darknet框架GPU编译安装

news2024/11/22 21:20:05

Darknet: Open Source Neural Networks in C

1、darknet下载

git clone https://github.com/pjreddie/darknet.git
cd darknet

设置makefile 

gpu=1 
cudnn=1 
opencv=1

【1】GPU=1;需要设置显卡驱动、cuda

  • 使用nvidia-smi 查看显卡型号和支持的cuda版本号

在这里插入图片描述

  • nvidia官网下载cuda,以及cudnn

在这里插入图片描述

安装cuda若提示

Existing package manager installation of the driver found. It is strongly recommended that you remove this before continuing

原因是驱动重复安装，卸载掉其他驱动

dpkg -l | grep Nvidia //查看驱动
sudo apt-get purge "nvidia*"  //卸载旧版本驱动

然后再次安装就正常了。成功之后显示

===========
= Summary =
===========

Driver:   Not Selected
Toolkit:  Installed in /usr/local/cuda-11.6/

Please make sure that
 -   PATH includes /usr/local/cuda-11.6/bin
 -   LD_LIBRARY_PATH includes /usr/local/cuda-11.6/lib64, or, add /usr/local/cuda-11.6/lib64 to /etc/ld.so.conf and run ldconfig as root

To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-11.6/bin
***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 510.00 is required for CUDA 11.6 functionality to work.
To install the driver using this installer, run the following command, replacing <CudaInstaller> with the name of this run file:
    sudo <CudaInstaller>.run --silent --driver

Logfile is /var/log/cuda-installer.log

添加cuda到系统路径，vim ~/.zshrv 

export PATH=/usr/local/cuda-11.6/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-11.6/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

运行source ~/.zshrc 让路径生效，此时可以输入命令nvcc -V验证一下cuda

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Thu_Feb_10_18:23:41_PST_2022
Cuda compilation tools, release 11.6, V11.6.112
Build cuda_11.6.r11.6/compiler.30978841_0

【2】cudnn=1

  • nvidia官网选择相应版本的cudnn,进行下载（建议下载可解压版本的，方便自己操作）
    在这里插入图片描述

将解压出来的cudnn文件copy到cuda路径中(usl/local中会有两个cuda路径，一个带版本号，一个不带，记得是copy到不带版本号的cuda路径中)

sudo cp include/cudnn*.h /usr/local/cuda/include
sudo cp lib/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*

copy完成之后用cat /usr/local/cuda/include/cudnn_version.h | grep CUDNN_MAJOR -A 2 验证一下

#define CUDNN_MAJOR 8
#define CUDNN_MINOR 7
#define CUDNN_PATCHLEVEL 0
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

2、darknet编译

由于版本问题，需要先修改几个文件

  • 用https://github.com/arnoldfychen/darknet/blob/master/src/convolutional_layer.c 直接替换darknet/src/convolutional_laye.c文件，老版本不支持cudnn8以上的
#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;
}
  • 修改/src/gemm.c中的cudaThreadSynchronize为cudaDeviceSynchronize
  • Makefile中添加,并删除掉低版本的信息
-gencode arch=compute_70,code=[sm_70,compute_70] \
-gencode arch=compute_75,code=[sm_75,compute_75] \
-gencode arch=compute_86,code=[sm_86,compute_86]

最后，用make命令编译

3、编译出错

/bin/sh: 1: nvcc: not found
sudo

修改Makefile中nvcc 的路径

NVCC=/usr/local/cuda/bin/nvcc

本文来自互联网用户投稿，该文观点仅代表作者本人，不代表本站立场。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如若转载，请注明出处：http://www.coloradmin.cn/o/58785.html

如若内容造成侵权/违法违规/事实不符，请联系多彩编程网进行投诉反馈，一经查实，立即删除！

相关文章

计算机网络学习笔记（Ⅱ）：物理层

计算机网络学习笔记（Ⅱ）：物理层

目录 1 物理层概念 1.1 物理层基本概念 1.定义 2.主要任务 3.特性 1.2 数据通信基础 1.典型模型 2.相关术语 3.三种通信方式 4.数据传输方式 1.3 物理层内容 1.码元 2.速率 3.带宽 1.4 奈氏准则与香农定理 1.失真 2.码间串扰 3.奈氏准则 4.香农定理 1.5 …
阅读更多...
蓝桥杯C/C++VIP试题每日一练之Huffman树

蓝桥杯C/C++VIP试题每日一练之Huffman树

💛作者主页：静Yu 🧡简介：CSDN全栈优质创作者、华为云享专家、阿里云社区博客专家，前端知识交流社区创建者 💛社区地址：前端知识交流社区 🧡博主的个人博客：静Yu的个人博客 🧡博主的个人笔记本：前端面试题 个人笔记本只记录前端领域的面试题目，项目总结，面试技…
阅读更多...
基于JSP的某餐厅点餐系统

基于JSP的某餐厅点餐系统

目 录 第一章 绪论 1 1.1系统研究背景和意义 1 1.2研究现状 1 1.3研究主要内容 2 第二章 相关技术说明 3 2.1 JSP(Java Server Page)简介 3 2.2 Spring框架简介 4 2.3 Spring MVC框架简介 5 2.4 MyBatis 框架简介 5 2.4 MySql数据库简介 6 2.6 Tomcat简介 7 2.7 jQuery简介 8 …
阅读更多...
Hadoop原理与技术——Hbase的基本操作

Hadoop原理与技术——Hbase的基本操作

点击链接查看文档 一、实验目的 上机实操&#xff0c;熟悉指令操作Hbase和java代码操作Hbase 二、实验环境 Windows 10 VMware Workstation Pro虚拟机 Hadoop环境 Jdk1.8 三、实验内容 1&#xff1a;指令操作Hbase (1)&#xff1a;start-all.sh&#xff0c;启动所有进程 (2)…
阅读更多...
Ansys（Maxwell、Simplorer）与Simulink联合仿真（二）直线电机

Ansys（Maxwell、Simplorer）与Simulink联合仿真（二）直线电机

Ansys&#xff08;Maxwell、Simplorer&#xff09;与Simulink联合仿真&#xff08;二&#xff09;直线电机 在仿真过程中&#xff0c;遇到了一个问题&#xff0c;卡了好久得到了解决。 关于 motion setup 提示 moving 找不到面 cannot find the sarface 所有的动态部件要隔开…
阅读更多...
【pen200-lab】10.11.1.21（实际获得22权限）

【pen200-lab】10.11.1.21（实际获得22权限）

pen200-lab 学习笔记 【pen200-lab】10.11.1.21 &#x1f525;系列专栏&#xff1a;pen200-lab &#x1f389;欢迎关注&#x1f50e;点赞&#x1f44d;收藏⭐️留言&#x1f4dd; &#x1f4c6;首发时间&#xff1a;&#x1f334;2022年11月27日&#x1f334; &#x1f36d;作…
阅读更多...
算法导论24章单源最短路径—Bellman-Ford算法 Dijkstra算法

算法导论24章单源最短路径—Bellman-Ford算法 Dijkstra算法

松弛操作 松弛操作就是判断从现在s到v的路径更近&#xff0c;还是我从s到u再到v更近&#xff0c;选一个更近的走。 松弛操作的例子 松弛是唯一导致最短路径估计和前驱结点变化的操作 Bellman-Ford算法 第一个循环&#xff0c;循环V-1次&#xff0c;每次循环对所有的边都松弛一…
阅读更多...
Python数据分析-matplotlib

Python数据分析-matplotlib

目录 一、折线图&#xff1a;plt.plot() 1.1 plt.plot()基本用法 1.2 设置坐标轴范围&#xff1a;plt.axis([xmin,xmax,ymin,ymax]) 1.3 plt.plot()绘制多个图形 1.4 linewidth设置线条宽度 1.5 使用plt.plot()的返回值设置线条属性 1.6 plt.setp()修改线条性质 1.7 对…
阅读更多...
软件测试的几种方法

软件测试的几种方法

1、从是否关心内部结构来看 (1)白盒测试&#xff1a;又称为结构测试或逻辑驱动测试&#xff0c;是一种按照程序内部逻辑结构和编码结构&#xff0c;设计测试数据并完成测试的一种测试方法。 (2)黑盒测试&#xff1a;又称为数据驱动测试&#xff0c;把测试对象当做看不见的黑盒…
阅读更多...
讲透金融风控建模全流程(附 Python 代码)

讲透金融风控建模全流程(附 Python 代码)

信贷风控是数据挖掘算法最成功的应用之一&#xff0c;这在于金融信贷行业的数据量很充足&#xff0c;需求场景清晰及丰富。 信贷风控简单来说就是判断一个人借了钱后面&#xff08;如下个月的还款日&#xff09;会不会按期还钱。更专业来说&#xff0c;信贷风控是还款能力及还…
阅读更多...
SQL 汇总统计及GROUP BY

SQL 汇总统计及GROUP BY

SQL 汇总统计1、汇总统计2、GROUT BY3、如何对分组统计的结果进行过滤&#xff1f; GROUP BY HAVING4、如何对分组统计的结果进行排序&#xff1f;GROUP BY ORDER BY5、介绍SELECT语句中各个子句的书写顺序6、备注&#xff1a; 上方用到的表1、汇总统计 介绍几个聚集函数 有…
阅读更多...
电脑误删Path环境变量后前端如何重新配置所需变量

电脑误删Path环境变量后前端如何重新配置所需变量

需求背景 &#xff1a; 当时公司需要我们安装一款软件 &#xff0c; 按照操作文档需要配置一下 Path 环境变量 &#xff0c; 但当时的云桌面操作系统是 window7系统 &#xff0c; 当时配置时并不知道新的变量配置时需要在前面一个的后面加 “ &#xff1b; ” 来间隔开来…
阅读更多...
【目的：windows下VS2017/2022使用MSVC编译GLFW库】

【目的：windows下VS2017/2022使用MSVC编译GLFW库】

目的&#xff1a;windows下VS2017/2022使用MSVC编译GLFW库 环境&#xff1a; 系统&#xff1a;Win10 环境&#xff1a;VS2017 64bit步骤&#xff1a; 1.下载GLFW源码 官网链接https://www.glfw.org/download.html&#xff0c; 下载glfw的源码&#xff0c;解压到本地&#x…
阅读更多...
考研数据结构大题整合_组二（TJP组）

考研数据结构大题整合_组二（TJP组）

考研数据结构大题整合 目录考研数据结构大题整合二、TJP组TJP组一TJP组二TJP组三二、TJP组 TJP组一 四、画图/计算/证明/算法分析&#xff08;30分&#xff09; &#xff08;1&#xff09;证明题&#xff08;8分&#xff09; 如果一棵树有n1个度为1的结点&#xff0c;n2个度为…
阅读更多...
（四）Vue之数据绑定

（四）Vue之数据绑定

文章目录数据绑定单向数据绑定双向数据绑定Vue学习目录上一篇&#xff1a;&#xff08;三&#xff09;Vue之模板语法 数据绑定 Vue中有2种数据绑定的方式&#xff1a; 1.单向绑定&#xff1a;数据只能从data流向页面。2.双向绑定&#xff1a;数据不仅能从data流向页面&#…
阅读更多...
著名书画家、中国书画院院士李适中

著名书画家、中国书画院院士李适中

著名书画家、中国书画院院士李适中 李适中　著名书画家、中国书画院院士 版画艺术家 文物复制专家 中国文物学会会员单位创始人 文化部科技进步奖获得者 艺术简历 李适中&#xff0c;1943年生&#xff0c;安徽颍上人&#xff0c;著名书画家、中国书画院院士。李适中先生师从著名…
阅读更多...
Vue3+nodejs全栈项目(资金管理系统)——前端篇

Vue3+nodejs全栈项目(资金管理系统)——前端篇

文章目录创建项目项目初始化使用element-plus设置Register和404组件搭建element注册表单验证表单和按钮加载动画和消息提醒路由守卫和token过期处理配置请求拦截和响应拦截解析token并存储到vuex中设计顶部导航设置首页和个人信息设置左侧导航栏展示资金管理页面添加按钮编辑和…
阅读更多...
返回当前系统串口名称

返回当前系统串口名称

主要针对当前的usb转串口进行了穷举。 方便判断串口对应哪个设备。 返回串口名称 类对象&#xff0c;&#xff08;包含了参考网址&#xff0c;以及对其进行了修改&#xff0c;防止出现蓝牙端口&#xff09; using System; using System.Collections.Generic; using System.L…
阅读更多...
VMware-KVM安装

VMware-KVM安装

目录 VMware-KVM安装 一、kvm虚拟化平台 KVM 网络管理&#xff08;以NAT网卡为例[ens33]&#xff09; VMware-KVM安装 一台Centos7、一个winSCP上传文件工具&#xff1b; 搭建KVM平台 一、kvm虚拟化平台 1 cat /etc/hosts ##查看主机…
阅读更多...
ecology修改Reisn的JDK目录

ecology修改Reisn的JDK目录

修改resin运行JDK&#xff1a; 用文本编辑器打开resin/bin/resin.sh文件&#xff0c;将JAVA_HOME改为要设置的JDK路径。
阅读更多...
推荐文章
最新文章

Copyright @ 2022~2023