在ROS环境下用yolov3-tiny训练自己的数据集

由于整个过程新建、修改、设置的文件和参数较多,所以我先放张整个流程分析图在这,方便大家结合下列步骤去更好理解。
ros训练yolov3的流程分析图

0.提前准备:

  • 下载darknet(yolov3)并编译:
git clone https://gitee.com/bingda-robot/darknet.git
make

如果编译的时候报以下错:

./src/convolutional_layer.c:153:13: error: ‘CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT’ undeclared (first use in this function)
  153 |             CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,

就修改./src/convolutional_layer.c文件为:

#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>

#define PRINT_CUDNN_ALGO 0
#define MEMORY_LIMIT 2000000000

#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;
        }
    }
}

int convolutional_out_height(convolutional_layer l)
{
    return (l.h + 2*l.pad - l.size) / l.stride + 1;
}

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
    #if CUDNN_MAJOR >= 8
    int returnedAlgoCount;
    cudnnConvolutionFwdAlgoPerf_t       fw_results[2 * CUDNN_CONVOLUTION_FWD_ALGO_COUNT];
    cudnnConvolutionBwdDataAlgoPerf_t   bd_results[2 * CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT];
    cudnnConvolutionBwdFilterAlgoPerf_t bf_results[2 * CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT];

    cudnnFindConvolutionForwardAlgorithm(cudnn_handle(),
            l->srcTensorDesc,
            l->weightDesc,
            l->convDesc,
            l->dstTensorDesc,
            CUDNN_CONVOLUTION_FWD_ALGO_COUNT,
            &returnedAlgoCount,
        fw_results);
    for(int algoIndex = 0; algoIndex < returnedAlgoCount; ++algoIndex){
        #if PRINT_CUDNN_ALGO > 0
        printf("^^^^ %s for Algo %d: %f time requiring %llu memory\n",
               cudnnGetErrorString(fw_results[algoIndex].status),
               fw_results[algoIndex].algo, fw_results[algoIndex].time,
               (unsigned long long)fw_results[algoIndex].memory);
        #endif
        if( fw_results[algoIndex].memory < MEMORY_LIMIT ){
            l->fw_algo = fw_results[algoIndex].algo;
            break;
    }
    }

    cudnnFindConvolutionBackwardDataAlgorithm(cudnn_handle(),
            l->weightDesc,
            l->ddstTensorDesc,
            l->convDesc,
            l->dsrcTensorDesc,
            CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT,
            &returnedAlgoCount,
            bd_results);
    for(int algoIndex = 0; algoIndex < returnedAlgoCount; ++algoIndex){
        #if PRINT_CUDNN_ALGO > 0
        printf("^^^^ %s for Algo %d: %f time requiring %llu memory\n",
               cudnnGetErrorString(bd_results[algoIndex].status),
               bd_results[algoIndex].algo, bd_results[algoIndex].time,
               (unsigned long long)bd_results[algoIndex].memory);
        #endif
        if( bd_results[algoIndex].memory < MEMORY_LIMIT ){
            l->bd_algo = bd_results[algoIndex].algo;
            break;
        }
    }

    cudnnFindConvolutionBackwardFilterAlgorithm(cudnn_handle(),
            l->srcTensorDesc,
            l->ddstTensorDesc,
            l->convDesc,
            l->dweightDesc,
            CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT,
            &returnedAlgoCount,
            bf_results);
    for(int algoIndex = 0; algoIndex < returnedAlgoCount; ++algoIndex){
        #if PRINT_CUDNN_ALGO > 0
        printf("^^^^ %s for Algo %d: %f time requiring %llu memory\n",
               cudnnGetErrorString(bf_results[algoIndex].status),
               bf_results[algoIndex].algo, bf_results[algoIndex].time,
               (unsigned long long)bf_results[algoIndex].memory);
        #endif
        if( bf_results[algoIndex].memory < MEMORY_LIMIT ){
            l->bf_algo = bf_results[algoIndex].algo;
            break;
        }
    }

    #else

    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
#endif

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;

    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));

    l.nweights = c/groups*n*size*size;
    l.nbiases = n;

    // float scale = 1./sqrt(size*size*c);
    float scale = sqrt(2./(size*size*c/l.groups));
    //printf("convscale %f\n", scale);
    //scale = .02;
    //for(i = 0; i < c*n*size*size; ++i) l.weights[i] = scale*rand_uniform(-1, 1);
    for(i = 0; i < l.nweights; ++i) l.weights[i] = scale*rand_normal();
    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;
    l.outputs = l.out_h * l.out_w * l.out_c;
    l.inputs = l.w * l.h * l.c;

    l.output = calloc(l.batch*l.outputs, sizeof(float));
    l.delta  = calloc(l.batch*l.outputs, sizeof(float));

    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 test_convolutional_layer()
{
    convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 0, 0, 0);
    l.batch_normalize = 1;
    float data[] = {1,1,1,1,1,
        1,1,1,1,1,
        1,1,1,1,1,
        1,1,1,1,1,
        1,1,1,1,1,
        2,2,2,2,2,
        2,2,2,2,2,
        2,2,2,2,2,
        2,2,2,2,2,
        2,2,2,2,2,
        3,3,3,3,3,
        3,3,3,3,3,
        3,3,3,3,3,
        3,3,3,3,3,
        3,3,3,3,3};
    //net.input = data;
    //forward_convolutional_layer(l);
}
*/

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];
            }
        }
    }
}

void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
{
    int i,b;
    for(b = 0; b < batch; ++b){
        for(i = 0; i < n; ++i){
            bias_updates[i] += sum_array(delta+size*(i+b*n), size);
        }
    }
}

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;
    int k = l.size*l.size*l.c/l.groups;
    int n = l.out_w*l.out_h;
    for(i = 0; i < l.batch; ++i){
        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);
        }
    }

    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);
}

void backward_convolutional_layer(convolutional_layer l, network net)
{
    int i, j;
    int m = l.n/l.groups;
    int n = l.size*l.size*l.c/l.groups;
    int k = l.out_w*l.out_h;

    gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);

    if(l.batch_normalize){
        backward_batchnorm_layer(l, net);
    } else {
        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;

            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) {
                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);
                }
            }
        }
    }
}

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;

    axpy_cpu(l.n, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
    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);
    }

    axpy_cpu(l.nweights, -decay*batch, l.weights, 1, l.weight_updates, 1);
    axpy_cpu(l.nweights, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
    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;
}
  • 数据集准备
    标准数据集的软件叫做labelimg,该软件可以标注图片数据集,每张图片标注完成后会生成对应的xml文件,软件下载和标注教程可以网上百度。如果大家懒得标记可以去网上下载数据集。反正最后需要以下两份数据集文件。
    数据集文件准备

1.创建myData文件夹

在darknet目录下创建myData文件夹,目录结构如下。将自己标注好的数据集图片拷贝到JPEGImages目录下,将数据集lxml文件拷贝到Annotations目录下。ImageSets/Main文件先建好,后续会用。

myData
  ...JPEGImages#存放图像
 
  ...Annotations#存放图像对应的xml文件
 
  ...ImageSets/Main # 存放训练/验证图像的名字

2.创建test.py文件

在myData文件下创建test.py,将下面代码拷贝进去,虚拟机终端myData下输入python test.py运行该文件,将在ImageSets/Main文件下生成四个文件:train.txt,val.txt,test.txt和trainval.txt。
在这里插入图片描述

import os
import random

trainval_percent = 0.1
train_percent = 0.9
xmlfilepath = 'Annotations'
txtsavepath = 'ImageSets\Main'
total_xml = os.listdir(xmlfilepath)
num = len(total_xml)
list = range(num)
tv = int(num * trainval_percent)
tr = int(tv * train_percent)
trainval = random.sample(list, tv)
train = random.sample(trainval, tr)

ftrainval = open('ImageSets/Main/trainval.txt', 'w')
ftest = open('ImageSets/Main/test.txt', 'w')
ftrain = open('ImageSets/Main/train.txt', 'w')
fval = open('ImageSets/Main/val.txt', 'w')

for i in list:
    name = total_xml[i][:-4] + '\n'
    if i in trainval:
        ftrainval.write(name)
        if i in train:
            ftest.write(name)
        else:
            fval.write(name)
    else:
        ftrain.write(name)

ftrainval.close()
ftrain.close()
fval.close()
ftest.close()

3.创建weights文件

在maData文件下创建weights文件,后续训练的权重文件会保存至此。
在这里插入图片描述

4.创建myData.names文件

在myData文件下创建myData.names文件,在该文件下添加数据集标注的种类,每一个种类占一行,我这里为了简单只标注了car。
大家如果标注种类多就换行继续添加就行,如下图所示。
标注种类多

5.将数据转换成darknet支持的格式

yolov3提供了将VOC数据集转为yolo训练所需要的格式的代码,在scripts/voc_label.py文件中。在darknet文件夹下新建一个my_lables.py文件,将voc_label.py内容复制到my_lables.py中,将classes内的内容修改为前面标注时的种类,修改后运行该程序。
修改内容:

classes = ["car"] 

完整代码:

# coding=gbk
import xml.etree.ElementTree as ET
import pickle
import os
from os import listdir, getcwd
from os.path import join
 
sets=[('myData', 'train')] 
 
classes = ["car"] 
 
def convert(size, box):
    dw = 1./(size[0])
    dh = 1./(size[1])
    x = (box[0] + box[1])/2.0 - 1
    y = (box[2] + box[3])/2.0 - 1
    w = box[1] - box[0]
    h = box[3] - box[2]
    x = x*dw
    w = w*dw
    y = y*dh
    h = h*dh
    return (x,y,w,h)
 
def convert_annotation(year, image_id):
    in_file = open('myData/Annotations/%s.xml'%(image_id)) 
    out_file = open('myData/labels/%s.txt'%(image_id), 'w')  
    tree=ET.parse(in_file)
    root = tree.getroot()
    size = root.find('size')
    w = int(size.find('width').text)
    h = int(size.find('height').text)
 
    for obj in root.iter('object'):
        difficult = obj.find('difficult').text
        cls = obj.find('name').text
        if cls not in classes or int(difficult)==1:
            continue
        cls_id = classes.index(cls)
        xmlbox = obj.find('bndbox')
        b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text), float(xmlbox.find('ymax').text))
        bb = convert((w,h), b)
        out_file.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')
 
wd = getcwd()
 
for year, image_set in sets:
    if not os.path.exists('myData/labels/'): 
        os.makedirs('myData/labels/')
    image_ids = open('myData/ImageSets/Main/%s.txt'%(image_set)).read().strip().split()
    list_file = open('myData/%s_%s.txt'%(year, image_set), 'w')
    for image_id in image_ids:
        list_file.write('%s/myData/JPEGImages/%s.jpg\n'%(wd, image_id))
        convert_annotation(year, image_id)
    list_file.close()

运行代码:

cd darknet
python my_lables.py

成功运行代码后会在myData目录下生成一个labels文件夹(里面是txt文件)和一个myData_train.txt文件。
在这里插入图片描述lables文件中的‘txt文件的含义为:类别的编码和目标的相对位置。
在这里插入图片描述

6.创建my_data.data文件

darknet/cfg下新建my_data.data,内容如下(可参考cfg/voc.data):

classes= 1 ##改为自己的分类个数
##下面都改为自己的路径
train  = /home/bingda/darknet/myData/myData_train.txt  
names = /home/bingda/darknet/myData/myData.names 
backup = /home/bingda/darknet/myData/weights

7.创建my_yolov3-tiny.cfg文件

darknet/cfg下新建my_yolov3-tiny.cfg,内容如下(可参考cfg/yolov3-tiny.cfg):

  • 搜索yolo关键词,找到两个地方,将每个[yolo]行上面的filters的值改为(种类数+5)*3=(1+5)*3=18,下面的classes改为种类数1,总共要改两个filters和两个classes,大家根据自己的种类数修改自己的两个filters和两个classes。

  • 可修改:random = 0:原来是1,显存小的话就改为0。

  • 然后找到文件最上面,将# Training 注释放开,使其为训练模式。然后修改batch=1,subdivisions=1。
    在这里插入图片描述在这里插入图片描述

8.下载预训练权重

cd darknet
wget https://pjreddie.com/media/files/darknet53.conv.74

9开始训练

cd darknet
./darknet detector train cfg/my_data.data cfg/my_yolov3-tiny.cfg darknet53.conv.74

在这里插入图片描述

  • 3
    点赞
  • 21
    收藏
    觉得还不错? 一键收藏
  • 5
    评论
告知:需要学习YOLOv4进行TT100K数据集上中国交通标志识别的学员请前往(1) Ubuntu系统《YOLOv4目标检测实战:中国交通标志识别》课程链接:https://edu.csdn.net/course/detail/29362(2)《Windows版YOLOv4目标检测实战:中国交通标志识别》课程链接:https://edu.csdn.net/course/detail/29363在无人驾驶中,交通标志识别是一项重要的任务。本课程中的项目以美国交通标志数据集LISA为训练对象,采用YOLOv3目标检测方法实现实时交通标志识别。具体项目过程包括包括:安装Darknet、下载LISA交通标志数据集数据集格式转换、修改配置文件、训练LISA数据集、测试训练出的网络模型、性能统计(mAP计算和画出PR曲线)和先验框聚类。YOLOv3基于深度学习,可以实时地进行端到端的目标检测,以速度快见长。本课程将手把手地教大家使用YOLOv3实现交通标志的多目标检测。本课程的YOLOv3使用Darknet,在Ubuntu系统上做项目演示。 Darknet是使用C语言实现的轻型开源深度学习框架,依赖少,可移植性好,值得深入学习和探究。除本课程《YOLOv3目标检测实战:交通标志识别》外,本人推出了有关YOLOv3目标检测的系列课程,请持续关注该系列的其它课程视频,包括:《YOLOv3目标检测实战:训练自己的数据集》《YOLOv3目标检测:原理与源码解析》《YOLOv3目标检测:网络模型改进方法》另一门课程《YOLOv3目标检测实战:训练自己的数据集》主要是介绍如何训练自己标注的数据集。而本课程的区别主要在于学习对已标注数据集的格式转换,即把LISA数据集从csv格式转换成YOLOv3所需要的PASCAL VOC格式和YOLO格式。本课程提供数据集格式转换的Python代码。请大家关注以上课程,并选择学习。下图是使用YOLOv3进行交通标志识别的测试结果
评论 5
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值