在这篇文章darknet-yolov3 识别自定义数据集中，已经实现了用 yolov3 模型的自定义数据集训练，以此为参考，介绍使用 GPU 训练 yolov3-tiny 模型识别自己的数据集。详细的参照上一篇，有变动的以这篇为基准。
一、准备工作
1.1 修改 Makefile
从 github 上拉取 darknet 的源代码，然后修改 Makefile 的内容如下，注意支持架构 ARCH 的要根据自己的GPU 所支持的进行调整：
GPU=1 #如果使用GPU设置为1，CPU设置为0
CUDNN=1 #如果使用CUDNN设置为1，否则为0
OPENCV=0
OPENMP=0
DEBUG=0
ARCH= -gencode arch=compute_30,code=sm_30 \
-gencode arch=compute_35,code=sm_35 \
-gencode arch=compute_50,code=[sm_50,compute_50] \
-gencode arch=compute_52,code=[sm_52,compute_52]
# This is what I use, uncomment if you know your arch and want to specify
# ARCH= -gencode arch=compute_52,code=compute_52
1.2 替换源码内容
修改 Makefile 后直接对源码进行编译可能会出现以下报错
原因是 darknet 较为老旧，其中包含的cuda相关的代码只写到7.0版本，需要修改。（建议直接将convolutional_layer.c 文件中的代码全部替换）
convolutional_layer.c 文件是 darknet 的源代码，在 darknet/src 目录下，清空源代码，将以下内容复制进去
#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;
}
之后再使用 make 命令进行编译，即可成功
1.3 下载 yolov3-tiny 权重文件
如果是训练 yolov3-tiny 模型，就不用下载 yolov3.weights 了，下载 yolov3-tiny.weights 到 darknet 目录下
wget https://pjreddie.com/media/files/yolov3-tiny.weights
测试是否能正常识别，注意是使用了 yolov3-tiny.weights 文件
./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg
1.4 数据集准备
按照 darknet-yolov3 模型那篇文章的数据集准备部分，准备好 myData 文件夹，结构如下所示：
myData/
├── Annotations # xml 标签文件夹
├── ImageSets
│ └── Main #花名册文件夹
│ └── train.txt
│ └── val.txt
│ └── test.txt
│ └── trainval.txt
├── JPEGImages #数据集照片
├── labels # txt 标签文件夹
├── weights #权重文件夹
└── myData_train.txt # train.txt花名册中的所有图片的绝对路径
二、修改训练配置文件
训练所需的配置文件主要在 darknet/cfg 中，这里用到 voc.data 和 yolov3-tiny.cfg 这两个配置文件
cp cfg/voc.data cfg/my_data.data
cp cfg/yolov3-tiny.cfg cfg/my_yolov3.cfg
2.1 修改 my_data.data 配置文件
修改 my_data.data 的内容如下：
classes= 3 ##改为自己的分类个数
##下面都改为自己的绝对路径
train = /home/XXX/darknet/myData/myData_train.txt #训练图片的绝对路径
names = /home/XXX/darknet/myData/myData.names #稍后需要创建这个文件
backup = /home/XXX/darknet/myData/weights #保存训练的权重文件的文件夹
2.2 修改 my_yolov3.cfg 配置文件
上面将 yolov3-tiny.cfg 复制了一份，重命名为 my_yolov3.cfg 。
使用 vi 编辑器，输入 /yolo，总共会搜出 2 个含有yolo的地方。
每个地方都必须要改2个参数（filters 和 classes）
classes 的值改为要识别的种类数目
filters 的值改为 3*（5+种类数目）
如要识别的种类有 3 种，则
filters = 24
classes = 3
可选修改：random = 1：原来是1，如果显存较小则改为0 （random 决定是否要多尺度输出）
除此之外，还可以修改训练轮数 max_batches ，下图是文件开头内容：
max_batches参数根据总类别修改，为类别数x2000，steps参数改为max_batches的80%和90%，大致如此
[net]
# Testing ### 测试模式，测试的时候把 Training 下面的 batch 和 subdivisions 注释掉，取消 Testing 下的 batch 和 subdivisions 的注释
# batch=1
# subdivisions=1
# Training ### 训练模式，每次前向的图片数目 = batch/subdivisions
batch=64 ### 每batch个样本更新一次参数
subdivisions=16 ### 如果内存不够大，将batch分割为subdivisions个子batch，每个子batch的大小为batch/subdivisions
width=416 ### 网络的输入宽、高、通道数
height=416
channels=3
momentum=0.9 ### 动量
decay=0.0005 ### 权重衰减
angle=0
saturation = 1.5 ### 饱和度
exposure = 1.5 ### 曝光度
hue=.1 ### 色调
learning_rate=0.001 ### 学习率
burn_in=1000 ### 学习率控制的参数
max_batches = 50200 ### 迭代次数
policy=steps ### 学习率策略
steps=40000,45000 ### 学习率变动步长
2.3 创建 myData.names 文件
在 myData文件夹下新建 myData.names 文件，内容为识别类型的名字，每个名字一行
people
foot
car
三、训练模型
3.1 生成预训练模型
在 darknet 目录下运行如下命令，生成预训练模型
./darknet partial cfg/my_yolov3.cfg yolov3-tiny.weights yolov3-tiny.conv.15 15
3.2 开始训练
在 darknet 目录下，用之前修改好的配置文件进行训练
./darknet detector train cfg/my_data.data cfg/my_yolov3.cfg yolov3-tiny.conv.15
从停止处开始训练
./darknet detector train cfg/my_data.data cfg/my_yolov3.cfg yolov3-tiny.conv.15 myData/weights/my_yolov3.backup
说明：myData/weights/my_yolov3.backup 是每一轮迭代后所权重文件缓存
四、测试
训练完成后，修改 my_yolov3.cfg 文件开头，使用 Testing ，注释掉 Training , 使用数据集图片和指定的权重文件进行测试，命令如下
./darknet detector test cfg/my_data.data cfg/my_yolov3.cfg myData/weights/my_yolov3_9000.weights myData/JPEGImages/1.jpg
3317
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
