SSD訓練網絡參數計算

一個預測層的網絡結構如下所示：

可以看到，是由三個分支組成的，分別是"PriorBox"層，以及conf、loc的預測層，其中，conf與loc的預測層的參數是由PriorBox的參數計算得到的，具體計算公式如下：

min_size與max_size分別對應一個尺度的預測框（有幾個就對應幾個預測框），in_size只管自己的預測，而max_size是與aspect_ratio聯系在一起的；

filp參數是對應aspect_ratio的預測框*2，以幾個max_size，再乘以幾；最終得到結果為A

conf、loc的參數是在A的基礎上再乘以類別數（加背景），以及4

如下，是需要預測兩類的其中一個尺度的網絡參數；

如上算出的是，每個格子需要預測的conf以及loc的個數；

每個預測層有H*W個格子，因此，總共預測的loc以及conf的個數是需要乘以H*W的；

如下是某一個層的例子（轉自：http://www.360doc.com/content/17/1013/16/42392246_694639090.shtml）

 

注意最后這里的num_priorbox的值與前面的並不一樣，這里是每個預測層所有的輸出框的個數：

 

 

 

 

 

 

 

 

layer {
  name: "combined_2_EltwisePROD_relu"
  type: "ReLU"
  bottom: "combined_2_EltwisePROD"
  top: "combined_2_EltwisePROD_relu"
}
###########################################
###################################################################

layer {
  name: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  type: "Convolution"
  bottom: "combined_2_EltwisePROD_relu"
  top: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
  }
}
layer {
  name: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter_bn"
  type: "BatchNorm"
  bottom: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  top: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    moving_average_fraction: 0.999
    eps: 0.001
  }
}
layer {
  name: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter_scale"
  type: "Scale"
  bottom: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  top: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    filler {
      type: "constant"
      value: 1.0
    }
    bias_term: true
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}



layer {
  name: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  type: "Convolution"
  bottom: "combined_2_EltwisePROD_relu"
  top: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  convolution_param {
    num_output: 128
    bias_term: false
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
  }
}
layer {
  name: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter_bn"
  type: "BatchNorm"
  bottom: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  top: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  batch_norm_param {
    moving_average_fraction: 0.999
    eps: 0.001
  }
}
layer {
  name: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter_scale"
  type: "Scale"
  bottom: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  top: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  scale_param {
    filler {
      type: "constant"
      value: 1.0
    }
    bias_term: true
    bias_filler {
      type: "constant"
      value: 0.0
    }
  }
}

layer {
  name: "combined_2_EltwisePROD_relu_mbox_loc"
  type: "Convolution"
  bottom: "rescombined_2_EltwisePROD_relu_inter256_mbox_locnew_inter"
  top: "combined_2_EltwisePROD_relu_mbox_loc"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    engine: CAFFE
    num_output: 64
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "combined_2_EltwisePROD_relu_mbox_loc_perm"
  type: "Permute"
  bottom: "combined_2_EltwisePROD_relu_mbox_loc"
  top: "combined_2_EltwisePROD_relu_mbox_loc_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "combined_2_EltwisePROD_relu_mbox_loc_flat"
  type: "Flatten"
  bottom: "combined_2_EltwisePROD_relu_mbox_loc_perm"
  top: "combined_2_EltwisePROD_relu_mbox_loc_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "combined_2_EltwisePROD_relu_mbox_conf_new"
  type: "Convolution"
  bottom: "rescombined_2i_EltwisePROD_relu_inter256_mbox_locnew_inter"
  top: "combined_2_EltwisePROD_relu_mbox_conf_new"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    engine: CAFFE
    num_output: 32
    pad: 1
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "combined_2_EltwisePROD_relu_mbox_conf_new_perm"
  type: "Permute"
  bottom: "combined_2_EltwisePROD_relu_mbox_conf_new"
  top: "combined_2_EltwisePROD_relu_mbox_conf_new_perm"
  permute_param {
    order: 0
    order: 2
    order: 3
    order: 1
  }
}
layer {
  name: "combined_2_EltwisePROD_relu_mbox_conf_new_flat"
  type: "Flatten"
  bottom: "combined_2_EltwisePROD_relu_mbox_conf_new_perm"
  top: "combined_2_EltwisePROD_relu_mbox_conf_new_flat"
  flatten_param {
    axis: 1
  }
}
layer {
  name: "combined_2_EltwisePROD_relu_mbox_priorbox"
  type: "PriorBox"
  bottom: "combined_2_EltwisePROD_relu"
  bottom: "data"
  top: "combined_2_EltwisePROD_relu_mbox_priorbox"
  prior_box_param {
    min_size: 12.0
    min_size: 6.0
    max_size: 30.0
    max_size: 20.0
    aspect_ratio: 2
    aspect_ratio: 2.5
    aspect_ratio: 3
    flip: true
    clip: false
    variance: 0.1
    variance: 0.1
    variance: 0.2
    variance: 0.2
    step: 4
    offset: 0.5
  }
}