/tiny_cnn

The convolution, relu and pooling is the basic units of the neural network. This test wants you to write the function of these three parts in C/C++: · Forward only, Backward is PLUS; · Support Conv2D, Pooling2D operator; Verify the results by test case and calculate the computation efficiency;

Primary LanguageC++

tiny_CNN性能测试

Convolution, Relu 以及Pooling是卷积神经网络的基本单元，这个项目采用C++实现了这三个基本单元的Forward以及Backward计算。并且实现了将这三个单元组合为一体的模块以提高性能，组合后有1.2倍的性能提升。该项目采用OpenMP支持多核并行。

目录

  1. Introduction
  2. Results

Introduction

1. 项目结构

./src 源码

./lib 开源第三方矩阵库Eigen(底层可调用MKL)

./data mnist数据集

./python numpy实现算法demo以及Pytorch搭建网络作为benchmark

./main.cpp 测试性能代码

2. Build

mkdir build
cd build
cmake ../
make

3. 算法简介

  1. Convolution Caffe中经典的im2col算法实现

  2. Convolution和MaxPooling实现中对最外层的n_sample循环加入parallel for优化

  3. 将Convolution,Relu,Max_Pooling融合为一个模块，同样在最外层加入parallel for优化

4. 测试过程

使用mnist测试数据集输入，采用如下Pytorch定义的简化CNN网络模型(./python/mnist_cnn_pytorch.py)。

class Net(nn.Module):
   def __init__(self):
       super(Net, self).__init__()
       self.conv1 = nn.Conv2d(1, 6, 5)
       self.pool = nn.MaxPool2d(2, 2)
  

   def forward(self, x):
       x = x.view(-1, 1, 28, 28)
       x = self.pool(F.relu(self.conv1(x)))
       x = x.view(-1, 6 * 12 * 12)
       return x

⬆ 回到顶部

Results

1. 测试环境

Mac CPU(2 GHz 四核Intel Core i5)

HPC (2.5GHz 24核Core) ICC编译优化

2. Forward results

调用n次从网络输入到输出的forward计算并计时测试程序性能，结果如下表所示

n Pytorch serial fuse 24-core 24-core-fuse naive-numpy fuse_numpy
n=1 0.582s 1.042s 0.995s 0.452s 0.431s 332.457s 64.329s
n=10 4.467s 7.357s 7.850s 1.433s 1.254s - -
n=100 42.883s 69.868s 77.798s 10.548s 9.938s - -

3. Backward results

调用n次从网络输出到输入的backward计算并计时测试程序性能,结果如下表所示

n serial fuse 24-core 24-core-fuse
n=1 1.082s 0.956s 0.470s 0.363s
n=10 9.861s 9.351s 4.496s 3.478s
n=100 97.247s 90.616s 41.811s 33.760s

⬆ back to top