【YOLO3代碼詳解系列05】卷積層

1 convolution_layer.h

#ifndef CONVOLUTIONAL_LAYER_H
#define CONVOLUTIONAL_LAYER_H
#include "cuda.h"
#include "image.h"
#include "activations.h"
#include "layer.h"
#include "network.h"

typedef layer convolutional_layer;

#ifdef GPU
void forward_convolutional_layer_gpu(convolutional_layer layer, network net);
void backward_convolutional_layer_gpu(convolutional_layer layer, network net);
void update_convolutional_layer_gpu(convolutional_layer layer, update_args a);

void push_convolutional_layer(convolutional_layer layer);
void pull_convolutional_layer(convolutional_layer layer);

void add_bias_gpu(float *output, float *biases, int batch, int n, int size);
void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size);
void adam_update_gpu(float *w, float *d, float *m, float *v, float B1, float B2, float eps, float decay, float rate, int n, int batch, int t);
#ifdef CUDNN
void cudnn_convolutional_setup(layer *l);
#endif
#endif

// 構建卷積層
convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int groups, int size, int stride, int padding, ACTIVATION activation, int batch_normalize, int binary, int xnor, int adam);
void resize_convolutional_layer(convolutional_layer *layer, int w, int h);

// 卷積層前向傳播函數 
void forward_convolutional_layer(const convolutional_layer layer, network net);

// 卷積層參數更新函數
void update_convolutional_layer(convolutional_layer layer, update_args a);

image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_weights);
void binarize_weights(float *weights, int n, int size, float *binary);
void swap_binary(convolutional_layer *l);
void binarize_weights2(float *weights, int n, int size, char *binary, float *scales);

// 卷積層後向傳播函數 
void backward_convolutional_layer(convolutional_layer layer, network net);

void add_bias(float *output, float *biases, int batch, int n, int size);
void backward_bias(float *bias_updates, float *delta, int batch, int n, int size);

image get_convolutional_image(convolutional_layer layer);
image get_convolutional_delta(convolutional_layer layer);
image get_convolutional_weight(convolutional_layer layer, int i);

int convolutional_out_height(convolutional_layer layer);
int convolutional_out_width(convolutional_layer layer);

#endif

2 convolution_layer.c

#include "convolutional_layer.h"
#include "utils.h"
#include "batchnorm_layer.h"
#include "im2col.h"
#include "col2im.h"
#include "blas.h"
#include "gemm.h"
#include <stdio.h>
#include <time.h>

#ifdef AI2
#include "xnor_layer.h"
#endif

void swap_binary(convolutional_layer *l)
{
    float *swap = l->weights;
    l->weights = l->binary_weights;
    l->binary_weights = swap;

#ifdef GPU
    swap = l->weights_gpu;
    l->weights_gpu = l->binary_weights_gpu;
    l->binary_weights_gpu = swap;
#endif
}

void binarize_weights(float *weights, int n, int size, float *binary)
{
    int i, f;
    for(f = 0; f < n; ++f){
        float mean = 0;
        for(i = 0; i < size; ++i){
            mean += fabs(weights[f*size + i]);
        }
        mean = mean / size;
        for(i = 0; i < size; ++i){
            binary[f*size + i] = (weights[f*size + i] > 0) ? mean : -mean;
        }
    }
}

void binarize_cpu(float *input, int n, float *binary)
{
    int i;
    for(i = 0; i < n; ++i){
        binary[i] = (input[i] > 0) ? 1 : -1;
    }
}

void binarize_input(float *input, int n, int size, float *binary)
{
    int i, s;
    for(s = 0; s < size; ++s){
        float mean = 0;
        for(i = 0; i < n; ++i){
            mean += fabs(input[i*size + s]);
        }
        mean = mean / n;
        for(i = 0; i < n; ++i){
            binary[i*size + s] = (input[i*size + s] > 0) ? mean : -mean;
        }
    }
}
/*      
**  功能：計算卷積層輸出的特徵圖高度函數
**  輸入：卷積層
**  輸出：輸出圖像的高度
*/
int convolutional_out_height(convolutional_layer l)
{
    return (l.h + 2*l.pad - l.size) / l.stride + 1;
}

// 計算卷積層輸出的特徵圖寬度函數
int convolutional_out_width(convolutional_layer l)
{
    return (l.w + 2*l.pad - l.size) / l.stride + 1;
}

image get_convolutional_image(convolutional_layer l)
{
    return float_to_image(l.out_w,l.out_h,l.out_c,l.output);
}

image get_convolutional_delta(convolutional_layer l)
{
    return float_to_image(l.out_w,l.out_h,l.out_c,l.delta);
}

static size_t get_workspace_size(layer l){
#ifdef CUDNN
    if(gpu_index >= 0){
        size_t most = 0;
        size_t s = 0;
        cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
                l.srcTensorDesc,
                l.weightDesc,
                l.convDesc,
                l.dstTensorDesc,
                l.fw_algo,
                &s);
        if (s > most) most = s;
        cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
                l.srcTensorDesc,
                l.ddstTensorDesc,
                l.convDesc,
                l.dweightDesc,
                l.bf_algo,
                &s);
        if (s > most) most = s;
        cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
                l.weightDesc,
                l.ddstTensorDesc,
                l.convDesc,
                l.dsrcTensorDesc,
                l.bd_algo,
                &s);
        if (s > most) most = s;
        return most;
    }
#endif
    return (size_t)l.out_h*l.out_w*l.size*l.size*l.c/l.groups*sizeof(float);
}

#ifdef GPU
#ifdef CUDNN
void cudnn_convolutional_setup(layer *l)
{
    cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w); 
    cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w); 

    cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w); 
    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); 

    cudnnSetFilter4dDescriptor(l->dweightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c/l->groups, l->size, l->size); 
    cudnnSetFilter4dDescriptor(l->weightDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c/l->groups, l->size, l->size); 
    #if CUDNN_MAJOR >= 6
    cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT);
    #else
    cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
    #endif

    #if CUDNN_MAJOR >= 7
    cudnnSetConvolutionGroupCount(l->convDesc, l->groups);
    #else
    if(l->groups > 1){
        error("CUDNN < 7 doesn't support groups, please upgrade!");
    }
    #endif

    cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
            l->srcTensorDesc,
            l->weightDesc,
            l->convDesc,
            l->dstTensorDesc,
            CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
            2000000000,
            &l->fw_algo);
    cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
            l->weightDesc,
            l->ddstTensorDesc,
            l->convDesc,
            l->dsrcTensorDesc,
            CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
            2000000000,
            &l->bd_algo);
    cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
            l->srcTensorDesc,
            l->ddstTensorDesc,
            l->convDesc,
            l->dweightDesc,
            CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
            2000000000,
            &l->bf_algo);
}
#endif
#endif

/*
**  功能： 構建卷積層
**  輸入： batch    每個batch含有的圖片數
**        h               圖片高度
**        w               圖片寬度
**        c               輸入圖片通道數
**        n               卷積核個數
**        size            卷積核尺寸
**        stride          步幅
**        padding         四周補0長度
**        activation      激活函數類別
**        batch_normalize 是否進行BN 
**        binary          是否對權重進行二值化
**        xnor            是否對權重以及輸入進行二值化
**        adam            是否使用adam
*/
convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int groups, int size, int stride, int padding, ACTIVATION activation, int batch_normalize, int binary, int xnor, int adam)
{
    int i;
    convolutional_layer l = {0};
    l.type = CONVOLUTIONAL; 

    l.groups = groups;
    l.h = h; 
    l.w = w;  
    l.c = c; 
    l.n = n; 
    l.binary = binary; 
    l.xnor = xnor; 
    l.batch = batch; 
    l.stride = stride; 
    l.size = size; 
    l.pad = padding; 
    l.batch_normalize = batch_normalize; 

	// 該卷積層權重元素個數=輸入圖像通道數*卷積核個數*卷積核尺寸
    l.weights = calloc(c/groups*n*size*size, sizeof(float));
    l.weight_updates = calloc(c/groups*n*size*size, sizeof(float));

    l.biases = calloc(n, sizeof(float));
    l.bias_updates = calloc(n, sizeof(float));

	// 該卷積層總的權重元素個數
    l.nweights = c/groups*n*size*size;
    l.nbiases = n;

	// 初始化權重：使用He初始化方法，縮放因子是標準正態分佈隨機數，縮放因子等於sqrt(2./(size*size*c)) 
	float scale = sqrt(2./(size*size*c/l.groups));
	for(i = 0; i < l.nweights; ++i) l.weights[i] = scale*rand_normal();

	// 計算輸出特徵圖的寬度和高度
    int out_w = convolutional_out_width(l);
    int out_h = convolutional_out_height(l);
    l.out_h = out_h; 
    l.out_w = out_w; 
    l.out_c = n; // 輸出圖像通道等於卷積核個數
    l.outputs = l.out_h * l.out_w * l.out_c; 
    l.inputs = l.w * l.h * l.c; 

    // 該卷積層所有的輸出
    l.output = calloc(l.batch*l.outputs, sizeof(float));
    l.delta  = calloc(l.batch*l.outputs, sizeof(float));

    // 卷積層三種指針函數：前向，反向，更新
    l.forward = forward_convolutional_layer;
    l.backward = backward_convolutional_layer;
    l.update = update_convolutional_layer;
    
    if(binary){
        l.binary_weights = calloc(l.nweights, sizeof(float));
        l.cweights = calloc(l.nweights, sizeof(char));
        l.scales = calloc(n, sizeof(float));
    }
    
    if(xnor){
        l.binary_weights = calloc(l.nweights, sizeof(float));
        l.binary_input = calloc(l.inputs*l.batch, sizeof(float));
    }

    if(batch_normalize){
        l.scales = calloc(n, sizeof(float));
        l.scale_updates = calloc(n, sizeof(float));
        for(i = 0; i < n; ++i){
            l.scales[i] = 1;
        }

        l.mean = calloc(n, sizeof(float));
        l.variance = calloc(n, sizeof(float));

        l.mean_delta = calloc(n, sizeof(float));
        l.variance_delta = calloc(n, sizeof(float));

        l.rolling_mean = calloc(n, sizeof(float));
        l.rolling_variance = calloc(n, sizeof(float));
        l.x = calloc(l.batch*l.outputs, sizeof(float));
        l.x_norm = calloc(l.batch*l.outputs, sizeof(float));
    }
    
    if(adam){
        l.m = calloc(l.nweights, sizeof(float));
        l.v = calloc(l.nweights, sizeof(float));
        l.bias_m = calloc(n, sizeof(float));
        l.scale_m = calloc(n, sizeof(float));
        l.bias_v = calloc(n, sizeof(float));
        l.scale_v = calloc(n, sizeof(float));
    }

#ifdef GPU
    l.forward_gpu = forward_convolutional_layer_gpu;
    l.backward_gpu = backward_convolutional_layer_gpu;
    l.update_gpu = update_convolutional_layer_gpu;

    if(gpu_index >= 0){
        if (adam) {
            l.m_gpu = cuda_make_array(l.m, l.nweights);
            l.v_gpu = cuda_make_array(l.v, l.nweights);
            l.bias_m_gpu = cuda_make_array(l.bias_m, n);
            l.bias_v_gpu = cuda_make_array(l.bias_v, n);
            l.scale_m_gpu = cuda_make_array(l.scale_m, n);
            l.scale_v_gpu = cuda_make_array(l.scale_v, n);
        }

        l.weights_gpu = cuda_make_array(l.weights, l.nweights);
        l.weight_updates_gpu = cuda_make_array(l.weight_updates, l.nweights);

        l.biases_gpu = cuda_make_array(l.biases, n);
        l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);

        l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n);
        l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);

        if(binary){
            l.binary_weights_gpu = cuda_make_array(l.weights, l.nweights);
        }
        if(xnor){
            l.binary_weights_gpu = cuda_make_array(l.weights, l.nweights);
            l.binary_input_gpu = cuda_make_array(0, l.inputs*l.batch);
        }

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

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

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

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

            l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
            l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
        }
#ifdef CUDNN
        cudnnCreateTensorDescriptor(&l.normTensorDesc);
        cudnnCreateTensorDescriptor(&l.srcTensorDesc);
        cudnnCreateTensorDescriptor(&l.dstTensorDesc);
        cudnnCreateFilterDescriptor(&l.weightDesc);
        cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
        cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
        cudnnCreateFilterDescriptor(&l.dweightDesc);
        cudnnCreateConvolutionDescriptor(&l.convDesc);
        cudnn_convolutional_setup(&l);
#endif
    }
#endif
    l.workspace_size = get_workspace_size(l);
    l.activation = activation;

    fprintf(stderr, "conv  %5d %2d x%2d /%2d  %4d x%4d x%4d   ->  %4d x%4d x%4d  %5.3f BFLOPs\n", n, size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c, (2.0 * l.n * l.size*l.size*l.c/l.groups * l.out_h*l.out_w)/1000000000.);

    return l;
}

void denormalize_convolutional_layer(convolutional_layer l)
{
    int i, j;
    for(i = 0; i < l.n; ++i){
        float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .00001);
        for(j = 0; j < l.c/l.groups*l.size*l.size; ++j){
            l.weights[i*l.c/l.groups*l.size*l.size + 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 resize_convolutional_layer(convolutional_layer *l, int w, int h)
{
    l->w = w;
    l->h = h;
    int out_w = convolutional_out_width(*l);
    int out_h = convolutional_out_height(*l);

    l->out_w = out_w;
    l->out_h = out_h;

    l->outputs = l->out_h * l->out_w * l->out_c;
    l->inputs = l->w * l->h * l->c;

    l->output = realloc(l->output, l->batch*l->outputs*sizeof(float));
    l->delta  = realloc(l->delta,  l->batch*l->outputs*sizeof(float));
    if(l->batch_normalize){
        l->x = realloc(l->x, l->batch*l->outputs*sizeof(float));
        l->x_norm  = realloc(l->x_norm, l->batch*l->outputs*sizeof(float));
    }

#ifdef GPU
    cuda_free(l->delta_gpu);
    cuda_free(l->output_gpu);

    l->delta_gpu =  cuda_make_array(l->delta,  l->batch*l->outputs);
    l->output_gpu = cuda_make_array(l->output, l->batch*l->outputs);

    if(l->batch_normalize){
        cuda_free(l->x_gpu);
        cuda_free(l->x_norm_gpu);

        l->x_gpu = cuda_make_array(l->output, l->batch*l->outputs);
        l->x_norm_gpu = cuda_make_array(l->output, l->batch*l->outputs);
    }
#ifdef CUDNN
    cudnn_convolutional_setup(l);
#endif
#endif
    l->workspace_size = get_workspace_size(*l);
}

void add_bias(float *output, float *biases, int batch, int n, int size)
{
    int i,j,b;
    for(b = 0; b < batch; ++b){
        for(i = 0; i < n; ++i){
            for(j = 0; j < size; ++j){
                output[(b*n + i)*size + j] += biases[i];
            }
        }
    }
}

void scale_bias(float *output, float *scales, int batch, int n, int size)
{
    int i,j,b;
    for(b = 0; b < batch; ++b){
        for(i = 0; i < n; ++i){
            for(j = 0; j < size; ++j){
                output[(b*n + i)*size + j] *= scales[i];
            }
        }
    }
}

/*
** 功能： 計算每個卷積核的偏置更新值
** 輸入： bias_updates     當前層所有偏置的更新值，維度爲l.n
**       delta            當前層的誤差項
**       batch            一個batch含有的圖片張數
**       n                當前層卷積核個數
**       size             當前層輸入特徵圖尺寸
*/
void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
{
    int i,b;
	for(b = 0; b < batch; ++b){
		// 求和得一張輸入圖片的總偏置更新值
        for(i = 0; i < n; ++i){ 
            bias_updates[i] += sum_array(delta+size*(i+b*n), size);
        }
    }
}

// 卷積層前向傳播函數
void forward_convolutional_layer(convolutional_layer l, network net)
{
    int i, j;
    fill_cpu(l.outputs*l.batch, 0, l.output, 1);

    if(l.xnor){
        binarize_weights(l.weights, l.n, l.c/l.groups*l.size*l.size, l.binary_weights);
        swap_binary(&l);
        binarize_cpu(net.input, l.c*l.h*l.w*l.batch, l.binary_input);
        net.input = l.binary_input;
    }

    int m = l.n/l.groups; // 該層卷積核個數
    int k = l.size*l.size*l.c/l.groups; // 該層每個卷積核的參數元素個數
    int n = l.out_w*l.out_h; // 該層每個特徵圖的尺寸
    
    // 所有卷積覈對batch中每張圖片進行卷積運算
    for(i = 0; i < l.batch; ++i){
		// 將多通道二維圖像net.input變成按一定存儲規則排列的數組b，以便高效地進行卷積計算
        // 注意net.input包含batch中所有圖片的數據，但是im2col_cpu僅會處理其中一張圖片
        // im2col_cpu進行重排，l.c爲每張圖片的通道數，l.h爲每張圖片的高度，l.w爲每張圖片的寬度，l.size爲卷積核尺寸，l.stride爲步幅
        // 得到的b爲一張圖片重排後的結果，也是按行存儲的一維數組（共有l.c*l.size*l.size行，l.out_w*l.out_h列）
        for(j = 0; j < l.groups; ++j){
            float *a = l.weights + j*l.nweights/l.groups;
            float *b = net.workspace;
            float *c = l.output + (i*l.groups + j)*n*m;
            float *im =  net.input + (i*l.groups + j)*l.c/l.groups*l.h*l.w;

            if (l.size == 1) {
                b = im;
            } else {
                im2col_cpu(im, l.c/l.groups, l.h, l.w, l.size, l.stride, l.pad, b);
            }
            gemm(0,0,m,n,k,1,a,k,b,n,1,c,n); // 不轉置的矩陣運算，詳見具體函數
        }
    }

    if(l.batch_normalize){
        forward_batchnorm_layer(l, net);
    } else {
        add_bias(l.output, l.biases, l.batch, l.n, l.out_h*l.out_w);
    }

    activate_array(l.output, l.outputs*l.batch, l.activation);
    if(l.binary || l.xnor) swap_binary(&l);
}

// 卷積神經網絡反向傳播函數
void backward_convolutional_layer(convolutional_layer l, network net)
{
    int i, j;
    int m = l.n/l.groups; // 卷積核個數

    int n = l.size*l.size*l.c/l.groups;
    int k = l.out_w*l.out_h; // 每張輸出特徵圖的元素個數

    // 計算當前層激活函數對加權輸入的導數值並乘以l.delta相應元素，得到當前層的誤差項l.delta
    gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);

    if(l.batch_normalize){
        backward_batchnorm_layer(l, net);
    } else {
        // 計算偏置的更新值
        backward_bias(l.bias_updates, l.delta, l.batch, l.n, k);
    }

    for(i = 0; i < l.batch; ++i){
        for(j = 0; j < l.groups; ++j){
            float *a = l.delta + (i*l.groups + j)*m*k;
            float *b = net.workspace;
            float *c = l.weight_updates + j*l.nweights/l.groups;
            float *im  = net.input + (i*l.groups + j)*l.c/l.groups*l.h*l.w;
            float *imd = net.delta + (i*l.groups + j)*l.c/l.groups*l.h*l.w;

			// im2col_cpu()與gemm()是爲了計算當前層的權重更新值
            if(l.size == 1){
                b = im;
            } else {
                im2col_cpu(im, l.c/l.groups, l.h, l.w, 
                        l.size, l.stride, l.pad, b);
            }
            gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);

            if (net.delta) {
				// 當前層還未更新的權重
                a = l.weights + j*l.nweights/l.groups;
                b = l.delta + (i*l.groups + j)*m*k;
                c = net.workspace;
                if (l.size == 1) {
                    c = imd;
                }
	
                gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);

                if (l.size != 1) {
                    col2im_cpu(net.workspace, l.c/l.groups, l.h, l.w, l.size, l.stride, l.pad, imd);
                }
            }
        }
    }
}

// 卷積層更新函數
void update_convolutional_layer(convolutional_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.n, 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.n, momentum, l.bias_updates, 1);

    if(l.scales){
        axpy_cpu(l.n, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
        scal_cpu(l.n, momentum, l.scale_updates, 1);
    }
	//計算權重衰減
    axpy_cpu(l.nweights, -decay*batch, l.weights, 1, l.weight_updates, 1);
	//更新權重
    axpy_cpu(l.nweights, 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.nweights, momentum, l.weight_updates, 1);
}


image get_convolutional_weight(convolutional_layer l, int i)
{
    int h = l.size;
    int w = l.size;
    int c = l.c/l.groups;
    return float_to_image(w,h,c,l.weights+i*h*w*c);
}

void rgbgr_weights(convolutional_layer l)
{
    int i;
    for(i = 0; i < l.n; ++i){
        image im = get_convolutional_weight(l, i);
        if (im.c == 3) {
            rgbgr_image(im);
        }
    }
}

void rescale_weights(convolutional_layer l, float scale, float trans)
{
    int i;
    for(i = 0; i < l.n; ++i){
        image im = get_convolutional_weight(l, i);
        if (im.c == 3) {
            scale_image(im, scale);
            float sum = sum_array(im.data, im.w*im.h*im.c);
            l.biases[i] += sum*trans;
        }
    }
}

image *get_weights(convolutional_layer l)
{
    image *weights = calloc(l.n, sizeof(image));
    int i;
    for(i = 0; i < l.n; ++i){
        weights[i] = copy_image(get_convolutional_weight(l, i));
        normalize_image(weights[i]);
        /*
           char buff[256];
           sprintf(buff, "filter%d", i);
           save_image(weights[i], buff);
         */
    }
    //error("hey");
    return weights;
}

image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_weights)
{
    image *single_weights = get_weights(l);
    show_images(single_weights, l.n, window);

    image delta = get_convolutional_image(l);
    image dc = collapse_image_layers(delta, 1);
    char buff[256];
    sprintf(buff, "%s: Output", window);
    //show_image(dc, buff);
    //save_image(dc, buff);
    free_image(dc);
    return single_weights;
}