Darknet YOLOV3 模型训练
没想到YOLOV3我还能用得到,这个模型非常的经典,这里先不去介绍它的理论,只记录下这个经典模型的训练过程!!
Darknet YOLOV3的地址
Darknet yolov3作者的网站: https://pjreddie.com/darknet/yolo/
Darknet yolov3官方权重的下载地址:https://pjreddie.com/media/files/yolov3.weights
Darknet yolov3官方源代码:https://github.com/pjreddie/darknet
环境准备
在https://www.autodl.com/market/list,租用的GPU服务器的显卡为2080TI,镜像随便选择,使用VsCode SSH链接进行操作最简便;
先下载编译安装darknet工程,详见官网https://pjreddie.com/darknet/yolo/
编译安装darkenet工程
git clone https://github.com/pjreddie/darknet
cd darknet
修改darknet/Makefile文件,使用GPU进行训练(也可以不修改,训练的时候用参数指定)
修改完成后进行编译 注意:如果没有没有修改Makefile就编译为cpu模式, 此时可以使用make clean 清楚之前的编译记录
make
环境配置可能会出现的问题
- 问题一:
如果出现下面的情况:
./src/convolutional_layer.c:153:13: error: ‘CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT’ undeclared (first use in this function); did you mean ‘CUDNN_CONVOLUTION_FWD_ALGO_DIRECT’?CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~CUDNN_CONVOLUTION_FWD_ALGO_DIRECT
compilation terminated due to -Wfatal-errors.
Makefile:89: recipe for target 'obj/convolutional_layer.o' failed
make: *** [obj/convolutional_layer.o] Error 1
(参考:https://blog.csdn.net/LHX19971114/article/details/126229887)要使用https://github.com/arnoldfychen/darknet/tree/master/src代码中的convolutional_layer.c文件替换原作者https://github.com/pjreddie/darknet/tree/master/src代码中的convolutional_layer.c。(Darknet使用的是pjreddie版本)
原因:现在的cuda版本(11以上)对应的cudnn(8.2以上)里没有CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT这个宏定义。NVIDIA给出了一个针对cudnn8的解决方案代码,就是修改出错的文件src/convolutional_layer.c的代码,增加针对CUDNN_MAJOR>=8的处理:
convolutional_layer.c的代码如下:
#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"
#endifvoid swap_binary(convolutional_layer *l)
{float *swap = l->weights;l->weights = l->binary_weights;l->binary_weights = swap;#ifdef GPUswap = 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 CUDNNif(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;}
#endifreturn (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 >= 6cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT);#elsecudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);#endif#if CUDNN_MAJOR >= 7cudnnSetConvolutionGroupCount(l->convDesc, l->groups);#elseif(l->groups > 1){error("CUDNN < 7 doesn't support groups, please upgrade!");}#endif#if CUDNN_MAJOR >= 8int 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 > 0printf("^^^^ %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);#endifif( 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 > 0printf("^^^^ %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);#endifif( 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 > 0printf("^^^^ %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);#endifif( bf_results[algoIndex].memory < MEMORY_LIMIT ){l->bf_algo = bf_results[algoIndex].algo;break;}}#elsecudnnGetConvolutionForwardAlgorithm(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
#endifconvolutional_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 GPUl.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 CUDNNcudnnCreateTensorDescriptor(&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}
#endifl.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 GPUcuda_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 CUDNNcudnn_convolutional_setup(l);
#endif
#endifl->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;
}
- 问题二
如果出现下面这个问题(参考https://blog.csdn.net/ttjbmkjsjyyjbbyg/article/details/114801686)
tional_kernels.cu -o obj/convolutional_kernels.o
nvcc fatal : Unsupported gpu architecture 'compute_30'
Makefile:93: recipe for target 'obj/convolutional_kernels.o' failed
make: *** [obj/convolutional_kernels.o] Error 1
原因:这是不支持compute_30的gpu构架,这是由于GPU太新,与CUDA版本不兼容导致
解决: 找到darknet/makefile的下图处,删除 -gencode arch=compute_30,code=sm_30
编译完成:
数据准备
准备一个文件夹,命名为my_data,内容如下:
拷贝到darknet文件下面,其中train_images目录和validate_images目录下存放的是我们自己要训练的图像数据
trainImageId.txt和validateImageId.txt文件为遍历train_images目录和validate_images目录下所有图片
trainImagePath.txt和validateImagePath.txt为图片路径文件
trainImageId.txt和validateImageId.txt以及trainImagePath.txt和validateImagePath.txt可以使用下面的简单脚本生成
import os # 指定文件夹路径
folder_path = 'D:/Users/XueLi_G/Desktop/工地检测/my_data/train_images' # 获取文件夹下的所有文件名
files = os.listdir(folder_path) # # 打印所有文件名
with open("D:/Users/XueLi_G/Desktop/工地检测/my_data/trainImageId.txt", 'w') as f: for file in files: print(file)# 用于生成trainImagePath.txt和valImagePath.txt# f.write("/root/darknet/my_data/val_images/" + file + '\n')# 用于生成trainImageId.txt和valImageId.txtf.write(file + '\n')
权重和网络配置文件准备
进入my_data目录
- 下载权重数据,例如以yolov3-tiny.weights为例子:
cd my_data
wget https://pjreddie.com/media/files/yolov3-tiny.weights
- 下载网络配置文件
进入https://github.com/pjreddie/darknet/blob/master/cfg/yolov3-tiny.cfg下载配置文件到my_data目录
修改配置文件
在my_data目录下,创建voc.names 和voc.data
- 修改voc.names文件,写入数据集中的所有类别名称
vi voc.namesclass_1class_2class_3
- 修改voc.data文件,N是你训练的物体种类,train和valid为图片路径文件,names为数据集的类别名文件,backup为生成的结果文件
vi voc.dataclasses = N # N为类别的数量train = /home/XXX/darknet/train_my_data/trainImagePath.txtvalid = /home/XXX/darknet/train_my_data/validateImagePath.txtnames = train_my_data/voc.namesbackup = train_my_data/backup
- 修改yolov3-tiny.cfg,修改所有yolo层的classes参数,修改yolo层前一层的filters参数,修改max_batches迭代次数
vi yolov3-tiny.cfg修改所有yolo层的classes参数 为数据集的类别数修改yolo层前一层的filters参数 为3*(N+1+4),其中N为数据集的类别数修改max_batches 为你想要的迭代次数
注释掉Testing 模式
# Testing ### 测试模式
# batch=1
# subdivisions=1
cfg参数的解读:
[net]
# Testing ### 测试模式
# batch=1
# subdivisions=1# Training ### 训练模式,每次前向的图片数目 = batch/subdivisions
batch=64 ### 训练的一批数据的个数,batch越大,训练效果越好
subdivisions=16 ### 在训练模型过程中,一批batch样本又被平均分成subdivision次送入网络参与训练,以减轻内存占用的压力,subdivision越大,占用内存压力越小
width=416 ### 网络的输入宽、高、通道数(可以直接修改,但是要修改网络避免深层网络的浪费)
height=416 ### 宽高比的比例可依据输入图像的尺寸进行设置(但需要为32的倍数)
channels=3 ### 训练数据的通道数 彩色图
momentum=0.9 ### 动量: DeepLearning1中最优化方法中的动量参数,这个值影响着梯度下降到最优值得速度。 详情参考:https://blog.csdn.net/qq_33270279/article/details/102796812
decay=0.0005 ### 权重衰减:权重衰减正则项,防止过拟合.每一次学习的过程中,将学习后的参数按照固定比例进行降低,为了防止过拟合,decay参数越大对过拟合的抑制能力越强。
angle=0 ###旋转角度(单位:度)
saturation = 1.5 ### 饱和度
exposure = 1.5 ### 曝光度
hue=.1 ### 色调
learning_rate=0.001 ### 学习率的调整参考 :https://blog.csdn.net/qq_33485434/article/details/80452941
burn_in=1000 ### 模型预热,小于1000batch时采用(0-learning_rate)递增的方式。
max_batches = 50200 ### 迭代次数
policy=steps ### 学习率策略: 这个是学习率调整的策略,有policy:constant, steps, exp, poly, step, sig, RANDOM,constant等方式参考[https://nanfei.ink/2018/01/23/YOLOv2%E8%B0%83%E5%8F%82%E6%80%BB%E7%BB%93/#more](https://nanfei.ink/2018/01/23/YOLOv2调参总结/#more)
steps=40000,45000 ### 学习率变动步长
scales=.1,.1 ### 学习率变动因子 [convolutional]
batch_normalize=1 ### BN
filters=32 ### 卷积核数目
size=3 ### 卷积核尺寸
stride=1 ### 卷积核步长
pad=1 ### pad
activation=leaky ###激活函数的类型:linear(线性:不做改变)relu(值>0时保持不变,小于零时置零)leaky(值>0时保持不变,小于零时*0.1)......[convolutional]
size=1
stride=1
pad=1
filters=45 #每一个[region/yolo]层前的最后一个卷积层中的 filters=(classes+cords+1)*anchors_num,其中anchors_num 是该层mask的数量。如果没有maskanchors_num=num。coords+1:论文中的tx,ty,tw,th,to
activation=linear[yolo]
mask = 6,7,8 #预测anchors[]中的后三个较大尺寸的anchor。因为其感受野较大
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 #(width, height)通过聚类脚本计算样本宽高
classes=2 #类别
num=9 #如果在配置文件中anchors的数量大于num时,仅使用前num个,小于时内存越界。
jitter=.3 #利用数据抖动产生更多数据, jitter就是crop的参数, jitter=.3,就是在0~0.3中进行crop(具体的裁剪方式请参考src/data.c中的load_data_detection()函数)
ignore_thresh = .5 #参数解释:ignore_thresh 指得是参与计算的IOU阈值大小。当预测的检测框与ground true的IOU大于ignore_thresh的时候,参与loss的计算,否则,检测框的不参与损失计算。参数目的和理解:目的是控制参与loss计算的检测框的规模,当ignore_thresh过于大,接近于1的时候,那么参与检测框回归loss的个数就会比较少,同时也容易造成过拟合;而如果ignore_thresh设置的过于小,那么参与计算的会数量规模就会很大。同时也容易在进行检测框回归的时候造成欠拟合。参数设置:一般选取0.5-0.7之间的一个值,之前的计算基础都是小尺(13*13)用的是0.7,(26*26)用的是0.5。这次先将0.5更改为0.7。 实验结果:AP=0.5121(有明显下降)
truth_thresh = 1
random=0 #1,如果显存很小,将random设置为0,关闭多尺度训练;......[yolo]
mask = 3,4,5 #预测anchors[]中的中间三个较小尺寸的anchor。因为其感受野适中
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326 # 三个尺寸的预设的anchor大小
classes=2 # 数据中的类别数量
num=9
jitter=.3
ignore_thresh = .5
truth_thresh = 1
random=0 ### 每隔几次迭代后就会微调网络的输入尺寸,如果为1,每次迭代图片大小随机从320到608,步长为32,如果为0,每次训练大小与输入大小......[yolo]
mask = 0,1,2 #预测anchors[]中的前三个较小尺寸的anchor。因为其感受野较小
anchors = 10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326
classes=2
num=9
jitter=.3
ignore_thresh = .5
truth_thresh = 1
random=0
参考:https://blog.csdn.net/qq_43211132/article/details/88679979?ops_request_misc=%257B%2522request%255Fid%2522%253A%252297B38269-B787-4CDE-96D0-900AD54CCB05%2522%252C%2522scm%2522%253A%252220140713.130102334…%2522%257D&request_id=97B38269-B787-4CDE-96D0-900AD54CCB05&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2allsobaiduend~default-2-88679979-null-null.142v100pc_search_result_base1&utm_term=yolov3%E7%9A%84subdivisions&spm=1018.2226.3001.4187
模型训练
提取卷积层数据
./darknet partial my_data/yolov3-tiny.cfg my_data/yolov3-tiny.weights my_data/yolov3-tiny.conv.15 15
开始训练
./darknet detector train my_data/voc.data my_data/yolov3-tiny.cfg my_data/yolov3-tiny.conv.15
参数含义:
- Region xx: cfg文件中yolo-layer的索引;
- Avg IOU:当前迭代中,预测的box与标注的box的平均交并比,越大越好,期望数值为1;
- Class: 标注物体的分类准确率,越大越好,期望数值为1;
- obj: 越大越好,期望数值为1;
- No obj: 越小越好;
- .5R: 以IOU=0.5为阈值时候的recall; recall = 检出的正样本/实际的正样本
- 0.75R: 以IOU=0.75为阈值时候的recall;
- count:正样本数目。
- 1:表示当前迭代的次数
- 280.736786: 是总体的 Loss(损失);
-
- 2550.5avg: 是平均 Loss, 这个数值应该越低越好, 一般来说, 一旦这个数值低于 0.060730 avg 就可以终止训练了;
- 0.000000 rate: 代表当前的学习率, 是在.cfg文件中定义的;
- 0.158851seconds: 表示当前批次训练花费的总时间;
- 1408 images: 这个数值是 表示到目前为止, 参与训练的图像数量
如果出现上图-nan的情况,不要慌:只是说明本组图像中在此尺度下的特征图中检测不到正样本而已。
训练可能会出现的问题
- 问题一
"annot load image "/root/darknet/my_data/train_images/fimg_919.jpg
STB Reason: can't fopen
解决的前提:先要确定数据和标签一一对应!!!
解决方法一:
出现问题的可能原因是路径问题:如果不是用脚本写的路径,可能会存在中英文字符不相同的情况
训练前:需要用notepad++修改,先点击视图->显示符号->显示所有字符,然后点编辑->文档格式转换->转liunx,保证每一行最后都只有一个LF,一般是最后一行的问题。
解决方法二:https://blog.csdn.net/hyl999/article/details/90765633?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522170305541316800185858935%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request_id=170305541316800185858935&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2allfirst_rank_ecpm_v1~rank_v31_ecpm-1-90765633-null-null.142v96pc_search_result_base9&utm_term=Resizing%20480%20annot%20load%20image%20%2Froot%2Fdarknet%2Fmy_data%2Ftrain_images%2F0027.png%20STB%20Reason%3A%20cant%20fopen&spm=1018.2226.3001.4187
sed -i 's/\r//g' my_data/trainImageId.txt
sed -i 's/\r//g' my_data/valImagePath.txt
sed -i 's/\r//g' my_data/trainImagePath.txt
sed -i 's/\r//g' my_data/valImageId.txt
- 问题二:
yolov3出现nun爆炸
https://blog.csdn.net/weixin_42234720/article/details/93751752
断点训练
如终止训练,权重会保存在backup文件夹下。如果要从检查点停止并重新启动训练
./darknet detector train my_data/voc.data my_data/yolov3-tiny.cfg backup/yolov3-tiny.backup ./darknet detector train cfg/coco.data cfg/yolov3.cfg backup/yolov3.backup -gpus 0,1,2,3
结果文件
训练产生的模型文件被保存在了voc.data文件中指定的backup = train_my_data/backup下面
推理验证
./darknet detector test my_data/voc.data my_data/yolov3-tiny.cfg backup/yolov3-tiny.weights test.jpg -thresh 0.1