ssmh后处理网络模型prototxt修改

本章节用于说明ssmh网络后处理算子的使用方式，后处理算子的详细参数规格参见《支持Caffe&TensorFlow&ONNX算子清单》>支持Caffe算子清单中的Yolov5FourInputDecodeBox，YoloNms算子，使用后处理算子的prototxt样例如下，算子有4个输入，对应检测头输出的4个feature_map，输入的顺序为feature_map的尺寸由小到大，有多个检测头就接多个后处理即可。

注意：caffe.proto的LayerParameter中已经添加了对应的参数层信息，用户只需要根据模型的具体参数规格修改网络模型prototxt中的参数。

在LayerParameter中添加如下参数层：

message LayerParameter {
...
  optional Yolov5FourInputDecodeBoxParameter yolov5_4_input_decode_box_param = 256;
  optional YoloNmsParameter yolo_nms_param = 257;
...
}

在caffe.proto中增加如下参数：

message Yolov5FourInputDecodeBoxParameter{
    repeated float bias_0 = 1;
    repeated float bias_1= 2;
    repeated float bias_2 = 3;
    repeated float bias_3 = 4;
    required int32 ori_size_h = 5;
    required int32 ori_size_w = 6;
    required int32 anchor = 7;
    required int32 class_num = 8[default = 8];
    required float thresh_front = 9 [default = 0.5];
}

message YoloNmsParameter{
    required int32 shape = 1;
    required float thresh = 2;
    required int32 num_anchor = 3;
    required int32 num_class = 4[default = 8];
    required int32 total_output_proposal_num = 5 [default = 10];
}

后处理对应prototxt文件如下：

layer {
    name: 'Yolov5FourInputDecodeBox_Head_0'
    type: 'Yolov5FourInputDecodeBox'
    bottom: 'Head0_Detect_3'
    bottom: 'Head0_Detect_2'
    bottom: 'Head0_Detect_1'
    bottom: 'Head0_Detect_0'
    top: 'bounding_box_0'
    yolov5_4_input_decode_box_param {
        bias_0: 102
        bias_0: 240
        bias_0: 252
        bias_0: 278
        bias_0: 228
        bias_0: 420
        bias_1: 78
        bias_1: 79
        bias_1: 64
        bias_1: 138
        bias_1: 122
        bias_1: 119
        bias_2: 31
        bias_2: 39
        bias_2: 54
        bias_2: 48
        bias_2: 37
        bias_2: 76
        bias_3: 11
        bias_3: 25
        bias_3: 18
        bias_3: 30
        bias_3: 20
        bias_3: 52
	ori_size_h: 1088
	ori_size_w: 1920
        anchor:3
        class_num: 3
        thresh_front: 0.01
    }
}

layer {
    type: 'YoloNms'
    name: 'yolo_nms_Head_0'
    bottom: 'bounding_box_0'
    top: 'output_proposal_0'
    yolo_nms_param {
        shape: 43350
        thresh: 0.5
        num_anchor: 3
        num_class: 3
        total_output_proposal_num: 20
    }
}

父主题： 样例参考