【YOLO3代码详解系列06】正则化层

1 dropout_layer.h

#ifndef DROPOUT_LAYER_H
#define DROPOUT_LAYER_H
#include "layer.h"
#include "network.h"

typedef layer dropout_layer;

// 构建dropout层
dropout_layer make_dropout_layer(int batch, int inputs, float probability);

// dropout层前向传播函数 
void forward_dropout_layer(dropout_layer l, network net);

// dropout层反向传播函数 
void backward_dropout_layer(dropout_layer l, network net);


void resize_dropout_layer(dropout_layer *l, int inputs);

#ifdef GPU
void forward_dropout_layer_gpu(dropout_layer l, network net);
void backward_dropout_layer_gpu(dropout_layer l, network net);

#endif
#endif

2 dropout_layer.c

#include "dropout_layer.h"
#include "utils.h"
#include "cuda.h"
#include <stdlib.h>
#include <stdio.h>

// 构建dropout层
dropout_layer make_dropout_layer(int batch, int inputs, float probability)
{
    dropout_layer l = {0};
    l.type = DROPOUT;
    l.probability = probability; // 神经元舍弃概率（1-probability为保留概率）
    l.inputs = inputs; 
    l.outputs = inputs; 
    l.batch = batch;
    l.rand = calloc(inputs*batch, sizeof(float));
    l.scale = 1./(1.-probability); 
    l.forward = forward_dropout_layer;
    l.backward = backward_dropout_layer;
    
    #ifdef GPU
    l.forward_gpu = forward_dropout_layer_gpu;
    l.backward_gpu = backward_dropout_layer_gpu;
    l.rand_gpu = cuda_make_array(l.rand, inputs*batch);
    #endif
    
    fprintf(stderr, "dropout       p = %.2f               %4d  ->  %4d\n", probability, inputs, inputs);
    return l;
} 

// 重新配置dropout的相关参数，主要修改的是l.rand的尺寸
void resize_dropout_layer(dropout_layer *l, int inputs)
{
    l->rand = realloc(l->rand, l->inputs*l->batch*sizeof(float));
    
    #ifdef GPU
    cuda_free(l->rand_gpu);
    l->rand_gpu = cuda_make_array(l->rand, inputs*l->batch);
    #endif
}

// dropout层前向传播函数
void forward_dropout_layer(dropout_layer l, network net)
{
    int i;
    if (!net.train) return;
	// 遍历dropout层的每一个输入元素，按照指定的概率l.probability置为0或者按l.scale缩放
    for(i = 0; i < l.batch * l.inputs; ++i){
        float r = rand_uniform(0, 1);
	    // 每个输入元素都对应一个随机数，保存在l.rand中
        l.rand[i] = r;
	    // 如果r小于l.probability，则舍弃该输入元素; 注意，舍弃并不是删除，所以输入元素个数总数没变
        if(r < l.probability) net.input[i] = 0;
		// 否则保留该输入元素，并乘以比例因子scale
        else net.input[i] *= l.scale;
    }
}

// dropout层反向传播函数
void backward_dropout_layer(dropout_layer l, network net)
{
    int i;
    if(!net.delta) return;，
    // 遍历当前层的误差项，并根据l.rand的指示反向计算上一层的误差项值
    for(i = 0; i < l.batch * l.inputs; ++i){
        float r = l.rand[i];
        if(r < l.probability) net.delta[i] = 0;
        else net.delta[i] *= l.scale;
    }
}