【YOLO3代碼詳解系列04】全連接層

1 connected_layer.h

#ifndef CONNECTED_LAYER_H
#define CONNECTED_LAYER_H
#include "activations.h"
#include "layer.h"
#include "network.h"

// 構建全連接層
layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize, int adam);

// 全連接層前向傳播函數
void forward_connected_layer(layer l, network net);

// 全連接層反向傳播函數
void backward_connected_layer(layer l, network net);

// 全連接層更新函數
void update_connected_layer(layer l, update_args a);

#ifdef GPU
void forward_connected_layer_gpu(layer l, network net);
void backward_connected_layer_gpu(layer l, network net);
void update_connected_layer_gpu(layer l, update_args a);
void push_connected_layer(layer l);
void pull_connected_layer(layer l);
#endif

#endif

2 connected_layer.c

#include "connected_layer.h"
#include "convolutional_layer.h"
#include "batchnorm_layer.h"
#include "utils.h"
#include "cuda.h"
#include "blas.h"
#include "gemm.h"
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

// 構建全連接層
layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize, int adam)
{
    int i;
    layer l = {0};
    l.learning_rate_scale = 1;
    l.type = CONNECTED;

    l.inputs = inputs; // 全連接層一張輸入圖片的元素個數
    l.outputs = outputs; // 全連接層對應一張輸入圖片的輸出元素個數
    l.batch=batch; // 一個batch中的圖片數
    l.batch_normalize = batch_normalize; // 是否進行BN
    l.h = 1;  // 全連接層輸入圖片高爲1, 寬也爲1
    l.w = 1;
    l.c = inputs;  // 全連接層的輸入通道數等於單張輸入圖片的元素個數
    l.out_h = 1;  // 全連接層的輸出圖高爲1, 寬也爲1
    l.out_w = 1;
    l.out_c = outputs; // 全連接層輸出圖片的通道數等於單張輸入圖片對應的輸出元素個數

    l.output = calloc(batch*outputs, sizeof(float)); // 全連接層所有輸出
    l.delta = calloc(batch*outputs, sizeof(float)); // 全連接層的誤差項

    l.weight_updates = calloc(inputs*outputs, sizeof(float));  // 全連接層權重係數更新值
    l.bias_updates = calloc(outputs, sizeof(float)); // 全連接層偏置更新值
	
    l.weights = calloc(outputs*inputs, sizeof(float)); // 全連接層權重係數
    l.biases = calloc(outputs, sizeof(float)); // 全連接層偏置
    
	// 全連接層前向、反向、更新函數
    l.forward = forward_connected_layer;
    l.backward = backward_connected_layer;
    l.update = update_connected_layer;
    
    //權重初始化:使用He初始化方法，縮放因子是-1~1之間的均勻分佈，縮放因子等於sqrt(2./inputs)
    float scale = sqrt(2./inputs);
    for(i = 0; i < outputs*inputs; ++i){
        l.weights[i] = scale*rand_uniform(-1, 1);
    }

    // 初始化所有偏置值爲0
    for(i = 0; i < outputs; ++i){
        l.biases[i] = 0;
    }

    if(adam){ //採用Adam優化器
        l.m = calloc(l.inputs*l.outputs, sizeof(float));
        l.v = calloc(l.inputs*l.outputs, sizeof(float));
        l.bias_m = calloc(l.outputs, sizeof(float));
        l.scale_m = calloc(l.outputs, sizeof(float));
        l.bias_v = calloc(l.outputs, sizeof(float));
        l.scale_v = calloc(l.outputs, sizeof(float));
    }
    
    if(batch_normalize){ //採用BN
        l.scales = calloc(outputs, sizeof(float));
        l.scale_updates = calloc(outputs, sizeof(float));
        for(i = 0; i < outputs; ++i){
            l.scales[i] = 1;
        }

        l.mean = calloc(outputs, sizeof(float));
        l.mean_delta = calloc(outputs, sizeof(float));
        l.variance = calloc(outputs, sizeof(float));
        l.variance_delta = calloc(outputs, sizeof(float));

        l.rolling_mean = calloc(outputs, sizeof(float));
        l.rolling_variance = calloc(outputs, sizeof(float));

        l.x = calloc(batch*outputs, sizeof(float));
        l.x_norm = calloc(batch*outputs, sizeof(float));
    }

#ifdef GPU
    l.forward_gpu = forward_connected_layer_gpu;
    l.backward_gpu = backward_connected_layer_gpu;
    l.update_gpu = update_connected_layer_gpu;

    l.weights_gpu = cuda_make_array(l.weights, outputs*inputs);
    l.biases_gpu = cuda_make_array(l.biases, outputs);

    l.weight_updates_gpu = cuda_make_array(l.weight_updates, outputs*inputs);
    l.bias_updates_gpu = cuda_make_array(l.bias_updates, outputs);

    l.output_gpu = cuda_make_array(l.output, outputs*batch);
    l.delta_gpu = cuda_make_array(l.delta, outputs*batch);
    if (adam) {
        l.m_gpu =       cuda_make_array(0, inputs*outputs);
        l.v_gpu =       cuda_make_array(0, inputs*outputs);
        l.bias_m_gpu =  cuda_make_array(0, outputs);
        l.bias_v_gpu =  cuda_make_array(0, outputs);
        l.scale_m_gpu = cuda_make_array(0, outputs);
        l.scale_v_gpu = cuda_make_array(0, outputs);
    }

    if(batch_normalize){
        l.mean_gpu = cuda_make_array(l.mean, outputs);
        l.variance_gpu = cuda_make_array(l.variance, outputs);

        l.rolling_mean_gpu = cuda_make_array(l.mean, outputs);
        l.rolling_variance_gpu = cuda_make_array(l.variance, outputs);

        l.mean_delta_gpu = cuda_make_array(l.mean, outputs);
        l.variance_delta_gpu = cuda_make_array(l.variance, outputs);

        l.scales_gpu = cuda_make_array(l.scales, outputs);
        l.scale_updates_gpu = cuda_make_array(l.scale_updates, outputs);

        l.x_gpu = cuda_make_array(l.output, l.batch*outputs);
        l.x_norm_gpu = cuda_make_array(l.output, l.batch*outputs);
#ifdef CUDNN
        cudnnCreateTensorDescriptor(&l.normTensorDesc);
        cudnnCreateTensorDescriptor(&l.dstTensorDesc);
        cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w); 
        cudnnSetTensor4dDescriptor(l.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l.out_c, 1, 1); 
#endif
    }
#endif
    l.activation = activation;
    fprintf(stderr, "connected                            %4d  ->  %4d\n", inputs, outputs);
    return l;
}

// 全連接層更新函數
void update_connected_layer(layer l, update_args a)
{
    float learning_rate = a.learning_rate*l.learning_rate_scale;
    float momentum = a.momentum;
    float decay = a.decay;
    int batch = a.batch;
	//更新偏置
    axpy_cpu(l.outputs, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
	//計算下次梯度需要的偏置的動量
    scal_cpu(l.outputs, momentum, l.bias_updates, 1);

    if(l.batch_normalize){
        axpy_cpu(l.outputs, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
        scal_cpu(l.outputs, momentum, l.scale_updates, 1);
    }
	//計算權重衰減
    axpy_cpu(l.inputs*l.outputs, -decay*batch, l.weights, 1, l.weight_updates, 1);
	//更新權重
    axpy_cpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
	//計算下次梯度需要的權重的動量
	scal_cpu(l.inputs*l.outputs, momentum, l.weight_updates, 1);
}

// 全連接層前向傳播函數
void forward_connected_layer(layer l, network net)
{
    // 初始化全連接層的所有輸出爲0
    fill_cpu(l.outputs*l.batch, 0, l.output, 1);
    int m = l.batch;
    int k = l.inputs;
    int n = l.outputs;
    float *a = net.input; // 全連接層的輸入數據，維度爲l.batch*l.inputs
    float *b = l.weights; // 全連接層的所有權重，維度爲l.outputs*l.inputs
    float *c = l.output; // 全連接層的所有輸出，維度爲l.batch*l.outputs
    
    gemm(0,1,m,n,k,1,a,k,b,k,1,c,n); // 對b轉置，後進行矩陣乘法得到輸出
    
    if(l.batch_normalize){
        forward_batchnorm_layer(l, net);
    } else {
        add_bias(l.output, l.biases, l.batch, l.outputs, 1);
    }
    
    // 最終得到全連接層的輸出f(Wx+b)
    activate_array(l.output, l.outputs*l.batch, l.activation);
}

// 全連接層反向傳播函數
void backward_connected_layer(layer l, network net)
{
    // 激活函數對加權輸入的導數，並乘以之前得到的l.delta，得到當前層最終的l.delta
	gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);

    if(l.batch_normalize){
        backward_batchnorm_layer(l, net);
    } else {
        // 計算當前全連接層的偏置梯度值
        backward_bias(l.bias_updates, l.delta, l.batch, l.outputs, 1);
    }
    
    // 計算當前全連接層的權重更新值
    int m = l.outputs;
    int k = l.batch;
    int n = l.inputs;
    float *a = l.delta;
    float *b = net.input;
    float *c = l.weight_updates;
    
   gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);

    // 由當前全連接層計算上一層的誤差項
    m = l.batch;
    k = l.outputs;
    n = l.inputs;

    a = l.delta;
    b = l.weights;
    c = net.delta; // 上一層誤差項，維度爲l.batch*l.inputs
	
    if(c) gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
}


void denormalize_connected_layer(layer l)
{
    int i, j;
    for(i = 0; i < l.outputs; ++i){
        float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .000001);
        for(j = 0; j < l.inputs; ++j){
            l.weights[i*l.inputs + j] *= scale;
        }
        l.biases[i] -= l.rolling_mean[i] * scale;
        l.scales[i] = 1;
        l.rolling_mean[i] = 0;
        l.rolling_variance[i] = 1;
    }
}


void statistics_connected_layer(layer l)
{
    if(l.batch_normalize){
        printf("Scales ");
        print_statistics(l.scales, l.outputs);
        /*
           printf("Rolling Mean ");
           print_statistics(l.rolling_mean, l.outputs);
           printf("Rolling Variance ");
           print_statistics(l.rolling_variance, l.outputs);
         */
    }
    printf("Biases ");
    print_statistics(l.biases, l.outputs);
    printf("Weights ");
    print_statistics(l.weights, l.outputs);
}

#ifdef GPU

void pull_connected_layer(layer l)
{
    cuda_pull_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
    cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
    cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
    cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
    if (l.batch_normalize){
        cuda_pull_array(l.scales_gpu, l.scales, l.outputs);
        cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
        cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
    }
}

void push_connected_layer(layer l)
{
    cuda_push_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
    cuda_push_array(l.biases_gpu, l.biases, l.outputs);
    cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
    cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
    if (l.batch_normalize){
        cuda_push_array(l.scales_gpu, l.scales, l.outputs);
        cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
        cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
    }
}

void update_connected_layer_gpu(layer l, update_args a)
{
    float learning_rate = a.learning_rate*l.learning_rate_scale;
    float momentum = a.momentum;
    float decay = a.decay;
    int batch = a.batch;
    if(a.adam){ 
        adam_update_gpu(l.weights_gpu, l.weight_updates_gpu, l.m_gpu, l.v_gpu, a.B1, a.B2, a.eps, decay, learning_rate, l.inputs*l.outputs, batch, a.t);
        adam_update_gpu(l.biases_gpu, l.bias_updates_gpu, l.bias_m_gpu, l.bias_v_gpu, a.B1, a.B2, a.eps, decay, learning_rate, l.outputs, batch, a.t);
        if(l.scales_gpu){
            adam_update_gpu(l.scales_gpu, l.scale_updates_gpu, l.scale_m_gpu, l.scale_v_gpu, a.B1, a.B2, a.eps, decay, learning_rate, l.outputs, batch, a.t);
        }
    }else{
        axpy_gpu(l.outputs, learning_rate/batch, l.bias_updates_gpu, 1, l.biases_gpu, 1);
        scal_gpu(l.outputs, momentum, l.bias_updates_gpu, 1);

        if(l.batch_normalize){
            axpy_gpu(l.outputs, learning_rate/batch, l.scale_updates_gpu, 1, l.scales_gpu, 1);
            scal_gpu(l.outputs, momentum, l.scale_updates_gpu, 1);
        }

        axpy_gpu(l.inputs*l.outputs, -decay*batch, l.weights_gpu, 1, l.weight_updates_gpu, 1);
        axpy_gpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates_gpu, 1, l.weights_gpu, 1);
        scal_gpu(l.inputs*l.outputs, momentum, l.weight_updates_gpu, 1);
    }
}

void forward_connected_layer_gpu(layer l, network net)
{
    fill_gpu(l.outputs*l.batch, 0, l.output_gpu, 1);

    int m = l.batch;
    int k = l.inputs;
    int n = l.outputs;
    float * a = net.input_gpu;
    float * b = l.weights_gpu;
    float * c = l.output_gpu;
    gemm_gpu(0,1,m,n,k,1,a,k,b,k,1,c,n);

    if (l.batch_normalize) {
        forward_batchnorm_layer_gpu(l, net);
    } else {
        add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.outputs, 1);
    }
    activate_array_gpu(l.output_gpu, l.outputs*l.batch, l.activation);
}

void backward_connected_layer_gpu(layer l, network net)
{
    constrain_gpu(l.outputs*l.batch, 1, l.delta_gpu, 1);
    gradient_array_gpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
    if(l.batch_normalize){
        backward_batchnorm_layer_gpu(l, net);
    } else {
        backward_bias_gpu(l.bias_updates_gpu, l.delta_gpu, l.batch, l.outputs, 1);
    }

    int m = l.outputs;
    int k = l.batch;
    int n = l.inputs;
    float * a = l.delta_gpu;
    float * b = net.input_gpu;
    float * c = l.weight_updates_gpu;
    gemm_gpu(1,0,m,n,k,1,a,m,b,n,1,c,n);

    m = l.batch;
    k = l.outputs;
    n = l.inputs;

    a = l.delta_gpu;
    b = l.weights_gpu;
    c = net.delta_gpu;

    if(c) gemm_gpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
}
#endif