darknet yolo average_pooling_layer

最新推荐文章于 2023-08-10 17:42:06 发布

wzw12315

最新推荐文章于 2023-08-10 17:42:06 发布

阅读量858

点赞数

CC 4.0 BY-SA版权

分类专栏： DeepLearning

本文链接：https://blog.youkuaiyun.com/wzw12315/article/details/80844799

DeepLearning 专栏收录该内容

20 篇文章

订阅专栏

本文介绍了一个用于Darknet框架的全局平均池化层的C语言实现。该层通过对输入特征图进行全局平均来简化模型并减少计算量，适用于计算机视觉任务中的特征提取部分。文中详细解释了如何创建、调整大小以及执行前向和后向传播。

#include "avgpool_layer.h"
#include "cuda.h"
#include <stdio.h>

avgpool_layer make_avgpool_layer(int batch, int w, int h, int c)
{
    fprintf(stderr, "avg                     %4d x%4d x%4d   ->  %4d\n",  w, h, c, c);
    avgpool_layer l = {0};
    l.type = AVGPOOL;
    l.batch = batch;
    l.h = h;
    l.w = w;
    l.c = c;
    l.out_w = 1;
    l.out_h = 1;
    l.out_c = c;
    l.outputs = l.out_c;
    l.inputs = h*w*c;
    int output_size = l.outputs * batch;
    l.output =  calloc(output_size, sizeof(float));
    l.delta =   calloc(output_size, sizeof(float));
    l.forward = forward_avgpool_layer;
    l.backward = backward_avgpool_layer;
    #ifdef GPU
    l.forward_gpu = forward_avgpool_layer_gpu;
    l.backward_gpu = backward_avgpool_layer_gpu;
    l.output_gpu  = cuda_make_array(l.output, output_size);
    l.delta_gpu   = cuda_make_array(l.delta, output_size);
    #endif
    return l;
}

void resize_avgpool_layer(avgpool_layer *l, int w, int h)
{
    l->w = w;
    l->h = h;
    l->inputs = h*w*l->c;
}

void forward_avgpool_layer(const avgpool_layer l, network net)
{
    int b,i,k;

	//看代码才发现,darknet实现avgpool_layer居然没有stride的，即对整个图像的某一个通道的所有像素做平均
	//global average pooling 可以替换FC层，减少参数和计算量
    for(b = 0; b < l.batch; ++b){
        for(k = 0; k < l.c; ++k){
            int out_index = k + b*l.c;
            l.output[out_index] = 0;
            for(i = 0; i < l.h*l.w; ++i){
                int in_index = i + l.h*l.w*(k + b*l.c);
                l.output[out_index] += net.input[in_index];
            }
            l.output[out_index] /= l.h*l.w;
        }
    }
}

void backward_avgpool_layer(const avgpool_layer l, network net)
{
    int b,i,k;

    for(b = 0; b < l.batch; ++b){
        for(k = 0; k < l.c; ++k){
            int out_index = k + b*l.c;
            for(i = 0; i < l.h*l.w; ++i){
                int in_index = i + l.h*l.w*(k + b*l.c);
                net.delta[in_index] += l.delta[out_index] / (l.h*l.w);
            }
        }
    }
}