首页 > 系统相关 >Ubuntu18.04 + Caffe + python3.7 + CUDA11 + cuDNN8编译记录 转载文章 非原创

Ubuntu18.04 + Caffe + python3.7 + CUDA11 + cuDNN8编译记录 转载文章 非原创

时间:2024-05-30 16:56:41浏览次数:28  
标签:cuDNN8 CUDNN filter python3.7 Caffe algo workspace bwd data

背景

这两天接手了一个在两年前基于caffe实现的交互式活体检测的项目,想要让他在python3 和CUDA 11的环境下运行。但是呢,caffe已经官方宣布不再继续更新,不支持最新版的cuDNN8,那需求摆在这边只好自行想办法,前前后后倒腾了两天,可算是编译成功把项目跑通了,在此记录一下自己配置辛酸史。
基础环境

    Ubuntu 18.04
    CUDA 11.0
    cuDNN 8

安装过程
Python3.7安装

在这里有一个注意点就是python3.7安装编译的时候一定要fPIC动态编译,否则后续编译caffe的时候会报fPIC的相关错误,安装指令:

apt-get update
apt-get upgrade
apt install build-essential -y
apt install libncurses5-dev libgdbm-dev libnss3-dev libssl-dev libreadline-dev libffi-dev -y
apt install zlib1g-dev
apt install wget
apt install openssl
apt install curl
apt install libsqlite3-dev
wget https://www.python.org/ftp/python/3.7.3/Python-3.7.3.tgz
tar -xvf Python-3.7.3.tgz
cd Python-3.7.3
./configure --enable-loadable-sqlite-extensions --prefix=/usr/local/  --enable-shared CFLAGS=-fPIC
make
make install
apt-get clean
rm -rf /var/lib/apt/lists/*
ln -s /usr/local/bin/pip3 /usr/bin/pip
ln -s /usr/local/bin/python3 /usr/bin/python

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19

caffe官方代码下载及配置文件修改

apt-get update
apt-get install git
cd /
git clone https://github.com/BVLC/caffe.git
cd caffe
vim CMakeLists.txt

    1
    2
    3
    4
    5
    6

第35行的set(python_version “2” CACHE STRING “Specify which Python version to use”)中的2改为3.7,保存退出

cp Makefile.config.example Makefile.config
vim Makefile.config

    1
    2

修改Makefile.config内容至如下后保存退出:

## Refer to http://caffe.berkeleyvision.org/installation.html
# Contributions simplifying and improving our build system are welcome!

# cuDNN acceleration switch (uncomment to build with cuDNN).
USE_CUDNN := 1

# CPU-only switch (uncomment to build without GPU support).
#CPU_ONLY := 1

# uncomment to disable IO dependencies and corresponding data layers
USE_OPENCV := 0
# USE_LEVELDB := 0
# USE_LMDB := 0
# This code is taken from https://github.com/sh1r0/caffe-android-lib
# USE_HDF5 := 0

# uncomment to allow MDB_NOLOCK when reading LMDB files (only if necessary)
#    You should not set this flag if you will be reading LMDBs with any
#    possibility of simultaneous read and write
# ALLOW_LMDB_NOLOCK := 1

# Uncomment if you're using OpenCV 3
OPENCV_VERSION := 3

# To customize your choice of compiler, uncomment and set the following.
# N.B. the default for Linux is g++ and the default for OSX is clang++
# CUSTOM_CXX := g++

# CUDA directory contains bin/ and lib/ directories that we need.
CUDA_DIR := /usr/local/cuda
# On Ubuntu 14.04, if cuda tools are installed via
# "sudo apt-get install nvidia-cuda-toolkit" then use this instead:
# CUDA_DIR := /usr

# CUDA architecture setting: going with all of them.
# For CUDA < 6.0, comment the *_50 through *_61 lines for compatibility.
# For CUDA < 8.0, comment the *_60 and *_61 lines for compatibility.
# For CUDA >= 9.0, comment the *_20 and *_21 lines for compatibility.
CUDA_ARCH := -gencode arch=compute_20,code=sm_20 \
        -gencode arch=compute_20,code=sm_21 \
        -gencode arch=compute_30,code=sm_30 \
        -gencode arch=compute_35,code=sm_35 \
        -gencode arch=compute_50,code=sm_50 \
        -gencode arch=compute_52,code=sm_52 \
        -gencode arch=compute_60,code=sm_60 \
        -gencode arch=compute_61,code=sm_61 \
        -gencode arch=compute_61,code=compute_61

# BLAS choice:
# atlas for ATLAS (default)
# mkl for MKL
# open for OpenBlas
BLAS := atlas
# Custom (MKL/ATLAS/OpenBLAS) include and lib directories.
# Leave commented to accept the defaults for your choice of BLAS
# (which should work)!
# BLAS_INCLUDE := /path/to/your/blas
# BLAS_LIB := /path/to/your/blas

# Homebrew puts openblas in a directory that is not on the standard search path
# BLAS_INCLUDE := $(shell brew --prefix openblas)/include
# BLAS_LIB := $(shell brew --prefix openblas)/lib

# This is required only if you will compile the matlab interface.
# MATLAB directory should contain the mex binary in /bin.
# MATLAB_DIR := /usr/local
# MATLAB_DIR := /Applications/MATLAB_R2012b.app

# NOTE: this is required only if you will compile the python interface.
# We need to be able to find Python.h and numpy/arrayobject.h.
#PYTHON_INCLUDE := /usr/include/python2.7 \
        /usr/lib/python2.7/dist-packages/numpy/core/include
# Anaconda Python distribution is quite popular. Include path:
# Verify anaconda location, sometimes it's in root.
# ANACONDA_HOME := $(HOME)/anaconda
# PYTHON_INCLUDE := $(ANACONDA_HOME)/include \
        # $(ANACONDA_HOME)/include/python2.7 \
        # $(ANACONDA_HOME)/lib/python2.7/site-packages/numpy/core/include

# Uncomment to use Python 3 (default is Python 2)
PYTHON_LIBRARIES := boost_python3 python3.7m
PYTHON_INCLUDE := /usr/local/include/python3.7m \
                /usr/local/lib/python3.7/dist-packages/numpy/core/include

# We need to be able to find libpythonX.X.so or .dylib.
PYTHON_LIB := /usr/local/lib
# PYTHON_LIB := $(ANACONDA_HOME)/lib

# Homebrew installs numpy in a non standard path (keg only)
# PYTHON_INCLUDE += $(dir $(shell python -c 'import numpy.core; print(numpy.core.__file__)'))/include
# PYTHON_LIB += $(shell brew --prefix numpy)/lib

# Uncomment to support layers written in Python (will link against Python libs)
WITH_PYTHON_LAYER := 1

# Whatever else you find you need goes here.
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /usr/include/hdf5/serial/
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib

# If Homebrew is installed at a non standard location (for example your home directory) and you use it for general dependencies
# INCLUDE_DIRS += $(shell brew --prefix)/include
# LIBRARY_DIRS += $(shell brew --prefix)/lib

# NCCL acceleration switch (uncomment to build with NCCL)
# https://github.com/NVIDIA/nccl (last tested version: v1.2.3-1+cuda8.0)
# USE_NCCL := 1

# Uncomment to use `pkg-config` to specify OpenCV library paths.
# (Usually not necessary -- OpenCV libraries are normally installed in one of the above $LIBRARY_DIRS.)
# USE_PKG_CONFIG := 1

# N.B. both build and distribute dirs are cleared on `make clean`
BUILD_DIR := build
DISTRIBUTE_DIR := distribute

# Uncomment for debugging. Does not work on OSX due to https://github.com/BVLC/caffe/issues/171
# DEBUG := 1

# The ID of the GPU that 'make runtest' will use to run unit tests.
TEST_GPUID := 0

# enable pretty build (comment to see full commands)
Q ?= @

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    30
    31
    32
    33
    34
    35
    36
    37
    38
    39
    40
    41
    42
    43
    44
    45
    46
    47
    48
    49
    50
    51
    52
    53
    54
    55
    56
    57
    58
    59
    60
    61
    62
    63
    64
    65
    66
    67
    68
    69
    70
    71
    72
    73
    74
    75
    76
    77
    78
    79
    80
    81
    82
    83
    84
    85
    86
    87
    88
    89
    90
    91
    92
    93
    94
    95
    96
    97
    98
    99
    100
    101
    102
    103
    104
    105
    106
    107
    108
    109
    110
    111
    112
    113
    114
    115
    116
    117
    118
    119
    120
    121
    122
    123

caffe与python3.7/cuDNN8适配
cudnn_conv_layer.cpp和cudnn_deconv_layer.cpp

caffe最后支持的版本是cuDNN7.6.5,为了能在cuDNN8的环境下编译通过,需要修改两个cpp文件,路径为/caffe/src/caffe/layers下的cudnn_conv_layer.cpp和cudnn_deconv_layer.cpp两个文件,分别将他们内容替换为:

cudnn_conv_layer.cpp

#ifdef USE_CUDNN
#include <algorithm>
#include <vector>

#include "caffe/layers/cudnn_conv_layer.hpp"

namespace caffe {

// Set to three for the benefit of the backward pass, which
// can use separate streams for calculating the gradient w.r.t.
// bias, filter weights, and bottom data for each group independently
#define CUDNN_STREAMS_PER_GROUP 3

/**
 * TODO(dox) explain cuDNN interface
 */
template <typename Dtype>
void CuDNNConvolutionLayer<Dtype>::LayerSetUp(
    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  ConvolutionLayer<Dtype>::LayerSetUp(bottom, top);
  // Initialize CUDA streams and cuDNN.
  stream_         = new cudaStream_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
  handle_         = new cudnnHandle_t[this->group_ * CUDNN_STREAMS_PER_GROUP];

  // Initialize algorithm arrays
  fwd_algo_       = new cudnnConvolutionFwdAlgo_t[bottom.size()];
  bwd_filter_algo_= new cudnnConvolutionBwdFilterAlgo_t[bottom.size()];
  bwd_data_algo_  = new cudnnConvolutionBwdDataAlgo_t[bottom.size()];

  // initialize size arrays
  workspace_fwd_sizes_ = new size_t[bottom.size()];
  workspace_bwd_filter_sizes_ = new size_t[bottom.size()];
  workspace_bwd_data_sizes_ = new size_t[bottom.size()];

  // workspace data
  workspaceSizeInBytes = 0;
  workspaceData = NULL;
  workspace = new void*[this->group_ * CUDNN_STREAMS_PER_GROUP];

  for (size_t i = 0; i < bottom.size(); ++i) {
    // initialize all to default algorithms
    fwd_algo_[i] = (cudnnConvolutionFwdAlgo_t)0;
    bwd_filter_algo_[i] = (cudnnConvolutionBwdFilterAlgo_t)0;
    bwd_data_algo_[i] = (cudnnConvolutionBwdDataAlgo_t)0;
    // default algorithms don't require workspace
    workspace_fwd_sizes_[i] = 0;
    workspace_bwd_data_sizes_[i] = 0;
    workspace_bwd_filter_sizes_[i] = 0;
  }

  for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
    CUDA_CHECK(cudaStreamCreate(&stream_[g]));
    CUDNN_CHECK(cudnnCreate(&handle_[g]));
    CUDNN_CHECK(cudnnSetStream(handle_[g], stream_[g]));
    workspace[g] = NULL;
  }

  // Set the indexing parameters.
  bias_offset_ = (this->num_output_ / this->group_);

  // Create filter descriptor.
  const int* kernel_shape_data = this->kernel_shape_.cpu_data();
  const int kernel_h = kernel_shape_data[0];
  const int kernel_w = kernel_shape_data[1];
  cudnn::createFilterDesc<Dtype>(&filter_desc_,
      this->num_output_ / this->group_, this->channels_ / this->group_,
      kernel_h, kernel_w);

  // Create tensor descriptor(s) for data and corresponding convolution(s).
  for (int i = 0; i < bottom.size(); i++) {
    cudnnTensorDescriptor_t bottom_desc;
    cudnn::createTensor4dDesc<Dtype>(&bottom_desc);
    bottom_descs_.push_back(bottom_desc);
    cudnnTensorDescriptor_t top_desc;
    cudnn::createTensor4dDesc<Dtype>(&top_desc);
    top_descs_.push_back(top_desc);
    cudnnConvolutionDescriptor_t conv_desc;
    cudnn::createConvolutionDesc<Dtype>(&conv_desc);
    conv_descs_.push_back(conv_desc);
  }

  // Tensor descriptor for bias.
  if (this->bias_term_) {
    cudnn::createTensor4dDesc<Dtype>(&bias_desc_);
  }

  handles_setup_ = true;
}

template <typename Dtype>
void CuDNNConvolutionLayer<Dtype>::Reshape(
    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  ConvolutionLayer<Dtype>::Reshape(bottom, top);
  CHECK_EQ(2, this->num_spatial_axes_)
      << "CuDNNConvolution input must have 2 spatial axes "
      << "(e.g., height and width). "
      << "Use 'engine: CAFFE' for general ND convolution.";
  bottom_offset_ = this->bottom_dim_ / this->group_;
  top_offset_ = this->top_dim_ / this->group_;
  const int height = bottom[0]->shape(this->channel_axis_ + 1);
  const int width = bottom[0]->shape(this->channel_axis_ + 2);
  const int height_out = top[0]->shape(this->channel_axis_ + 1);
  const int width_out = top[0]->shape(this->channel_axis_ + 2);
  const int* pad_data = this->pad_.cpu_data();
  const int pad_h = pad_data[0];
  const int pad_w = pad_data[1];
  const int* stride_data = this->stride_.cpu_data();
  const int stride_h = stride_data[0];
  const int stride_w = stride_data[1];
#if CUDNN_VERSION_MIN(8, 0, 0)
  int RetCnt;
  bool found_conv_algorithm;
  size_t free_memory, total_memory;
  cudnnConvolutionFwdAlgoPerf_t     fwd_algo_pref_[4];
  cudnnConvolutionBwdDataAlgoPerf_t bwd_data_algo_pref_[4];

  //get memory sizes
  cudaMemGetInfo(&free_memory, &total_memory);
#else
  // Specify workspace limit for kernels directly until we have a
  // planning strategy and a rewrite of Caffe's GPU memory mangagement
  size_t workspace_limit_bytes = 8*1024*1024;
#endif
  for (int i = 0; i < bottom.size(); i++) {
    cudnn::setTensor4dDesc<Dtype>(&bottom_descs_[i],
        this->num_,
        this->channels_ / this->group_, height, width,
        this->channels_ * height * width,
        height * width, width, 1);
    cudnn::setTensor4dDesc<Dtype>(&top_descs_[i],
        this->num_,
        this->num_output_ / this->group_, height_out, width_out,
        this->num_output_ * this->out_spatial_dim_,
        this->out_spatial_dim_, width_out, 1);
    cudnn::setConvolutionDesc<Dtype>(&conv_descs_[i], bottom_descs_[i],
        filter_desc_, pad_h, pad_w,
        stride_h, stride_w);

#if CUDNN_VERSION_MIN(8, 0, 0)
    // choose forward algorithm for filter
    // in forward filter the CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED is not implemented in cuDNN 8
    CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm_v7(handle_[0],
      bottom_descs_[i],
      filter_desc_,
      conv_descs_[i],
      top_descs_[i],
      4,
      &RetCnt,
      fwd_algo_pref_));
        
    found_conv_algorithm = false;
    for(int n=0;n<RetCnt;n++){
      if (fwd_algo_pref_[n].status == CUDNN_STATUS_SUCCESS &&
          fwd_algo_pref_[n].algo != CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED &&
          fwd_algo_pref_[n].memory < free_memory){
        found_conv_algorithm = true;
    fwd_algo_[i]                   = fwd_algo_pref_[n].algo;
     workspace_fwd_sizes_[i]        = fwd_algo_pref_[n].memory;
        break;
      }
    }
    if(!found_conv_algorithm) LOG(ERROR) << "cuDNN did not return a suitable algorithm for convolution.";
    else{
    // choose backward algorithm for filter
        // for better or worse, just a fixed constant due to the missing
        // cudnnGetConvolutionBackwardFilterAlgorithm in cuDNN version 8.0
    bwd_filter_algo_[i] = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0;
    //twice the amount of the forward search to be save     
        workspace_bwd_filter_sizes_[i] = 2*workspace_fwd_sizes_[i];
    }

    // choose backward algo for data
    CUDNN_CHECK(cudnnGetConvolutionBackwardDataAlgorithm_v7(handle_[0],
      filter_desc_,
      top_descs_[i],
      conv_descs_[i],
      bottom_descs_[i],
      4,
      &RetCnt,
      bwd_data_algo_pref_));

    found_conv_algorithm = false;
    for(int n=0;n<RetCnt;n++){
      if (bwd_data_algo_pref_[n].status == CUDNN_STATUS_SUCCESS &&
          bwd_data_algo_pref_[n].algo != CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD &&
          bwd_data_algo_pref_[n].algo != CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED &&
          bwd_data_algo_pref_[n].memory < free_memory){
        found_conv_algorithm = true;
    bwd_data_algo_[i]              = bwd_data_algo_pref_[n].algo;
     workspace_bwd_data_sizes_[i]   = bwd_data_algo_pref_[n].memory;
        break;
      }
    }
    if(!found_conv_algorithm) LOG(ERROR) << "cuDNN did not return a suitable algorithm for convolution.";
#else
    // choose forward and backward algorithms + workspace(s)
    CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm(handle_[0],
      bottom_descs_[i],
      filter_desc_,
      conv_descs_[i],
      top_descs_[i],
      CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
      workspace_limit_bytes,
      &fwd_algo_[i]));

    CUDNN_CHECK(cudnnGetConvolutionForwardWorkspaceSize(handle_[0],
      bottom_descs_[i],
      filter_desc_,
      conv_descs_[i],
      top_descs_[i],
      fwd_algo_[i],
      &(workspace_fwd_sizes_[i])));

    // choose backward algorithm for filter
    CUDNN_CHECK(cudnnGetConvolutionBackwardFilterAlgorithm(handle_[0],
          bottom_descs_[i], top_descs_[i], conv_descs_[i], filter_desc_,
          CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
          workspace_limit_bytes, &bwd_filter_algo_[i]) );

    // get workspace for backwards filter algorithm
    CUDNN_CHECK(cudnnGetConvolutionBackwardFilterWorkspaceSize(handle_[0],
          bottom_descs_[i], top_descs_[i], conv_descs_[i], filter_desc_,
          bwd_filter_algo_[i], &workspace_bwd_filter_sizes_[i]));

    // choose backward algo for data
    CUDNN_CHECK(cudnnGetConvolutionBackwardDataAlgorithm(handle_[0],
          filter_desc_, top_descs_[i], conv_descs_[i], bottom_descs_[i],
          CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
        workspace_limit_bytes, &bwd_data_algo_[i]));

    // get workspace size
    CUDNN_CHECK(cudnnGetConvolutionBackwardDataWorkspaceSize(handle_[0],
          filter_desc_, top_descs_[i], conv_descs_[i], bottom_descs_[i],
          bwd_data_algo_[i], &workspace_bwd_data_sizes_[i]) );
#endif
  }
  // reduce over all workspace sizes to get a maximum to allocate / reallocate
  size_t total_workspace_fwd = 0;
  size_t total_workspace_bwd_data = 0;
  size_t total_workspace_bwd_filter = 0;

  for (size_t i = 0; i < bottom.size(); i++) {
    total_workspace_fwd        = std::max(total_workspace_fwd,
                                     workspace_fwd_sizes_[i]);
    total_workspace_bwd_data   = std::max(total_workspace_bwd_data,
                                     workspace_bwd_data_sizes_[i]);
    total_workspace_bwd_filter = std::max(total_workspace_bwd_filter,
                                     workspace_bwd_filter_sizes_[i]);
  }
  // get max over all operations
  size_t max_workspace = std::max(total_workspace_fwd,
                             total_workspace_bwd_data);
  max_workspace = std::max(max_workspace, total_workspace_bwd_filter);
  // ensure all groups have enough workspace
  size_t total_max_workspace = max_workspace *
                               (this->group_ * CUDNN_STREAMS_PER_GROUP);

  // this is the total amount of storage needed over all groups + streams
  if (total_max_workspace > workspaceSizeInBytes) {
    DLOG(INFO) << "Reallocating workspace storage: " << total_max_workspace;
    workspaceSizeInBytes = total_max_workspace;

    // free the existing workspace and allocate a new (larger) one
    cudaFree(this->workspaceData);

    cudaError_t err = cudaMalloc(&(this->workspaceData), workspaceSizeInBytes);
    if (err != cudaSuccess) {
      // force zero memory path
      for (int i = 0; i < bottom.size(); i++) {
        workspace_fwd_sizes_[i] = 0;
        workspace_bwd_filter_sizes_[i] = 0;
        workspace_bwd_data_sizes_[i] = 0;
        fwd_algo_[i] = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
        bwd_filter_algo_[i] = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0;
        bwd_data_algo_[i] = CUDNN_CONVOLUTION_BWD_DATA_ALGO_0;
      }

      // NULL out all workspace pointers
      for (int g = 0; g < (this->group_ * CUDNN_STREAMS_PER_GROUP); g++) {
        workspace[g] = NULL;
      }
      // NULL out underlying data
      workspaceData = NULL;
      workspaceSizeInBytes = 0;
    }

    // if we succeed in the allocation, set pointer aliases for workspaces
    for (int g = 0; g < (this->group_ * CUDNN_STREAMS_PER_GROUP); g++) {
      workspace[g] = reinterpret_cast<char *>(workspaceData) + g*max_workspace;
    }
  }

  // Tensor descriptor for bias.
  if (this->bias_term_) {
    cudnn::setTensor4dDesc<Dtype>(&bias_desc_,
        1, this->num_output_ / this->group_, 1, 1);
  }
}

template <typename Dtype>
CuDNNConvolutionLayer<Dtype>::~CuDNNConvolutionLayer() {
  // Check that handles have been setup before destroying.
  if (!handles_setup_) { return; }

  for (int i = 0; i < bottom_descs_.size(); i++) {
    cudnnDestroyTensorDescriptor(bottom_descs_[i]);
    cudnnDestroyTensorDescriptor(top_descs_[i]);
    cudnnDestroyConvolutionDescriptor(conv_descs_[i]);
  }
  if (this->bias_term_) {
    cudnnDestroyTensorDescriptor(bias_desc_);
  }
  cudnnDestroyFilterDescriptor(filter_desc_);

  for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
    cudaStreamDestroy(stream_[g]);
    cudnnDestroy(handle_[g]);
  }

  cudaFree(workspaceData);
  delete [] stream_;
  delete [] handle_;
  delete [] fwd_algo_;
  delete [] bwd_filter_algo_;
  delete [] bwd_data_algo_;
  delete [] workspace_fwd_sizes_;
  delete [] workspace_bwd_data_sizes_;
  delete [] workspace_bwd_filter_sizes_;
}

INSTANTIATE_CLASS(CuDNNConvolutionLayer);

}   // namespace caffe
#endif

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    30
    31
    32
    33
    34
    35
    36
    37
    38
    39
    40
    41
    42
    43
    44
    45
    46
    47
    48
    49
    50
    51
    52
    53
    54
    55
    56
    57
    58
    59
    60
    61
    62
    63
    64
    65
    66
    67
    68
    69
    70
    71
    72
    73
    74
    75
    76
    77
    78
    79
    80
    81
    82
    83
    84
    85
    86
    87
    88
    89
    90
    91
    92
    93
    94
    95
    96
    97
    98
    99
    100
    101
    102
    103
    104
    105
    106
    107
    108
    109
    110
    111
    112
    113
    114
    115
    116
    117
    118
    119
    120
    121
    122
    123
    124
    125
    126
    127
    128
    129
    130
    131
    132
    133
    134
    135
    136
    137
    138
    139
    140
    141
    142
    143
    144
    145
    146
    147
    148
    149
    150
    151
    152
    153
    154
    155
    156
    157
    158
    159
    160
    161
    162
    163
    164
    165
    166
    167
    168
    169
    170
    171
    172
    173
    174
    175
    176
    177
    178
    179
    180
    181
    182
    183
    184
    185
    186
    187
    188
    189
    190
    191
    192
    193
    194
    195
    196
    197
    198
    199
    200
    201
    202
    203
    204
    205
    206
    207
    208
    209
    210
    211
    212
    213
    214
    215
    216
    217
    218
    219
    220
    221
    222
    223
    224
    225
    226
    227
    228
    229
    230
    231
    232
    233
    234
    235
    236
    237
    238
    239
    240
    241
    242
    243
    244
    245
    246
    247
    248
    249
    250
    251
    252
    253
    254
    255
    256
    257
    258
    259
    260
    261
    262
    263
    264
    265
    266
    267
    268
    269
    270
    271
    272
    273
    274
    275
    276
    277
    278
    279
    280
    281
    282
    283
    284
    285
    286
    287
    288
    289
    290
    291
    292
    293
    294
    295
    296
    297
    298
    299
    300
    301
    302
    303
    304
    305
    306
    307
    308
    309
    310
    311
    312
    313
    314
    315
    316
    317
    318
    319
    320
    321
    322
    323
    324
    325
    326
    327
    328
    329
    330
    331
    332
    333
    334

cudnn_deconv_layer.cpp

#ifdef USE_CUDNN
#include <algorithm>
#include <vector>

#include "caffe/layers/cudnn_deconv_layer.hpp"

namespace caffe {

// Set to three for the benefit of the backward pass, which
// can use separate streams for calculating the gradient w.r.t.
// bias, filter weights, and bottom data for each group independently
#define CUDNN_STREAMS_PER_GROUP 3

/**
 * TODO(dox) explain cuDNN interface
 */
template <typename Dtype>
void CuDNNDeconvolutionLayer<Dtype>::LayerSetUp(
    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  DeconvolutionLayer<Dtype>::LayerSetUp(bottom, top);
  // Initialize CUDA streams and cuDNN.
  stream_         = new cudaStream_t[this->group_ * CUDNN_STREAMS_PER_GROUP];
  handle_         = new cudnnHandle_t[this->group_ * CUDNN_STREAMS_PER_GROUP];

  // Initialize algorithm arrays
  fwd_algo_       = new cudnnConvolutionFwdAlgo_t[bottom.size()];
  bwd_filter_algo_= new cudnnConvolutionBwdFilterAlgo_t[bottom.size()];
  bwd_data_algo_  = new cudnnConvolutionBwdDataAlgo_t[bottom.size()];

  // initialize size arrays
  workspace_fwd_sizes_ = new size_t[bottom.size()];
  workspace_bwd_filter_sizes_ = new size_t[bottom.size()];
  workspace_bwd_data_sizes_ = new size_t[bottom.size()];

  // workspace data
  workspaceSizeInBytes = 0;
  workspaceData = NULL;
  workspace = new void*[this->group_ * CUDNN_STREAMS_PER_GROUP];

  for (size_t i = 0; i < bottom.size(); ++i) {
    // initialize all to default algorithms
    fwd_algo_[i] = (cudnnConvolutionFwdAlgo_t)0;
    bwd_filter_algo_[i] = (cudnnConvolutionBwdFilterAlgo_t)0;
    bwd_data_algo_[i] = (cudnnConvolutionBwdDataAlgo_t)0;
    // default algorithms don't require workspace
    workspace_fwd_sizes_[i] = 0;
    workspace_bwd_data_sizes_[i] = 0;
    workspace_bwd_filter_sizes_[i] = 0;
  }

  for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
    CUDA_CHECK(cudaStreamCreate(&stream_[g]));
    CUDNN_CHECK(cudnnCreate(&handle_[g]));
    CUDNN_CHECK(cudnnSetStream(handle_[g], stream_[g]));
    workspace[g] = NULL;
  }

  // Set the indexing parameters.
  bias_offset_ = (this->num_output_ / this->group_);

  // Create filter descriptor.
  const int* kernel_shape_data = this->kernel_shape_.cpu_data();
  const int kernel_h = kernel_shape_data[0];
  const int kernel_w = kernel_shape_data[1];
  cudnn::createFilterDesc<Dtype>(&filter_desc_,
                                 this->channels_ / this->group_,
                                 this->num_output_ / this->group_,
                                 kernel_h,
                                 kernel_w);

  // Create tensor descriptor(s) for data and corresponding convolution(s).
  for (int i = 0; i < bottom.size(); i++) {
    cudnnTensorDescriptor_t bottom_desc;
    cudnn::createTensor4dDesc<Dtype>(&bottom_desc);
    bottom_descs_.push_back(bottom_desc);
    cudnnTensorDescriptor_t top_desc;
    cudnn::createTensor4dDesc<Dtype>(&top_desc);
    top_descs_.push_back(top_desc);
    cudnnConvolutionDescriptor_t conv_desc;
    cudnn::createConvolutionDesc<Dtype>(&conv_desc);
    conv_descs_.push_back(conv_desc);
  }

  // Tensor descriptor for bias.
  if (this->bias_term_) {
    cudnn::createTensor4dDesc<Dtype>(&bias_desc_);
  }

  handles_setup_ = true;
}

template <typename Dtype>
void CuDNNDeconvolutionLayer<Dtype>::Reshape(
    const vector<Blob<Dtype>*>& bottom, const vector<Blob<Dtype>*>& top) {
  DeconvolutionLayer<Dtype>::Reshape(bottom, top);
  CHECK_EQ(2, this->num_spatial_axes_)
      << "CuDNNDeconvolutionLayer input must have 2 spatial axes "
      << "(e.g., height and width). "
      << "Use 'engine: CAFFE' for general ND convolution.";
  bottom_offset_ = this->bottom_dim_ / this->group_;
  top_offset_ = this->top_dim_ / this->group_;
  const int height = bottom[0]->shape(this->channel_axis_ + 1);
  const int width = bottom[0]->shape(this->channel_axis_ + 2);
  const int height_out = top[0]->shape(this->channel_axis_ + 1);
  const int width_out = top[0]->shape(this->channel_axis_ + 2);
  const int* pad_data = this->pad_.cpu_data();
  const int pad_h = pad_data[0];
  const int pad_w = pad_data[1];
  const int* stride_data = this->stride_.cpu_data();
  const int stride_h = stride_data[0];
  const int stride_w = stride_data[1];
  #if  CUDNN_VERSION_MIN(8, 0, 0)
  int RetCnt;
  bool found_conv_algorithm;
  size_t free_memory, total_memory;
  cudnnConvolutionFwdAlgoPerf_t     fwd_algo_pref_[4];
  cudnnConvolutionBwdDataAlgoPerf_t bwd_data_algo_pref_[4];
 
  //get memory sizes
  cudaMemGetInfo(&free_memory, &total_memory);
  #else
  // Specify workspace limit for kernels directly until we have a
  // planning strategy and a rewrite of Caffe's GPU memory mangagement
  size_t workspace_limit_bytes = 8*1024*1024;
  #endif
  for (int i = 0; i < bottom.size(); i++) {
    cudnn::setTensor4dDesc<Dtype>(&bottom_descs_[i],
                                  this->num_,
                                  this->channels_ / this->group_,
                                  height,
                                  width,
                                  this->channels_ * height * width,
                                  height * width,
                                  width,
                                  1);
    cudnn::setTensor4dDesc<Dtype>(&top_descs_[i],
                                  this->num_,
                                  this->num_output_ / this->group_,
                                  height_out,
                                  width_out,
                                  this->num_output_ * height_out * width_out,
                                  height_out * width_out,
                                  width_out,
                                  1);
    cudnn::setConvolutionDesc<Dtype>(&conv_descs_[i],
                                     top_descs_[i],
                                     filter_desc_,
                                     pad_h,
                                     pad_w,
                                     stride_h,
                                     stride_w);
                     
    #if  CUDNN_VERSION_MIN(8, 0, 0)
    // choose forward algorithm for filter
    // in forward filter the CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED is not implemented in cuDNN 8
    CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm_v7(handle_[0],
      top_descs_[i],
      filter_desc_,
      conv_descs_[i],
      bottom_descs_[i],
      4,
      &RetCnt,
      fwd_algo_pref_));
 
    found_conv_algorithm = false;
    for(int n=0;n<RetCnt;n++){
      if (fwd_algo_pref_[n].status == CUDNN_STATUS_SUCCESS &&
          fwd_algo_pref_[n].algo != CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED &&
          fwd_algo_pref_[n].memory < free_memory){
        found_conv_algorithm = true;
        fwd_algo_[i]                   = fwd_algo_pref_[n].algo;
        workspace_fwd_sizes_[i]        = fwd_algo_pref_[n].memory;
        break;
      }
    }
    if(!found_conv_algorithm) LOG(ERROR) << "cuDNN did not return a suitable algorithm for convolution.";
    else{
        // choose backward algorithm for filter
        // for better or worse, just a fixed constant due to the missing
        // cudnnGetConvolutionBackwardFilterAlgorithm in cuDNN version 8.0
        bwd_filter_algo_[i] = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0;
        //twice the amount of the forward search to be save
        workspace_bwd_filter_sizes_[i] = 2*workspace_fwd_sizes_[i];
    }
 
    // choose backward algo for data
    CUDNN_CHECK(cudnnGetConvolutionBackwardDataAlgorithm_v7(handle_[0],
      filter_desc_,
      bottom_descs_[i],
      conv_descs_[i],
      top_descs_[i],
      4,
      &RetCnt,
      bwd_data_algo_pref_));
    
    found_conv_algorithm = false;
    for(int n=0;n<RetCnt;n++){
      if (bwd_data_algo_pref_[n].status == CUDNN_STATUS_SUCCESS &&
          bwd_data_algo_pref_[n].algo != CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD &&
          bwd_data_algo_pref_[n].algo != CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED &&
          bwd_data_algo_pref_[n].memory < free_memory){
        found_conv_algorithm = true;
        bwd_data_algo_[i]              = bwd_data_algo_pref_[n].algo;
        workspace_bwd_data_sizes_[i]   = bwd_data_algo_pref_[n].memory;
        break;
      }
    }
    if(!found_conv_algorithm) LOG(ERROR) << "cuDNN did not return a suitable algorithm for convolution.";
   #else
    // choose forward and backward algorithms + workspace(s)
    CUDNN_CHECK(cudnnGetConvolutionForwardAlgorithm(
        handle_[0],
        top_descs_[i],
        filter_desc_,
        conv_descs_[i],
        bottom_descs_[i],
        CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
        workspace_limit_bytes,
        &fwd_algo_[i]));

    // We have found that CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM is
    // buggy. Thus, if this algo was chosen, choose winograd instead. If
    // winograd is not supported or workspace is larger than threshold, choose
    // implicit_gemm instead.
    if (fwd_algo_[i] == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM) {
      size_t winograd_workspace_size;
      cudnnStatus_t status = cudnnGetConvolutionForwardWorkspaceSize(
          handle_[0],
          top_descs_[i],
          filter_desc_,
          conv_descs_[i],
          bottom_descs_[i],
          CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD,
          &winograd_workspace_size);
      if (status != CUDNN_STATUS_SUCCESS ||
          winograd_workspace_size >= workspace_limit_bytes) {
        fwd_algo_[i] = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM;
      } else {
        fwd_algo_[i] = CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD;
      }
    }

    CUDNN_CHECK(cudnnGetConvolutionForwardWorkspaceSize(
        handle_[0],
        top_descs_[i],
        filter_desc_,
        conv_descs_[i],
        bottom_descs_[i],
        fwd_algo_[i],
        &(workspace_fwd_sizes_[i])));

    // choose backward algorithm for filter
    CUDNN_CHECK(cudnnGetConvolutionBackwardFilterAlgorithm(
        handle_[0],
        top_descs_[i],
        bottom_descs_[i],
        conv_descs_[i],
        filter_desc_,
        CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
        workspace_limit_bytes,
        &bwd_filter_algo_[i]));

    // get workspace for backwards filter algorithm
    CUDNN_CHECK(cudnnGetConvolutionBackwardFilterWorkspaceSize(
        handle_[0],
        top_descs_[i],
        bottom_descs_[i],
        conv_descs_[i],
        filter_desc_,
        bwd_filter_algo_[i],
        &workspace_bwd_filter_sizes_[i]));

    // choose backward algo for data
    CUDNN_CHECK(cudnnGetConvolutionBackwardDataAlgorithm(
        handle_[0],
        filter_desc_,
        bottom_descs_[i],
        conv_descs_[i],
        top_descs_[i],
        CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
        workspace_limit_bytes,
        &bwd_data_algo_[i]));

    // get workspace size
    CUDNN_CHECK(cudnnGetConvolutionBackwardDataWorkspaceSize(
        handle_[0],
        filter_desc_,
        bottom_descs_[i],
        conv_descs_[i],
        top_descs_[i],
        bwd_data_algo_[i],
        &workspace_bwd_data_sizes_[i]));
    #endif
  }

  // reduce over all workspace sizes to get a maximum to allocate / reallocate
  size_t total_workspace_fwd = 0;
  size_t total_workspace_bwd_data = 0;
  size_t total_workspace_bwd_filter = 0;

  for (size_t i = 0; i < bottom.size(); i++) {
    total_workspace_fwd        = std::max(total_workspace_fwd,
                                     workspace_fwd_sizes_[i]);
    total_workspace_bwd_data   = std::max(total_workspace_bwd_data,
                                     workspace_bwd_data_sizes_[i]);
    total_workspace_bwd_filter = std::max(total_workspace_bwd_filter,
                                     workspace_bwd_filter_sizes_[i]);
  }
  // get max over all operations
  size_t max_workspace = std::max(total_workspace_fwd,
                             total_workspace_bwd_data);
  max_workspace = std::max(max_workspace, total_workspace_bwd_filter);
  // ensure all groups have enough workspace
  size_t total_max_workspace = max_workspace *
                               (this->group_ * CUDNN_STREAMS_PER_GROUP);

  // this is the total amount of storage needed over all groups + streams
  if (total_max_workspace > workspaceSizeInBytes) {
    DLOG(INFO) << "Reallocating workspace storage: " << total_max_workspace;
    workspaceSizeInBytes = total_max_workspace;

    // free the existing workspace and allocate a new (larger) one
    cudaFree(this->workspaceData);

    cudaError_t err = cudaMalloc(&(this->workspaceData), workspaceSizeInBytes);
    if (err != cudaSuccess) {
      // force zero memory path
      for (int i = 0; i < bottom.size(); i++) {
        workspace_fwd_sizes_[i] = 0;
        workspace_bwd_filter_sizes_[i] = 0;
        workspace_bwd_data_sizes_[i] = 0;
        fwd_algo_[i] = CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING;
        bwd_filter_algo_[i] = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0;
        bwd_data_algo_[i] = CUDNN_CONVOLUTION_BWD_DATA_ALGO_0;
      }

      // NULL out all workspace pointers
      for (int g = 0; g < (this->group_ * CUDNN_STREAMS_PER_GROUP); g++) {
        workspace[g] = NULL;
      }
      // NULL out underlying data
      workspaceData = NULL;
      workspaceSizeInBytes = 0;
    }

    // if we succeed in the allocation, set pointer aliases for workspaces
    for (int g = 0; g < (this->group_ * CUDNN_STREAMS_PER_GROUP); g++) {
      workspace[g] = reinterpret_cast<char *>(workspaceData) + g*max_workspace;
    }
  }

  // Tensor descriptor for bias.
  if (this->bias_term_) {
    cudnn::setTensor4dDesc<Dtype>(
        &bias_desc_, 1, this->num_output_ / this->group_, 1, 1);
  }
}

template <typename Dtype>
CuDNNDeconvolutionLayer<Dtype>::~CuDNNDeconvolutionLayer() {
  // Check that handles have been setup before destroying.
  if (!handles_setup_) { return; }

  for (int i = 0; i < bottom_descs_.size(); i++) {
    cudnnDestroyTensorDescriptor(bottom_descs_[i]);
    cudnnDestroyTensorDescriptor(top_descs_[i]);
    cudnnDestroyConvolutionDescriptor(conv_descs_[i]);
  }
  if (this->bias_term_) {
    cudnnDestroyTensorDescriptor(bias_desc_);
  }
  cudnnDestroyFilterDescriptor(filter_desc_);

  for (int g = 0; g < this->group_ * CUDNN_STREAMS_PER_GROUP; g++) {
    cudaStreamDestroy(stream_[g]);
    cudnnDestroy(handle_[g]);
  }

  cudaFree(workspaceData);
  delete [] workspace;
  delete [] stream_;
  delete [] handle_;
  delete [] fwd_algo_;
  delete [] bwd_filter_algo_;
  delete [] bwd_data_algo_;
  delete [] workspace_fwd_sizes_;
  delete [] workspace_bwd_data_sizes_;
  delete [] workspace_bwd_filter_sizes_;
}

INSTANTIATE_CLASS(CuDNNDeconvolutionLayer);

}   // namespace caffe
#endif

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15
    16
    17
    18
    19
    20
    21
    22
    23
    24
    25
    26
    27
    28
    29
    30
    31
    32
    33
    34
    35
    36
    37
    38
    39
    40
    41
    42
    43
    44
    45
    46
    47
    48
    49
    50
    51
    52
    53
    54
    55
    56
    57
    58
    59
    60
    61
    62
    63
    64
    65
    66
    67
    68
    69
    70
    71
    72
    73
    74
    75
    76
    77
    78
    79
    80
    81
    82
    83
    84
    85
    86
    87
    88
    89
    90
    91
    92
    93
    94
    95
    96
    97
    98
    99
    100
    101
    102
    103
    104
    105
    106
    107
    108
    109
    110
    111
    112
    113
    114
    115
    116
    117
    118
    119
    120
    121
    122
    123
    124
    125
    126
    127
    128
    129
    130
    131
    132
    133
    134
    135
    136
    137
    138
    139
    140
    141
    142
    143
    144
    145
    146
    147
    148
    149
    150
    151
    152
    153
    154
    155
    156
    157
    158
    159
    160
    161
    162
    163
    164
    165
    166
    167
    168
    169
    170
    171
    172
    173
    174
    175
    176
    177
    178
    179
    180
    181
    182
    183
    184
    185
    186
    187
    188
    189
    190
    191
    192
    193
    194
    195
    196
    197
    198
    199
    200
    201
    202
    203
    204
    205
    206
    207
    208
    209
    210
    211
    212
    213
    214
    215
    216
    217
    218
    219
    220
    221
    222
    223
    224
    225
    226
    227
    228
    229
    230
    231
    232
    233
    234
    235
    236
    237
    238
    239
    240
    241
    242
    243
    244
    245
    246
    247
    248
    249
    250
    251
    252
    253
    254
    255
    256
    257
    258
    259
    260
    261
    262
    263
    264
    265
    266
    267
    268
    269
    270
    271
    272
    273
    274
    275
    276
    277
    278
    279
    280
    281
    282
    283
    284
    285
    286
    287
    288
    289
    290
    291
    292
    293
    294
    295
    296
    297
    298
    299
    300
    301
    302
    303
    304
    305
    306
    307
    308
    309
    310
    311
    312
    313
    314
    315
    316
    317
    318
    319
    320
    321
    322
    323
    324
    325
    326
    327
    328
    329
    330
    331
    332
    333
    334
    335
    336
    337
    338
    339
    340
    341
    342
    343
    344
    345
    346
    347
    348
    349
    350
    351
    352
    353
    354
    355
    356
    357
    358
    359
    360
    361
    362
    363
    364
    365
    366
    367
    368
    369
    370
    371
    372
    373
    374
    375
    376
    377
    378
    379
    380
    381
    382
    383
    384
    385
    386
    387
    388
    389
    390
    391
    392
    393
    394

cuDNN版本指定

由于cuDNN对代码进行了改版,在cudnn.h文件中不再指出cudnn的版本号,而是放在了cudnn_version.h文件中,这样的指明版本方式caffe完全不买账,你不魔改一下子他就给你error警告。所以,将cudnn_version.h中对于版本段的代码复制到cudnn.h文件中,代码如下:

#ifndef CUDNN_VERSION_H_
#define CUDNN_VERSION_H_

#define CUDNN_MAJOR 8
#define CUDNN_MINOR 2
#define CUDNN_PATCHLEVEL 1

#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)

#endif /* CUDNN_VERSION_H */

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10

后将cudnn.h复制到/usr/include/目录下caffe才能找到并进行编译:

rm /usr/include/cudnn.h
cp /usr/local/cuda-11.0/include/cudnn.h /usr/include/

    1
    2

libboost_python.so链接版本修改

系统中libboost_python.so默认链接的是libboost_python-py27.so,而我们编译需要的是python3版本的libboost_python.so,修改命令如下:

cd /usr/lib/x86_64-linux-gnu/
rm libboost_python.so
ln -s libboost_python-py36.so libboost_python.so

    1
    2
    3

开始编译caffe

回到根目录

cd caffe
mkdir build/
cd build/
apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler
apt-get install --no-install-recommends libboost-all-dev
apt-get install python-dev
apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev
apt-get install libatlas-base-dev
apt-get install the python-matplotlib python-scipy python-numpy
pip3.7 install boost cmake pytest numpy
cmake ..
make all
make pycaffe
make install
make runtest

    1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    12
    13
    14
    15

调用caffe

先安装一些调用caffe用到的库:

pip3.7 install scikit-image
pip3.7 install google
pip3.7 install protobuf

    1
    2
    3

再在代码块中指定caffe的路径:

import sys
sys.path.insert(0,'/caffe/python')
import caffe

    1
    2
    3

调用时记得使用python3.7调用caffe,不然会报错,指令为 python3.7 a_name.py

开始愉快的caffe调用之旅吧!
————————————————

                            版权声明:本文为博主原创文章,遵循 CC 4.0 BY-SA 版权协议,转载请附上原文出处链接和本声明。
                        
原文链接:https://blog.csdn.net/weixin_39161727/article/details/120136500

标签:cuDNN8,CUDNN,filter,python3.7,Caffe,algo,workspace,bwd,data
From: https://www.cnblogs.com/eastgeneral/p/18222700

相关文章

  • 深度学习:Caffe之经典模型详解与实战 (乐毅 王斌)
    链接:pan.baidu.com/s/1tIHXj9HmIYojAHqje09DTA?pwd=jqso提取码:jqsoCaffe框架介绍:简要介绍Caffe框架的起源、发展及其在深度学习领域的重要地位,阐述其模块化、可扩展性和易用性的设计原则。深度学习与Caffe基础:概述深度学习的基本概念、原理以及Caffe框架的核心组件,包括网络层......
  • 环境搭建 ubuntu22.04+gtx1070+cuda12.0+cudnn8.8.0+TensorRT8.6
    构建基础             cuda12.0的.deb包会强制安装所依赖的nvidia-525.60.13版本驱动,但是对于ubuntu22.04来说,linux内核为6.5.0,其与该nvidia驱动不兼容,会报错,所以要先安装所支持的驱动,然后再使用runfile进行安装cuda12.0。cuda与驱动版本对应可查如下官网:1.C......
  • python3.7.4 RV1126 交叉编译
    首先,十分感谢大佬分享的交叉编译攻略,原文链接如下:https://www.cnblogs.com/Se7eN-HOU/p/16736164.html在此基础上,本人经过一周时间的踩坑,终于在RV1126开发板上实现了交叉编译。现记录本人实际编译版本与上述版本略有不同之处。主要是由于3.5.2版本的python和openssl1.0.2g在......
  • SpringBoot整合Caffeine本地缓存
    一、Caffeine性能二、Caffeine配置注意:1、weakValues和softValues不可以同时使用。2、maximumSize和maximumWeight不可以同时使用。3、expireAfterWrite和expireAfterAccess同事存在时,以expireAfterWrite为准。三、软引用和弱引用软引用:如果一个对象只具有......
  • 通过注解实现本地缓存caffeine的学习
    注解源码如下1@Target(ElementType.METHOD)2@Retention(RetensionPolicy.RUNTIME)3@Documented4public@interfaceRvcCache{5Strngkey();6Stringid()defaultStringUtils.EMPTY;7}1@Component2@Aspect3@RequiredArgsConstructor4......
  • Caffeine 介绍
    Caffeine是一个高性能Java缓存库,提供了近乎最佳的命中率,它是GuavaCache的升级版;本文主要介绍它的相关功能及基本使用,文中所使用到的软件版本:Java1.8.0_341、Caffeine2.9.3。1、简介Caffeine和ConcurrentMap有点相似,但还是有所区别。最根本的区别是ConcurrentMap将会持......
  • Caffeine 介绍
    Caffeine是一个高性能Java缓存库,提供了近乎最佳的命中率,它是GuavaCache的升级版;本文主要介绍它的相关功能及基本使用,文中所使用到的软件版本:Java1.8.0_341、Caffeine2.9.3。1、简介Caffeine和ConcurrentMap有点相似,但还是有所区别。最根本的区别是ConcurrentMap将......
  • 性能利器Caffeine缓存全面指南
    第1章:引言大家好,我是小黑,今天咱们聊聊Caffeine缓存,小黑在网上购物,每次查看商品都要等几秒钟,那体验肯定不咋地。但如果用了缓存,常见的商品信息就像放在口袋里一样,随时取用,速度自然就快多了。这就是缓存的魔力,它通过存储临时数据,减少数据库的重复读写,提升系统的响应速度和性能。......
  • Spring Boot 中使用Caffeine缓存的简单例子
    Caffeine缓存是Java的高性能缓存库。本文简单记录下Caffeine缓存的用法。依赖配置<dependencies><dependency><groupId>org.springframework.boot</groupId><artifactId>spring-boot-starter-cache</artifactId></dependency&g......
  • 记录 | ubuntu源码编译python3.7.3(指定版本)
    一、安装依赖包sudoapt-getinstall-ymakebuild-essentiallibssl-devzlib1g-devsudoapt-getinstall-ylibbz2-devlibreadline-devlibsqlite3-devwgetcurlllvmsudoapt-getinstall-ylibncurses5-devlibncursesw5-devxz-utilstk-dev二、从Python网页中找......