YOLOv3：Darknet代码解析（三）卷积操作

目的：读懂Yolo_V3中的卷积操作

1.卷积内容

卷积层中核心代码，前馈的卷积层，反向传播的卷积层。

void forward_convolutional_layer(convolutional_layer l, network_state state)
{
    int out_h = convolutional_out_height(l);
    int out_w = convolutional_out_width(l);
    int i;

    fill_cpu(l.outputs*l.batch, 0, l.output, 1);

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

    int m = l.n;
    int k = l.size*l.size*l.c;
    int n = out_h*out_w;
    printf("Layer number - %d\n", l.layer_num);
    int hw = 0;

    if (l.layer_num == 0) hw = 1;
    else hw = 0;

    printf("\nValue of m (l.n) = %d, k (l.size*l.size*l.c) = %d, n (out_h*out_w) = %d, l.c = %d\n", m,k,n,l.c);
    if (l.xnor && l.c%32 == 0 && AI2) {
        forward_xnor_layer(l, state);
        printf("xnor\n");
    } else {

        float *a = l.weights;
        float *b = state.workspace;
        float *c = l.output;

        for(i = 0; i < l.batch; ++i){
            im2col_cpu(state.input, l.c, l.h, l.w, 
                    l.size, l.stride, l.pad, b);
            if(hw)
              gemm_cpu_hw(0,0,m,n,k,1,a,k,b,n,1,c,n);
            else
              gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
 
            c += n*m;
            state.input += l.c*l.h*l.w;
        }
    }

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

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

其中，核心语句为  gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);

2.gemm

为卷积操作中的矩阵乘法。在gemm.c中

void gemm(int TA, int TB, int M, int N, int K, float ALPHA, 
        float *A, int lda, 
        float *B, int ldb,
        float BETA,
        float *C, int ldc)
{
    gemm_cpu( TA,  TB,  M, N, K, ALPHA,A,lda, B, ldb,BETA,C,ldc);
}

void gemm_cpu(int TA, int TB, int M, int N, int K, float ALPHA, 
        float *A, int lda, 
        float *B, int ldb,
        float BETA,
        float *C, int ldc)
{
    //printf("cpu: %d %d %d %d %d %f %d %d %f %d\n",TA, TB, M, N, K, ALPHA, lda, ldb, BETA, ldc);
    int i, j;
    for(i = 0; i < M; ++i){
        for(j = 0; j < N; ++j){
            C[i*ldc + j] *= BETA;
        }
    }
    if(!TA && !TB)
        gemm_nn(M, N, K, ALPHA,A,lda, B, ldb,C,ldc);
    else if(TA && !TB)
        gemm_tn(M, N, K, ALPHA,A,lda, B, ldb,C,ldc);
    else if(!TA && TB)
        gemm_nt(M, N, K, ALPHA,A,lda, B, ldb,C,ldc);
    else
        gemm_tt(M, N, K, ALPHA,A,lda, B, ldb,C,ldc);
}

前两个TA，TB决定了后面运算函数为gemm_nn还是gemm_tn，gemm_nt，gemm_tt。输入层中输入TA，TB都为0，因此用到的函数为gemm_nn，

卷积的语句： gemm_nn

void gemm_nn(int M, int N, int K, float ALPHA, 
        float *A, int lda, 
        float *B, int ldb,
        float *C, int ldc)
{
    printf ("GEMM NN SW\n");
    int i,j,k;
    for(i = 0; i < M; ++i){
        for(k = 0; k < K; ++k){
            register float A_PART = ALPHA*A[i*lda+k];
            for(j = 0; j < N; ++j){
                C[i*ldc+j] += A_PART*B[k*ldb+j];
            }
        }
    }
}

A,B,C都为指针，在forward_convolutional_layer.c中，A为权重，B为图像值，C为输出

float *a = l.weights;
float *b = state.workspace;
float *c = l.output;

M，N，K为输出的featureMap的坐标，对应于下面的m，n，k。lda，ldb，ldc对应于下面的K，N，N

    int M = l.n;
    int K = l.size*l.size*l.c;
    int N = out_h*out_w;

整个卷积过程有两个重要步骤，一个是M，K，N的值的大小，另一个是指针a，b，c指向的位置。

核心语句为三个for循环里面，

C[i*ldc+j] += A[i*lda+k]  *  B[k*ldb+j];

即输入C[ i*N + j ]+=A[ i*K +k ]  *  B[k*N + j ]

权重M*K，输入 K*N， 输出 M*N

3.其他内容

运行YOLO时的命令行为：./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg

我们需要找到m,k,n的大小，和a，b，c指针的位置。

darknet.c
int main(int argc, char **argv)
{
    if(argc < 2){
        fprintf(stderr, "usage: %s <function>\n", argv[0]);
        return 0;
    }
    gpu_index = find_int_arg(argc, argv, "-i", 0);
    if(find_arg(argc, argv, "-nogpu")) {
        gpu_index = -1;
    }

#ifndef GPU
    gpu_index = -1;
#else
    if(gpu_index >= 0){
        cuda_set_device(gpu_index);
    }
#endif

    if (0 == strcmp(argv[1], "average")){
        average(argc, argv);
    } else if (0 == strcmp(argv[1], "detect")){
        float thresh = find_float_arg(argc, argv, "-thresh", .25);
        char *filename = (argc > 4) ? argv[4]: 0;
        test_detector("cfg/coco.data", argv[2], argv[3], filename, thresh);
    } else {
        fprintf(stderr, "Not an option: %s\n", argv[1]);
    }
    return 0;
}

主函数中test_detector子函数中，argv[2]为yolov3-tiny.cfg,  argv[3]为yolov3-tiny.weights

//detector.c
void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh)
{
    list *options = read_data_cfg(datacfg);
    char *name_list = option_find_str(options, "names", "data/names.list");
    char **names = get_labels(name_list);

    image **alphabet = load_alphabet();
    network net = parse_network_cfg(cfgfile);
    if(weightfile){
        load_weights(&net, weightfile);
    }
    layer l = net.layers[net.n-1];
    set_batch_network(&net, 1);
    srand(2222222);
    clock_t time;
    char buff[256];
    char *input = buff;
    int j;
    float nms=.4;
    box *boxes = calloc(l.w*l.h*l.n, sizeof(box));
    float **probs = calloc(l.w*l.h*l.n, sizeof(float *));
    for(j = 0; j < l.w*l.h*l.n; ++j) probs[j] = calloc(l.classes, sizeof(float *));
    while(1){
        if(filename){
            strncpy(input, filename, 256);
        } else {
            printf("Enter Image Path: ");
            fflush(stdout);
            input = fgets(input, 256, stdin);
            if(!input) return;
            strtok(input, "\n");
        }
        image im = load_image_color(input,0,0);
        image sized = resize_image(im, net.w, net.h);
        float *X = sized.data;
        time=clock();
        network_predict(net, X);
        printf("%s: Predicted in %f seconds.\n", input, sec(clock()-time));
        get_region_boxes(l, 1, 1, thresh, probs, boxes, 0);
        if (nms) do_nms_sort(boxes, probs, l.w*l.h*l.n, l.classes, nms);
        draw_detections(im, l.w*l.h*l.n, thresh, boxes, probs, names, alphabet, l.classes);
        save_image(im, "predictions");
        show_image(im, "predictions");

        free_image(im);
        free_image(sized);

        if (filename) break;
    }
}

函数中，确定网络的语句：

    network net = parse_network_cfg(cfgfile);
    if(weightfile){
        load_weights(&net, weightfile);
    }
    layer l = net.layers[net.n-1];
    set_batch_network(&net, 1);

函数中重要的为parse_network_cfg,此函数为根据相应的cfg文件创建神经网络，根据相应的load_weights加载相应的参数。

//parser.c
network parse_network_cfg(char *filename)
{
    list *sections = read_cfg(filename);
    node *n = sections->front;
    if(!n) error("Config file has no sections");
    network net = make_network(sections->size - 1);
    net.gpu_index = gpu_index;
    size_params params;

    section *s = (section *)n->val;
    list *options = s->options;
    if(!is_network(s)) error("First section must be [net] or [network]");
    parse_net_options(options, &net);

    params.h = net.h;
    params.w = net.w;
    params.c = net.c;
    params.inputs = net.inputs;
    params.batch = net.batch;
    params.time_steps = net.time_steps;
    params.net = net;

    size_t workspace_size = 0;
    n = n->next;
    int count = 0;
    free_section(s);
    fprintf(stderr, "layer     filters    size              input                output\n");
    while(n){
        params.index = count;
        fprintf(stderr, "%5d ", count);
        s = (section *)n->val;
        options = s->options;
        layer l = {0};
        LAYER_TYPE lt = string_to_layer_type(s->type);
        if(lt == CONVOLUTIONAL){
            l = parse_convolutional(options, params);
        }else if(lt == LOCAL){
            l = parse_local(options, params);
        }else if(lt == DROPOUT){
            l = parse_dropout(options, params);
            l.output = net.layers[count-1].output;
            l.delta = net.layers[count-1].delta;
        }else{
            fprintf(stderr, "Type not recognized: %s\n", s->type);
        }
        l.dontload = option_find_int_quiet(options, "dontload", 0);
        l.dontloadscales = option_find_int_quiet(options, "dontloadscales", 0);
        option_unused(options);
        net.layers[count] = l;
        if (l.workspace_size > workspace_size) workspace_size = l.workspace_size;
        free_section(s);
        n = n->next;
        ++count;
        if(n){
            params.h = l.out_h;
            params.w = l.out_w;
            params.c = l.out_c;
            params.inputs = l.outputs;
        }
    }   
    free_list(sections);
    net.outputs = get_network_output_size(net);
    net.output = get_network_output(net);
    if(workspace_size){
        net.workspace = calloc(1, workspace_size);
    }
    return net;
}

//parser.c
convolutional_layer parse_convolutional(list *options, size_params params)
{
    int n = option_find_int(options, "filters",1);
    int size = option_find_int(options, "size",1);
    int stride = option_find_int(options, "stride",1);
    int pad = option_find_int_quiet(options, "pad",0);
    int padding = option_find_int_quiet(options, "padding",0);
    if(pad) padding = size/2;

    char *activation_s = option_find_str(options, "activation", "logistic");
    ACTIVATION activation = get_activation(activation_s);

    int batch,h,w,c;
    h = params.h;
    w = params.w;
    c = params.c;
    batch=params.batch;
    if(!(h && w && c)) error("Layer before convolutional layer must output image.");
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
    int binary = option_find_int_quiet(options, "binary", 0);
    int xnor = option_find_int_quiet(options, "xnor", 0);

    convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,padding,activation, batch_normalize, binary, xnor, params.net.adam);
    layer.flipped = option_find_int_quiet(options, "flipped", 0);
    layer.dot = option_find_float_quiet(options, "dot", 0);
    if(params.net.adam){
        layer.B1 = params.net.B1;
        layer.B2 = params.net.B2;
        layer.eps = params.net.eps;
    }

    return layer;
}