caffe添加层：Focal Loss的caffe实现_caffe focalloss设置

版权声明：本文为博主原创文章，转载请注明出处，谢谢。    https://blog.csdn.net/wfei101/article/details/79477542
原代码见： 
https://github.com/chuanqi305/FocalLoss

1，caffe.proto 
源文件在src/caffe/proto/目录里 
从492行这些optional里，作者添加了两行:

optional ReLU6Parameter relu6_param = 208;
optional FocalLossParameter focal_loss_param = 147;
从895行这里添加了一行:

optional bool half_pad = 19 [default = false];
从1425行这里添加一行:

optional bool reduce_boxes = 14 [default = false];
从1505行添加了一段:

message ReLU6Parameter{
    enum Engine {
        DEFAULT = 0;
        CAFEE = 1;
        CUDNN = 2;
    }
    optional Engine engine = 2[default = DEFAULT];
}
从1641行添加一段:

message FocalLossParameter{
    enum Engine{
        DEFAULT = 0;
        CAFFE = 1;
        CUDNN = 2;
    }
    optional Engine engine = 1[default = DEFAULT];
 
    //The axis along which to perform the softmax -- may be negative to index
    //from the end(e.g., -1 for the last axis).
    //Any other axes will be evaluated as independent softmaxes.
    optional int32 axis = 2[default = 1];
    optional float alpha = 3[default = 0.25];
    optional float gamma = 4[default = 2.0];
}
2.在src/caffe/layers/下放入focal_loss_layer.cpp和focal_loss_layer.cu文件

3.在include/caffe/layers/下放入focla_loss_layer.hpp和multibox_focal_loss_layer.hpp

4. 修改prototxt：只需要修改一下层名

layer {
  name: "mbox_loss"
  type: "MultiBoxFocalLoss" #change the type
  bottom: "mbox_loc"
  bottom: "mbox_conf"
  bottom: "mbox_priorbox"
  bottom: "label"
  top: "mbox_loss"
  include {
    phase: TRAIN
  }
  propagate_down: true
  propagate_down: true
  propagate_down: false
  propagate_down: false
  loss_param {
    normalization: VALID
  }
  focal_loss_param { #set the alpha and gamma, default is alpha=0.25, gamma=2.0
    alpha: 0.25
    gamma: 2.0
  }
  multibox_loss_param {
    loc_loss_type: SMOOTH_L1
    conf_loss_type: SOFTMAX
    loc_weight: 1.0
    num_classes: 2
    share_location: true
    match_type: PER_PREDICTION
    overlap_threshold: 0.5
    use_prior_for_matching: true
    background_label_id: 0
    use_difficult_gt: true
    neg_pos_ratio: 3.0
    neg_overlap: 0.5
    code_type: CENTER_SIZE
    ignore_cross_boundary_bbox: false
    mining_type: NONE #do not use OHEM
  }
}
重新编译Caffe
--------------------- 
作者：BigCowPeking 
来源：CSDN 
原文：https://blog.csdn.net/wfei101/article/details/79477542 
版权声明：本文为博主原创文章，转载请附上博文链接！