Yolov3 spp Darknet version to caffemodel and then to OM model

yolov3-spp darknet Version transfer caffemodel Re turn om Model

Some time ago , I studied how to put YOLOv4 darknet Version of the model is converted to caffemodel It's converting to om Model in Atlas Reasoning on the server , If a passing friend has no impression , You can review the following two posts first .
YOLOv4 darknet Model into caffemodel Re turn om stay Atlas Reasoning on the server
YOLOv4 caffemodel turn om stay Atlas Reasoning on the server

When I was working today, I needed to YOLOv3-spp Model conversion to offline om Model in Altas Reasoning on the server , After practical experiments, it has been successfully converted , Here, write down the process , I hope that's helpful , If you think it's useful , Remember to like, share and pay attention .

YOLOv3-spp And YOLOv3 Only one more than spp(Spatial Pyramid Pooling Spatial pyramid pool structure ), It was proposed by he Kaiming of Microsoft Research Institute , Mainly to solve two problems ：
The clipping area is avoided 、 Problems such as image distortion caused by zoom operation ;
It solves the problem of image repetitive feature extraction by convolution neural network , It greatly improves the speed of generating candidate boxes , And the calculation cost is saved .

YOLOv3-spp It's converting to caffemodel Time and YOLOv4 It's similar , Just follow up prototxt The modification of the file is slightly different , The different parts will be described directly below .

1. modify prototxt file

1.1 Input description

Description of the input section ,YOLOv3-spp and YOLOv4 It's exactly the same , The network size here is 1024*1024, So the

name: "Darkent2Caffe"
input: "data"
input_dim: 1
input_dim: 3
input_dim: 608
input_dim: 608

It is amended as follows

input: "data"
input_shape {
    
  dim: 1
  dim: 3
  dim: 1024
  dim: 1024
}
input: "img_info"
input_shape {
    
  dim: 1
  dim: 4
}

1.2 unsample layer

And before YOLOv4 Of demo The same with "upsample_param" Keyword as search , That is, we can find the unsample layer , Namely :

layer92-conv

layer {
    
    bottom: "layer92-conv"
    top: "layer93-upsample"
    name: "layer93-upsample"
    type: "Upsample"
    upsample_param {
    
        scale: 2
    }
}

Change to

layer {
    
    bottom: "layer92-conv"
    top: "layer93-upsample"
    name: "layer93-upsample"
    type: "Upsample"
    upsample_param {
    
        scale: 1
        stride: 2
    }
}

as well as layer104-conv

layer {
    
    bottom: "layer104-conv"
    top: "layer105-upsample"
    name: "layer105-upsample"
    type: "Upsample"
    upsample_param {
    
        scale: 2
    }
}

Change to

layer {
    
    bottom: "layer104-conv"
    top: "layer105-upsample"
    name: "layer105-upsample"
    type: "Upsample"
    upsample_param {
    
        scale: 1
        stride: 2
    }
}

1.3 YOLO The correlation layer

We all know YOLOv3 Will be output 3 It's a different dimension featuremap, So we need to find these three first Featuremap. Here we use Netron The software opens the completed featuremap Look at these three output layers , Namely :

layer89-conv

layer101-conv

as well as layer113-conv

Then according to , We can write the rest 3 individual YOLO Layer with a YOLODetectionOutput layer , Pay attention to relevant parameters such as classes, boxes,thresh It needs to be modified according to the actual situation of the model , If you don't understand , Please read my previous article , Put the address here again :
YOLOv4 caffemodel turn om stay Atlas Reasoning on the server

Three YOLO layer

layer {
    
	bottom: "layer89-conv"
	top: "yolo1_coords"
	top: "yolo1_obj"
	top: "yolo1_classes"
	name: "yolo1"
	type: "Yolo"
	yolo_param {
    
		boxes: 3
		coords: 4
		classes: 4
		yolo_version: "V3"
		softmax: true
		background: false
    }
}

layer {
    
	bottom: "layer101-conv"
	top: "yolo2_coords"
	top: "yolo2_obj"
	top: "yolo2_classes"
	name: "yolo2"
	type: "Yolo"
	yolo_param {
    
		boxes: 3
		coords: 4
		classes: 4
		yolo_version: "V3"
		softmax: true
		background: false
	}
}

layer {
    
	bottom: "layer113-conv"
	top: "yolo3_coords"
	top: "yolo3_obj"
	top: "yolo3_classes"
	name: "yolo3"
	type: "Yolo"
	yolo_param {
    
		boxes: 3
		coords: 4
		classes: 4
		yolo_version: "V3"
		softmax: true
		background: false
	}
}

One YoloV3DetectionOutput layer

layer {
    
       name: "detection_out3"
       type: "YoloV3DetectionOutput"
       bottom: "yolo1_coords"
       bottom: "yolo2_coords"
       bottom: "yolo3_coords"
       bottom: "yolo1_obj"
       bottom: "yolo2_obj"
       bottom: "yolo3_obj"
       bottom: "yolo1_classes"
       bottom: "yolo2_classes"
       bottom: "yolo3_classes"
       bottom: "img_info"
       top: "box_out"
       top: "box_out_num"
       yolov3_detection_output_param {
    
                           boxes: 3
                           classes: 4
                           relative: true
                           obj_threshold: 0.25
                           score_threshold: 0.5
                           iou_threshold: 0.2
                           pre_nms_topn: 512
                           post_nms_topn: 1024
                           biases_high: 116
                           biases_high: 90
                           biases_high: 156
                           biases_high: 198
                           biases_high: 373
                           biases_high: 326
                           biases_high: 30
                           biases_high: 61
                           biases_high: 62
                           biases_high: 45
                           biases_high: 59
                           biases_high: 119
                           biases_high: 10
                           biases_high: 13
                           biases_high: 16
                           biases_high: 30
                           biases_high: 33
                           biases_high: 23
       }
}

About YoloV3DetectionOutput Of Anchor

stay YOLOv3-spp, If you follow the previous biases_high, biases_mid, biases_low Writing , The size of the detection frame is not suitable for the actual measurement , I put all of them through experiments anchor It's written in biases_high( See... Above YoloV3DetectionOutput) Be able to make om The model is closest to the original model , I don't understand the specific principle , If there are clear students can comment on the area to teach me the principle , thank you ~! As for the following content, it is related to YOLOv4 Partially consistent .

ALL DONE!
The content of this issue is finished , If you think it helps , Hope you can like, share and pay attention , Your attention can give me the motivation to write .

2021/12/22
I write down on my birthday , Sobbing , It's miserable to have to go to work on your birthday .