Caffe Layer Library及理解

Convolution layer  

# convolution
layer {
  name: "loss1/conv"
  type: "Convolution"
  bottom: "loss1/ave_pool"
  top: "loss1/conv"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 128
    kernel_size: 1
    stride:1 # default: stride=1
    pad: 1
    weight_filler {
      # xavier type
      type: "xavier"

      # gaussian type
      #type: "gaussian"
      #std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0.2
    }
  }
}

参数:bottom（输入），top(输出)，num_output(通道数，卷积的个数)，Kernel_size(卷积核的大小)，stride(滑动步长)，pad(填充)，特征输出大小为out_h = image_h + 2*pad_h – kernel_h）/stride_h+ 1，out_w = image_w +2*pad_w – kernel_w）/stride_w + 1。

Deconvolution

layer {
  name: "score2"
  type: "Deconvolution"
  bottom: "score"
  top: "score2"
  param {
    lr_mult: 1
  }
  convolution_param {
    num_output: 21
    kernel_size: 4
    stride: 2
    weight_filler: { type: "bilinear" }
  }
}

参数与卷积相同，out_h = (in_h - 1) * stride_h + kernel_h，out_w = (in_w - 1) * stride_w + kernel_w。

Dilation Convolution

layer {
  name: "conv5_3"
  type: "Convolution"
  bottom: "conv5_2"
  top: "conv5_3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 2
    dilation: 2 # Actually pad = dilation
  }
}

相对卷积层，多了参数dilation，dilation表示hole size，空洞卷积的特性。

Pooling

max pool

layer {
  name: "pool1_3x3_s2"
  type: "Pooling"
  bottom: "conv1_3_3x3"
  top: "pool1_3x3_s2"
  pooling_param {
    pool: MAX
    kernel_size: 3
    stride: 2
    pad: 1
  }
}

参数与卷积相近

ave pool

layer {
  name: "conv5_3_pool1"
  type: "Pooling"
  bottom: "conv5_3"
  top: "conv5_3_pool1"
  pooling_param {
    pool: AVE
    kernel_size: 60
    stride: 60
  }
}

Upsample

layer {
  name: "upsample4"
  type: "Upsample"
  bottom: "conv5_1_D"
  top: "pool4_D"
  bottom: "pool4_mask"
  upsample_param {
    scale: 2
    upsample_w: 60
    upsample_h: 45
  }
}

参数增加upsample_w和upsample_h，如果该层设定了该参数则输出特征图大小为upsample_w*upsample_h。

Eltwise

layer {
    bottom: "conv4_3"
    bottom: "res_conv4"
    top: "fusion_res_cov4"
    name: "fusion_res_cov4"
    type: "Eltwise"
    eltwise_param { operation: SUM } # PROD SUM MAX
} 

将两个特征特征加和，参数包含两个输入bottom和一个输出top。

Concat

layer {
  name: "inception_4a/output"
  type: "Concat"
  bottom: "inception_4a/1x1"
  bottom: "inception_4a/3x3"
  bottom: "inception_4a/5x5"
  bottom: "inception_4a/pool_proj"
  top: "inception_4a/output"
}

在使模型变宽时，常需要把多个分支合并起来作为后续层的输入，参数包括多个输入bottom（输入至少为2）和一个输出。

InnerProduct

layer {
  name: "imagenet_fc"
  type: "InnerProduct"
  bottom: "fc7"
  top: "imagenet_fc"
  param {
    lr_mult: 1
    decay_mult: 250
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  inner_product_param {
    num_output: ${NUM_LABELS}
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0.7
    }
  }
}

inner_product_layer也即全连接层，如下示意图，每个输出都连接到所有的输入。 

Dropout

layer {
  name: "drop6"
  type: "Dropout"
  bottom: "fc6"
  top: "fc6"
  dropout_param {
    dropout_ratio: 0.5
  }
}

参数dropout_ratio即为一个神经元被保留的概率。

Batch Normaliztion

# BatchNorm2  
layer {
  name: "BatchNorm2" 
  #type: "LRN"
  type: "BatchNorm" include { phase: TRAIN}
  bottom: "Concat1"
  top: "BatchNorm2"
  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: false
  }
}

# BatchNorm
layer {
  name: "bn3"
  type: "BatchNorm"
  bottom: "conv3"
  top: "bn3"
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
  param {
    lr_mult: 0
  }
}

# BN
layer {
  name: "spp3_bn"
  type: "BN"
  bottom: "conv_spp_3_ave_pool"
  top: "spp3_bn"
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  param {
    lr_mult: 0
    decay_mult: 0
  }
  bn_param {
    slope_filler {
      type: "constant"
      value: 1
    }
    bias_filler {
      type: "constant"
      value: 0
    }
    frozen: true
    momentum: 0.95
  }
}

LRN

layer {
  name: "norm1"
  type: "LRN"
  bottom: "pool1"
  top: "norm1"
  lrn_param {
    local_size: 5
    alpha: 0.0001
    beta: 0.75
  }
}

  caffe中有一个LRN 层，全称为Local Response Normalization,即局部响应归一化层。

该层参数有：

  normal_region：选择对相邻通道间归一化还是对通道内的空间区域归一化，默认为ACROSS_CHANNELS,

即通道间归一化；

  local_size:两种表示，（1）：通道间归一化是表示求和的通道数；（2）：通道内归一化示表示归一化操作的

区间的边长；local_size的默认值为5；

  alpha:缩放因子，默认值为1；

  beta:指数项，默认值为5；

  局部响应归一化层完成一种“临近抑制”操作，对局部输入区域归一化。

  在通道间归一化模式中，局部区域范围在相邻通道间，但没有空间上的扩张（即尺寸为local_sizeX1X1);

  在通道内归一化模式中，局部区域范围在当前通道内,有空间上的扩张（即1XlocalXloacl);

  对输入值都将除以;其中n为局部尺寸大小：local_size;alpha和beta前面已经经定义。

求和将在当前处于中间的位置的局部区域内进行（如有必要将进行补零）；

Scale

layer {
  name: "slice"
  type: "Slice"
  bottom: "input"
  top: "output1"
  top: "output2"
  top: "output3"
  top: "output4"
  slice_param {
    axis: 1
    slice_point: 1
    slice_point: 3
    slice_point: 4
  }
}

这里假设input的维度是 N*5*H*W ,tops输出的维度分别为 N*1*H*W   N*2*H*W   N*1*H*W   N*1*H*W  。 
这里需要注意的是，如果有slice_point，slice_point的个数一定要等于top的个数减一。 
axis表示要进行分解的维度。 
slice_point的作用是将axis按照slic_point 进行分解。 

slice_point没有设置的时候则对axis进行均匀分解。

label interpolation(差值)

Threshold

layer {
  name: "threshold"
  type: "Threshold"
  bottom: "soft_prob_s1"
  top: "threshold"
  threshold_param {  
    threshold: 1e-36
  }
}

SigmoidGateLayer

layer {
  name: "gate"
  type: "SigmoidGate"
  bottom: "soft_prob_s1"
  top: "gate"
  gate_param {  
    threshold: 0.5
  }
}

ReLU

layer {
  name: "relu1_1"
  type: "ReLU"
  bottom: "conv1_1"
  top: "conv1_1"
}

PReLU

layer {
  name: "relu6"
  bottom: "fc6"
  top: "relu6"
  type: "PReLU"
  prelu_param { 
    filler {
      type: "constant" 
      value: 0.3 
      } 
    channel_shared: false 
  }
}

interpolation

layer{
  bottom:"input"
  top:"output"
  name:"interp_layer"
  type:"Interp"
  interp_param{
     shrink_factor:4
     zoom_factor:3
     pad_beg:0
     pad_end:0
}

#读入图片的高度和宽度
 height_in_ = bottom[0]->height();
 width_in_ = bottom[0]->width();

#根据设定参数调整后的高度，pad_beg,pad_end只能设置成0及0以下。
 height_in_eff_ = height_in_ + pad_beg_ + pad_end_;
 width_in_eff_ = width_in_ + pad_beg_ + pad_end_;

#interp的顺序是先缩小，再放大
 height_out_ = (height_in_eff_ - 1) / shrink_factor + 1;
 width_out_ = (width_in_eff_ - 1) / shrink_factor + 1;
 height_out_ = height_out_ + (height_out_ - 1) * (zoom_factor - 1);
 width_out_ = width_out_ + (width_out_ - 1) * (zoom_factor - 1);

真题思路是先缩小再放大。

参考：

https://blog.csdn.net/grief_of_the_nazgul/article/details/62043799