7,49:4,11760) to Float(1,7,49,47040)
Adding reformat layer: conv6_3/linear reformatted input 0 (conv6_3/dwise/bn) from Float(1,7,49,47040) to Int8(1,7,49:4,11760)

mobilenetv2.prototxt:

input: "data"
input_dim: 1
input_dim: 3
input_dim: 224
input_dim: 224
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 2
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv1/bn"
  type: "BatchNorm"
  bottom: "conv1"
  top: "conv1/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv1/scale"
  type: "Scale"
  bottom: "conv1/bn"
  top: "conv1/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "conv1/bn"
  top: "conv1/bn"
}
layer {
  name: "conv2_1/expand"
  type: "Convolution"
  bottom: "conv1/bn"
  top: "conv2_1/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2_1/expand/bn"
  type: "BatchNorm"
  bottom: "conv2_1/expand"
  top: "conv2_1/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv2_1/expand/scale"
  type: "Scale"
  bottom: "conv2_1/expand/bn"
  top: "conv2_1/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu2_1/expand"
  type: "ReLU"
  bottom: "conv2_1/expand/bn"
  top: "conv2_1/expand/bn"
}
layer {
  name: "conv2_1/dwise"
  type: "Convolution"
  bottom: "conv2_1/expand/bn"
  top: "conv2_1/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 32
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv2_1/dwise/bn"
  type: "BatchNorm"
  bottom: "conv2_1/dwise"
  top: "conv2_1/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv2_1/dwise/scale"
  type: "Scale"
  bottom: "conv2_1/dwise/bn"
  top: "conv2_1/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu2_1/dwise"
  type: "ReLU"
  bottom: "conv2_1/dwise/bn"
  top: "conv2_1/dwise/bn"
}
layer {
  name: "conv2_1/linear"
  type: "Convolution"
  bottom: "conv2_1/dwise/bn"
  top: "conv2_1/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 16
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2_1/linear/bn"
  type: "BatchNorm"
  bottom: "conv2_1/linear"
  top: "conv2_1/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv2_1/linear/scale"
  type: "Scale"
  bottom: "conv2_1/linear/bn"
  top: "conv2_1/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv2_2/expand"
  type: "Convolution"
  bottom: "conv2_1/linear/bn"
  top: "conv2_2/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 96
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2_2/expand/bn"
  type: "BatchNorm"
  bottom: "conv2_2/expand"
  top: "conv2_2/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv2_2/expand/scale"
  type: "Scale"
  bottom: "conv2_2/expand/bn"
  top: "conv2_2/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu2_2/expand"
  type: "ReLU"
  bottom: "conv2_2/expand/bn"
  top: "conv2_2/expand/bn"
}
layer {
  name: "conv2_2/dwise"
  type: "Convolution"
  bottom: "conv2_2/expand/bn"
  top: "conv2_2/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 96
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 96
    stride: 2
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv2_2/dwise/bn"
  type: "BatchNorm"
  bottom: "conv2_2/dwise"
  top: "conv2_2/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv2_2/dwise/scale"
  type: "Scale"
  bottom: "conv2_2/dwise/bn"
  top: "conv2_2/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu2_2/dwise"
  type: "ReLU"
  bottom: "conv2_2/dwise/bn"
  top: "conv2_2/dwise/bn"
}
layer {
  name: "conv2_2/linear"
  type: "Convolution"
  bottom: "conv2_2/dwise/bn"
  top: "conv2_2/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 24
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv2_2/linear/bn"
  type: "BatchNorm"
  bottom: "conv2_2/linear"
  top: "conv2_2/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv2_2/linear/scale"
  type: "Scale"
  bottom: "conv2_2/linear/bn"
  top: "conv2_2/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv3_1/expand"
  type: "Convolution"
  bottom: "conv2_2/linear/bn"
  top: "conv3_1/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 144
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3_1/expand/bn"
  type: "BatchNorm"
  bottom: "conv3_1/expand"
  top: "conv3_1/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv3_1/expand/scale"
  type: "Scale"
  bottom: "conv3_1/expand/bn"
  top: "conv3_1/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu3_1/expand"
  type: "ReLU"
  bottom: "conv3_1/expand/bn"
  top: "conv3_1/expand/bn"
}
layer {
  name: "conv3_1/dwise"
  type: "Convolution"
  bottom: "conv3_1/expand/bn"
  top: "conv3_1/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 144
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 144
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv3_1/dwise/bn"
  type: "BatchNorm"
  bottom: "conv3_1/dwise"
  top: "conv3_1/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv3_1/dwise/scale"
  type: "Scale"
  bottom: "conv3_1/dwise/bn"
  top: "conv3_1/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu3_1/dwise"
  type: "ReLU"
  bottom: "conv3_1/dwise/bn"
  top: "conv3_1/dwise/bn"
}
layer {
  name: "conv3_1/linear"
  type: "Convolution"
  bottom: "conv3_1/dwise/bn"
  top: "conv3_1/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 24
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3_1/linear/bn"
  type: "BatchNorm"
  bottom: "conv3_1/linear"
  top: "conv3_1/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv3_1/linear/scale"
  type: "Scale"
  bottom: "conv3_1/linear/bn"
  top: "conv3_1/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_3_1"
  type: "Eltwise"
  bottom: "conv2_2/linear/bn"
  bottom: "conv3_1/linear/bn"
  top: "block_3_1"
}
layer {
  name: "conv3_2/expand"
  type: "Convolution"
  bottom: "block_3_1"
  top: "conv3_2/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 144
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3_2/expand/bn"
  type: "BatchNorm"
  bottom: "conv3_2/expand"
  top: "conv3_2/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv3_2/expand/scale"
  type: "Scale"
  bottom: "conv3_2/expand/bn"
  top: "conv3_2/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu3_2/expand"
  type: "ReLU"
  bottom: "conv3_2/expand/bn"
  top: "conv3_2/expand/bn"
}
layer {
  name: "conv3_2/dwise"
  type: "Convolution"
  bottom: "conv3_2/expand/bn"
  top: "conv3_2/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 144
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 144
    stride: 2
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv3_2/dwise/bn"
  type: "BatchNorm"
  bottom: "conv3_2/dwise"
  top: "conv3_2/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv3_2/dwise/scale"
  type: "Scale"
  bottom: "conv3_2/dwise/bn"
  top: "conv3_2/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu3_2/dwise"
  type: "ReLU"
  bottom: "conv3_2/dwise/bn"
  top: "conv3_2/dwise/bn"
}
layer {
  name: "conv3_2/linear"
  type: "Convolution"
  bottom: "conv3_2/dwise/bn"
  top: "conv3_2/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv3_2/linear/bn"
  type: "BatchNorm"
  bottom: "conv3_2/linear"
  top: "conv3_2/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv3_2/linear/scale"
  type: "Scale"
  bottom: "conv3_2/linear/bn"
  top: "conv3_2/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv4_1/expand"
  type: "Convolution"
  bottom: "conv3_2/linear/bn"
  top: "conv4_1/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 192
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_1/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_1/expand"
  top: "conv4_1/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_1/expand/scale"
  type: "Scale"
  bottom: "conv4_1/expand/bn"
  top: "conv4_1/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_1/expand"
  type: "ReLU"
  bottom: "conv4_1/expand/bn"
  top: "conv4_1/expand/bn"
}
layer {
  name: "conv4_1/dwise"
  type: "Convolution"
  bottom: "conv4_1/expand/bn"
  top: "conv4_1/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 192
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 192
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_1/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_1/dwise"
  top: "conv4_1/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_1/dwise/scale"
  type: "Scale"
  bottom: "conv4_1/dwise/bn"
  top: "conv4_1/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_1/dwise"
  type: "ReLU"
  bottom: "conv4_1/dwise/bn"
  top: "conv4_1/dwise/bn"
}
layer {
  name: "conv4_1/linear"
  type: "Convolution"
  bottom: "conv4_1/dwise/bn"
  top: "conv4_1/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_1/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_1/linear"
  top: "conv4_1/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_1/linear/scale"
  type: "Scale"
  bottom: "conv4_1/linear/bn"
  top: "conv4_1/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_4_1"
  type: "Eltwise"
  bottom: "conv3_2/linear/bn"
  bottom: "conv4_1/linear/bn"
  top: "block_4_1"
}
layer {
  name: "conv4_2/expand"
  type: "Convolution"
  bottom: "block_4_1"
  top: "conv4_2/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 192
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_2/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_2/expand"
  top: "conv4_2/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_2/expand/scale"
  type: "Scale"
  bottom: "conv4_2/expand/bn"
  top: "conv4_2/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_2/expand"
  type: "ReLU"
  bottom: "conv4_2/expand/bn"
  top: "conv4_2/expand/bn"
}
layer {
  name: "conv4_2/dwise"
  type: "Convolution"
  bottom: "conv4_2/expand/bn"
  top: "conv4_2/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 192
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 192
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_2/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_2/dwise"
  top: "conv4_2/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_2/dwise/scale"
  type: "Scale"
  bottom: "conv4_2/dwise/bn"
  top: "conv4_2/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_2/dwise"
  type: "ReLU"
  bottom: "conv4_2/dwise/bn"
  top: "conv4_2/dwise/bn"
}
layer {
  name: "conv4_2/linear"
  type: "Convolution"
  bottom: "conv4_2/dwise/bn"
  top: "conv4_2/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 32
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_2/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_2/linear"
  top: "conv4_2/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_2/linear/scale"
  type: "Scale"
  bottom: "conv4_2/linear/bn"
  top: "conv4_2/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_4_2"
  type: "Eltwise"
  bottom: "block_4_1"
  bottom: "conv4_2/linear/bn"
  top: "block_4_2"
}
layer {
  name: "conv4_3/expand"
  type: "Convolution"
  bottom: "block_4_2"
  top: "conv4_3/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 192
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_3/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_3/expand"
  top: "conv4_3/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_3/expand/scale"
  type: "Scale"
  bottom: "conv4_3/expand/bn"
  top: "conv4_3/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_3/expand"
  type: "ReLU"
  bottom: "conv4_3/expand/bn"
  top: "conv4_3/expand/bn"
}
layer {
  name: "conv4_3/dwise"
  type: "Convolution"
  bottom: "conv4_3/expand/bn"
  top: "conv4_3/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 192
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 192
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_3/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_3/dwise"
  top: "conv4_3/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_3/dwise/scale"
  type: "Scale"
  bottom: "conv4_3/dwise/bn"
  top: "conv4_3/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_3/dwise"
  type: "ReLU"
  bottom: "conv4_3/dwise/bn"
  top: "conv4_3/dwise/bn"
}
layer {
  name: "conv4_3/linear"
  type: "Convolution"
  bottom: "conv4_3/dwise/bn"
  top: "conv4_3/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_3/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_3/linear"
  top: "conv4_3/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_3/linear/scale"
  type: "Scale"
  bottom: "conv4_3/linear/bn"
  top: "conv4_3/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv4_4/expand"
  type: "Convolution"
  bottom: "conv4_3/linear/bn"
  top: "conv4_4/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_4/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_4/expand"
  top: "conv4_4/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_4/expand/scale"
  type: "Scale"
  bottom: "conv4_4/expand/bn"
  top: "conv4_4/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_4/expand"
  type: "ReLU"
  bottom: "conv4_4/expand/bn"
  top: "conv4_4/expand/bn"
}
layer {
  name: "conv4_4/dwise"
  type: "Convolution"
  bottom: "conv4_4/expand/bn"
  top: "conv4_4/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 384
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_4/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_4/dwise"
  top: "conv4_4/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_4/dwise/scale"
  type: "Scale"
  bottom: "conv4_4/dwise/bn"
  top: "conv4_4/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_4/dwise"
  type: "ReLU"
  bottom: "conv4_4/dwise/bn"
  top: "conv4_4/dwise/bn"
}
layer {
  name: "conv4_4/linear"
  type: "Convolution"
  bottom: "conv4_4/dwise/bn"
  top: "conv4_4/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_4/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_4/linear"
  top: "conv4_4/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_4/linear/scale"
  type: "Scale"
  bottom: "conv4_4/linear/bn"
  top: "conv4_4/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_4_4"
  type: "Eltwise"
  bottom: "conv4_3/linear/bn"
  bottom: "conv4_4/linear/bn"
  top: "block_4_4"
}
layer {
  name: "conv4_5/expand"
  type: "Convolution"
  bottom: "block_4_4"
  top: "conv4_5/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_5/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_5/expand"
  top: "conv4_5/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_5/expand/scale"
  type: "Scale"
  bottom: "conv4_5/expand/bn"
  top: "conv4_5/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_5/expand"
  type: "ReLU"
  bottom: "conv4_5/expand/bn"
  top: "conv4_5/expand/bn"
}
layer {
  name: "conv4_5/dwise"
  type: "Convolution"
  bottom: "conv4_5/expand/bn"
  top: "conv4_5/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 384
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_5/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_5/dwise"
  top: "conv4_5/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_5/dwise/scale"
  type: "Scale"
  bottom: "conv4_5/dwise/bn"
  top: "conv4_5/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_5/dwise"
  type: "ReLU"
  bottom: "conv4_5/dwise/bn"
  top: "conv4_5/dwise/bn"
}
layer {
  name: "conv4_5/linear"
  type: "Convolution"
  bottom: "conv4_5/dwise/bn"
  top: "conv4_5/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_5/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_5/linear"
  top: "conv4_5/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_5/linear/scale"
  type: "Scale"
  bottom: "conv4_5/linear/bn"
  top: "conv4_5/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_4_5"
  type: "Eltwise"
  bottom: "block_4_4"
  bottom: "conv4_5/linear/bn"
  top: "block_4_5"
}
layer {
  name: "conv4_6/expand"
  type: "Convolution"
  bottom: "block_4_5"
  top: "conv4_6/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_6/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_6/expand"
  top: "conv4_6/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_6/expand/scale"
  type: "Scale"
  bottom: "conv4_6/expand/bn"
  top: "conv4_6/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_6/expand"
  type: "ReLU"
  bottom: "conv4_6/expand/bn"
  top: "conv4_6/expand/bn"
}
layer {
  name: "conv4_6/dwise"
  type: "Convolution"
  bottom: "conv4_6/expand/bn"
  top: "conv4_6/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 384
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_6/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_6/dwise"
  top: "conv4_6/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_6/dwise/scale"
  type: "Scale"
  bottom: "conv4_6/dwise/bn"
  top: "conv4_6/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_6/dwise"
  type: "ReLU"
  bottom: "conv4_6/dwise/bn"
  top: "conv4_6/dwise/bn"
}
layer {
  name: "conv4_6/linear"
  type: "Convolution"
  bottom: "conv4_6/dwise/bn"
  top: "conv4_6/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 64
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_6/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_6/linear"
  top: "conv4_6/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_6/linear/scale"
  type: "Scale"
  bottom: "conv4_6/linear/bn"
  top: "conv4_6/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_4_6"
  type: "Eltwise"
  bottom: "block_4_5"
  bottom: "conv4_6/linear/bn"
  top: "block_4_6"
}
layer {
  name: "conv4_7/expand"
  type: "Convolution"
  bottom: "block_4_6"
  top: "conv4_7/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_7/expand/bn"
  type: "BatchNorm"
  bottom: "conv4_7/expand"
  top: "conv4_7/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_7/expand/scale"
  type: "Scale"
  bottom: "conv4_7/expand/bn"
  top: "conv4_7/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_7/expand"
  type: "ReLU"
  bottom: "conv4_7/expand/bn"
  top: "conv4_7/expand/bn"
}
layer {
  name: "conv4_7/dwise"
  type: "Convolution"
  bottom: "conv4_7/expand/bn"
  top: "conv4_7/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 384
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 384
    stride: 2
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv4_7/dwise/bn"
  type: "BatchNorm"
  bottom: "conv4_7/dwise"
  top: "conv4_7/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_7/dwise/scale"
  type: "Scale"
  bottom: "conv4_7/dwise/bn"
  top: "conv4_7/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu4_7/dwise"
  type: "ReLU"
  bottom: "conv4_7/dwise/bn"
  top: "conv4_7/dwise/bn"
}
layer {
  name: "conv4_7/linear"
  type: "Convolution"
  bottom: "conv4_7/dwise/bn"
  top: "conv4_7/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 96
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv4_7/linear/bn"
  type: "BatchNorm"
  bottom: "conv4_7/linear"
  top: "conv4_7/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv4_7/linear/scale"
  type: "Scale"
  bottom: "conv4_7/linear/bn"
  top: "conv4_7/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv5_1/expand"
  type: "Convolution"
  bottom: "conv4_7/linear/bn"
  top: "conv5_1/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 576
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5_1/expand/bn"
  type: "BatchNorm"
  bottom: "conv5_1/expand"
  top: "conv5_1/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_1/expand/scale"
  type: "Scale"
  bottom: "conv5_1/expand/bn"
  top: "conv5_1/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu5_1/expand"
  type: "ReLU"
  bottom: "conv5_1/expand/bn"
  top: "conv5_1/expand/bn"
}
layer {
  name: "conv5_1/dwise"
  type: "Convolution"
  bottom: "conv5_1/expand/bn"
  top: "conv5_1/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 576
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 576
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv5_1/dwise/bn"
  type: "BatchNorm"
  bottom: "conv5_1/dwise"
  top: "conv5_1/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_1/dwise/scale"
  type: "Scale"
  bottom: "conv5_1/dwise/bn"
  top: "conv5_1/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu5_1/dwise"
  type: "ReLU"
  bottom: "conv5_1/dwise/bn"
  top: "conv5_1/dwise/bn"
}
layer {
  name: "conv5_1/linear"
  type: "Convolution"
  bottom: "conv5_1/dwise/bn"
  top: "conv5_1/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 96
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5_1/linear/bn"
  type: "BatchNorm"
  bottom: "conv5_1/linear"
  top: "conv5_1/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_1/linear/scale"
  type: "Scale"
  bottom: "conv5_1/linear/bn"
  top: "conv5_1/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_5_1"
  type: "Eltwise"
  bottom: "conv4_7/linear/bn"
  bottom: "conv5_1/linear/bn"
  top: "block_5_1"
}
layer {
  name: "conv5_2/expand"
  type: "Convolution"
  bottom: "block_5_1"
  top: "conv5_2/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 576
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5_2/expand/bn"
  type: "BatchNorm"
  bottom: "conv5_2/expand"
  top: "conv5_2/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_2/expand/scale"
  type: "Scale"
  bottom: "conv5_2/expand/bn"
  top: "conv5_2/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu5_2/expand"
  type: "ReLU"
  bottom: "conv5_2/expand/bn"
  top: "conv5_2/expand/bn"
}
layer {
  name: "conv5_2/dwise"
  type: "Convolution"
  bottom: "conv5_2/expand/bn"
  top: "conv5_2/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 576
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 576
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv5_2/dwise/bn"
  type: "BatchNorm"
  bottom: "conv5_2/dwise"
  top: "conv5_2/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_2/dwise/scale"
  type: "Scale"
  bottom: "conv5_2/dwise/bn"
  top: "conv5_2/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu5_2/dwise"
  type: "ReLU"
  bottom: "conv5_2/dwise/bn"
  top: "conv5_2/dwise/bn"
}
layer {
  name: "conv5_2/linear"
  type: "Convolution"
  bottom: "conv5_2/dwise/bn"
  top: "conv5_2/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 96
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5_2/linear/bn"
  type: "BatchNorm"
  bottom: "conv5_2/linear"
  top: "conv5_2/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_2/linear/scale"
  type: "Scale"
  bottom: "conv5_2/linear/bn"
  top: "conv5_2/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_5_2"
  type: "Eltwise"
  bottom: "block_5_1"
  bottom: "conv5_2/linear/bn"
  top: "block_5_2"
}
layer {
  name: "conv5_3/expand"
  type: "Convolution"
  bottom: "block_5_2"
  top: "conv5_3/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 576
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5_3/expand/bn"
  type: "BatchNorm"
  bottom: "conv5_3/expand"
  top: "conv5_3/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_3/expand/scale"
  type: "Scale"
  bottom: "conv5_3/expand/bn"
  top: "conv5_3/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu5_3/expand"
  type: "ReLU"
  bottom: "conv5_3/expand/bn"
  top: "conv5_3/expand/bn"
}
layer {
  name: "conv5_3/dwise"
  type: "Convolution"
  bottom: "conv5_3/expand/bn"
  top: "conv5_3/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 576
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 576
    stride: 2
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv5_3/dwise/bn"
  type: "BatchNorm"
  bottom: "conv5_3/dwise"
  top: "conv5_3/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_3/dwise/scale"
  type: "Scale"
  bottom: "conv5_3/dwise/bn"
  top: "conv5_3/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu5_3/dwise"
  type: "ReLU"
  bottom: "conv5_3/dwise/bn"
  top: "conv5_3/dwise/bn"
}
layer {
  name: "conv5_3/linear"
  type: "Convolution"
  bottom: "conv5_3/dwise/bn"
  top: "conv5_3/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 160
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv5_3/linear/bn"
  type: "BatchNorm"
  bottom: "conv5_3/linear"
  top: "conv5_3/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv5_3/linear/scale"
  type: "Scale"
  bottom: "conv5_3/linear/bn"
  top: "conv5_3/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv6_1/expand"
  type: "Convolution"
  bottom: "conv5_3/linear/bn"
  top: "conv6_1/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 960
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_1/expand/bn"
  type: "BatchNorm"
  bottom: "conv6_1/expand"
  top: "conv6_1/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_1/expand/scale"
  type: "Scale"
  bottom: "conv6_1/expand/bn"
  top: "conv6_1/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_1/expand"
  type: "ReLU"
  bottom: "conv6_1/expand/bn"
  top: "conv6_1/expand/bn"
}
layer {
  name: "conv6_1/dwise"
  type: "Convolution"
  bottom: "conv6_1/expand/bn"
  top: "conv6_1/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 960
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 960
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv6_1/dwise/bn"
  type: "BatchNorm"
  bottom: "conv6_1/dwise"
  top: "conv6_1/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_1/dwise/scale"
  type: "Scale"
  bottom: "conv6_1/dwise/bn"
  top: "conv6_1/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_1/dwise"
  type: "ReLU"
  bottom: "conv6_1/dwise/bn"
  top: "conv6_1/dwise/bn"
}
layer {
  name: "conv6_1/linear"
  type: "Convolution"
  bottom: "conv6_1/dwise/bn"
  top: "conv6_1/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 160
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_1/linear/bn"
  type: "BatchNorm"
  bottom: "conv6_1/linear"
  top: "conv6_1/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_1/linear/scale"
  type: "Scale"
  bottom: "conv6_1/linear/bn"
  top: "conv6_1/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_6_1"
  type: "Eltwise"
  bottom: "conv5_3/linear/bn"
  bottom: "conv6_1/linear/bn"
  top: "block_6_1"
}
layer {
  name: "conv6_2/expand"
  type: "Convolution"
  bottom: "block_6_1"
  top: "conv6_2/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 960
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_2/expand/bn"
  type: "BatchNorm"
  bottom: "conv6_2/expand"
  top: "conv6_2/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_2/expand/scale"
  type: "Scale"
  bottom: "conv6_2/expand/bn"
  top: "conv6_2/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_2/expand"
  type: "ReLU"
  bottom: "conv6_2/expand/bn"
  top: "conv6_2/expand/bn"
}
layer {
  name: "conv6_2/dwise"
  type: "Convolution"
  bottom: "conv6_2/expand/bn"
  top: "conv6_2/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 960
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 960
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv6_2/dwise/bn"
  type: "BatchNorm"
  bottom: "conv6_2/dwise"
  top: "conv6_2/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_2/dwise/scale"
  type: "Scale"
  bottom: "conv6_2/dwise/bn"
  top: "conv6_2/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_2/dwise"
  type: "ReLU"
  bottom: "conv6_2/dwise/bn"
  top: "conv6_2/dwise/bn"
}
layer {
  name: "conv6_2/linear"
  type: "Convolution"
  bottom: "conv6_2/dwise/bn"
  top: "conv6_2/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 160
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_2/linear/bn"
  type: "BatchNorm"
  bottom: "conv6_2/linear"
  top: "conv6_2/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_2/linear/scale"
  type: "Scale"
  bottom: "conv6_2/linear/bn"
  top: "conv6_2/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "block_6_2"
  type: "Eltwise"
  bottom: "block_6_1"
  bottom: "conv6_2/linear/bn"
  top: "block_6_2"
}
layer {
  name: "conv6_3/expand"
  type: "Convolution"
  bottom: "block_6_2"
  top: "conv6_3/expand"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 960
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_3/expand/bn"
  type: "BatchNorm"
  bottom: "conv6_3/expand"
  top: "conv6_3/expand/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_3/expand/scale"
  type: "Scale"
  bottom: "conv6_3/expand/bn"
  top: "conv6_3/expand/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_3/expand"
  type: "ReLU"
  bottom: "conv6_3/expand/bn"
  top: "conv6_3/expand/bn"
}
layer {
  name: "conv6_3/dwise"
  type: "Convolution"
  bottom: "conv6_3/expand/bn"
  top: "conv6_3/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 960
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 960
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}
layer {
  name: "conv6_3/dwise/bn"
  type: "BatchNorm"
  bottom: "conv6_3/dwise"
  top: "conv6_3/dwise/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_3/dwise/scale"
  type: "Scale"
  bottom: "conv6_3/dwise/bn"
  top: "conv6_3/dwise/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_3/dwise"
  type: "ReLU"
  bottom: "conv6_3/dwise/bn"
  top: "conv6_3/dwise/bn"
}
layer {
  name: "conv6_3/linear"
  type: "Convolution"
  bottom: "conv6_3/dwise/bn"
  top: "conv6_3/linear"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 320
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_3/linear/bn"
  type: "BatchNorm"
  bottom: "conv6_3/linear"
  top: "conv6_3/linear/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_3/linear/scale"
  type: "Scale"
  bottom: "conv6_3/linear/bn"
  top: "conv6_3/linear/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "conv6_4"
  type: "Convolution"
  bottom: "conv6_3/linear/bn"
  top: "conv6_4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 1280
    bias_term: false
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
  }
}
layer {
  name: "conv6_4/bn"
  type: "BatchNorm"
  bottom: "conv6_4"
  top: "conv6_4/bn"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    use_global_stats: true
    eps: 1e-5
  }
}
layer {
  name: "conv6_4/scale"
  type: "Scale"
  bottom: "conv6_4/bn"
  top: "conv6_4/bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    bias_term: true
  }
}
layer {
  name: "relu6_4"
  type: "ReLU"
  bottom: "conv6_4/bn"
  top: "conv6_4/bn"
}
layer {
  name: "pool6"
  type: "Pooling"
  bottom: "conv6_4/bn"
  top: "pool6"
  pooling_param {
    pool: AVE
    global_pooling: true
  }
}
layer {
  name: "fc7"
  type: "Convolution"
  bottom: "pool6"
  top: "fc7"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 1000
    kernel_size: 1
    weight_filler {
      type: "msra"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "prob"
  type: "Softmax"
  bottom: "fc7"
  top: "prob"
}

Hi,

We are checking this internally.
Will update information with you later.

Hi,

What error do you meet on int8 mode?
We have checked your model on a p4 server and it works correctly.

deploy: /home/vickteam/topic_1038633.prototxt
output: prob
int8
Input "data": 3x224x224
Output "prob": 1000x1x1
name=data, bindingIndex=0, buffers.size()=2
name=prob, bindingIndex=1, buffers.size()=2
Average over 10 runs is 7.68952 ms.
Average over 10 runs is 7.68287 ms.
Average over 10 runs is 7.69537 ms.
Average over 10 runs is 7.69495 ms.
Average over 10 runs is 7.69044 ms.
Average over 10 runs is 7.69127 ms.
Average over 10 runs is 7.69137 ms.
Average over 10 runs is 7.68379 ms.
Average over 10 runs is 7.69823 ms.
Average over 10 runs is 7.6928 ms.

Thanks.

set --verbose=true, YOU find that grouped convolutions can’t be int8. input 224x224 ,the network run time is 7.68ms that is very slow.

layer {
  name: "conv2_2/dwise"
  type: "Convolution"
  bottom: "conv2_2/expand/bn"
  top: "conv2_2/dwise"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 96
    bias_term: false
    pad: 1
    kernel_size: 3
    group: 96
    stride: 2
    weight_filler {
      type: "msra"
    }
    engine: CAFFE
  }
}

You can check the run time of some convolutions with the group parameter.these layers are calculated using fp32.

Hi,

Sorry for the late reply.

The INT8 support of depthwise-separable convolutions is in our plan but we cannot disclose any schedule here.
Please pay attention to our announcement for the next release.

Thanks.