在Altera SoC DE1板卡上跑完整的卷积神经网络
这次为大家详细展示一个利用卷积神经网络实现图片自动分类的例程。
神经网络的优点:自动从数据中学习经验知识,无需复(👙)杂的模型和算法。
缺点:有监督学习(💻),需要大量的带(🚄)标签数据;参数量太少时容易过拟合,泛化能力差,参数量太大时(😘)训练收敛(🌲)很慢(有(😏)可能需要几个月到几年)。
为了克服上述缺点,人们发掘了各(📰)种计算资源,包括多核CPU、GPU、DSP、ASIC、FPGA,甚至使用(🤑)模拟电路。
使用CPU实现卷积神经(🕗)网络比较方便调试,但性能太差,一般人(🤹)们都选用更快的(🦑)GPU实现。目前开源的框架大多都支持GPU,如伯克利大学Caffe和Google Convnet。
微软在2015年2月宣布使(🖨)用Stratix V完成了CNN加速器,处理 CIFAR10 图片速(🐧)度可达每秒2300多张。
这里我们也使用CIFAR10图片数据,在Cyclone V板子上跑一个卷积神经网络CNN demo。由于板子上计算资源太少(DSP Slice只有80多个),实现完整的网络不(🥗)太现实,只能在FPGA上实现基本计算单(🚹)元,然后由HPS统一调度。性能预期不会太高,后面给出。
CIFAR10图片都是什么呢?先来张图!
有兴趣的朋友可以到官网下载(CIFAR10官网)。上(🏈)面提到过,CNN是有监督学习系统(📱),需要大量带label的数据,CIFAR10就(🍃)是这样一个开放的数据库,提供(🉑)了60000张(🤩)不同类别的图片,分为10个类(如上图左侧(⚓)所示),每个类别有600张图。这个数(👡)据集不算特别大,适合在嵌入式平台上实现。而更大的数据集有ImageNet-1000(ImageNet官网),拥有120多万张高清无码大图,我下载到硬盘,占(📮)用了近200GB空间(只能忍痛将其他rmvb和avi删掉(❣)了)!
有朋友会问,不用这(🥑)些数据行不行,我们的智能手机(🚼)里面照片能不能用于CNN做训练?
答案(🍅)是可以的,只是你的数据集很不(💶)“均匀”,采样不(🔤)够“完备”,训练出的模型是真实模型的“有偏估计”,而上述两个数据集经过了种种考验,已经是学术界(🍊)公认的优质数据集,一年一(🧔)度的ILSVRC比赛就采用了这些数据集。
说完数据,再说模型。先来看一张经典的CNN结构:
这是世界上第一个将CNN实用化的例子,实现了手写体字母自动识别。在这个CNN模型中,可以看到输入是一张32 x 32的二维图像,经过卷积(💎)层(Convolution)、下采样层(Subsampling,也称Pooling)、全连接层(Full Connection,也称Inner Product)后,得到一组(💬)概率密度,我们选其中概率最大的元素作为该模型对输入图像的分类结果。所以(➿)实现CNN时,只需要实现三种基本算法:卷积、(🎙)下采样、矩(🔵)阵乘。除此之外,每层输出都可选择是否经过(🙄)非线(😒)性变换,常用的非线性变换有ReLU和Sigmoid,前者计算较为简单,使用较为广泛。
Caffe框架中提供了专门为cifar10数据定制的模型,是用proto格式写的,我们的demo也基于这个模型。内容如下:
name: "CIFAR10_quick_test"input: "data"input_dim: 1input_dim: 3input_dim: 32input_dim: 32layers {name: "conv1"type: CONVOLUTIONbottom: "data"top: "conv1"blobs_lr: 1blobs_lr: 2convolution_param {num_output: 32pad: 2kernel_size: 5stride: 1}}layers {name: "pool1"type: POOLINGbottom: "conv1"top: "pool1"pooling_param {pool: MAXkernel_size: 3stride: 2}}layers {name: "relu1"type: RELUbottom: "pool1"top: "pool1"}layers {name: "conv2"type: CONVOLUTIONbottom: "pool1"top: "conv2"blobs_lr: 1blobs_lr: 2convolution_param {num_output: 32pad: 2kernel_size: 5stride: 1}}layers {name: "relu2"type: RELUbottom: "conv2"top: "conv2"}layers {name: "pool2"type: POOLINGbottom: "conv2"top: "pool2"pooling_param {pool: AVEkernel_size: 3stride: 2}}layers {name: "conv3"type: CONVOLUTIONbottom: "pool2"top: "conv3"blobs_lr: 1blobs_lr: 2convolution_param {num_output: 64pad: 2kernel_size: 5stride: 1}}layers {name: "relu3"type: RELUbottom: "conv3"top: "conv3"}layers {name: "pool3"type: POOLINGbottom: "conv3"top: "pool3"pooling_param {pool: AVEkernel_size: 3stride: 2}}layers {name: "ip1"type: INNER_PRODUCTbottom: "pool3"top: "ip1"blobs_lr: 1blobs_lr: 2inner_product_param {num_output: 64}}layers {name: "ip2"type: INNER_PRODUCTbottom: "ip1"top: "ip2"blobs_lr: 1blobs_lr: 2inner_product_param {num_output: 10}}layers {name: "prob"type: SOFTMAXbottom: "ip2"top: "prob"}复制代码
可(🌐)见,上述模型经过了3个卷(🖲)积层(conv1, conv2, conv3),每个卷积层后面都跟着下(🍭)采样层(🤛)(pool1, pool2, pool3),之后有两个全连接层(ip1, ip2),最后一层(💤)prob为SOFTMAX分类层,是计算概率密度的,这里我们不需要(🏴)关心。
下面三张图分别统计了CNN模型各层的参数量、数据量和计算量。
可以看出,卷积层的参数量很少,但数据量很大;全连接层刚好相反,参数量(🗑)较大,但数据量很少。
通过计算量(🈚)统计发现conv2计算量最大,其次是conv3和conv1。全连接层的计算量相(🎢)对卷积层较小,但不可忽略。其他层(pool1, pool2以及各级relu)由于计算量(🤖)太小,本设计中没有将其实现为Open CL kernel,而是直接CPU端实现。
综上所述,我们重点实(🔮)现两个算法:卷积和矩阵乘,分别对应卷积层、全连接层的实现。
在DE1-SOC上我利用了友晶提供的Open CL BSP,支持C语言开发FPGA。
卷积层计算kernel函数如下:
__attribute__((num_compute_units(4)))__kernelvoid conv(__global float * a, __global float * b, __global float * c, const int M, const int N, const int K){int gx = get_global_id(0);int gy = get_global_id(1);float tmp=0.0f;for(int x = 0; x < K; x ++){for(int y = 0; y < K; y ++){tmp += a[(gx + x) * M + (gy + y)] * b[x * K + y];}}复制代码
全连接层计(🍃)算采用矩阵乘实现,kernel函数如下:
__attribute__((num_compute_units(4)))__kernelvoid gemm(__global float * a, __global float * b, __global float * c, const int M, const int N, const int K){int gx = get_global_id(0);int gy = get_global_id(1);int sy = get_global_size(1);int sx = get_global_size(0);int s = sx * sy;for(int x = gx; x < M; x += sx){for(int y = gy; y < N; y += sy){float tmp=0.0f;for(int z = 0; z < K; z++){tmp += a[z * M + x] * b[y * K + z];}c[y * M + x] = tmp;}}}复制代码
编译kernel函数需要使用Altera SDK for OpenCL,我用的版本是14.0.0.200,申请了两个月的license。编译使用命令行(🧥)aoc,得到*.aocx文件。
Open CL编译输出报告中给(🤖)出了资源占用情况:
+--------------------------------------------------------------------+; Estimated Resource Usage Summary ;+----------------------------------------+---------------------------+; Resource + Usage ;+----------------------------------------+---------------------------+; Logic utilization ; 83% ;; Dedicated logic registers ; 46% ;; Memory blocks ; 57% ;; DSP blocks ; 25% ;+----------------------------------------+---------------------------;复制代码
可见,逻辑(😚)资源、存储器资源消耗较为明显,而DSP资源并未用尽,说明(📛)还有优化的空(🍩)间。
编译主程序需要使用SoCEDS,我用的版本为14.0.2.274,也是命令行(🌶)方式,在工程目(🈹)录下执行make,结束后得到可(🚌)执行文件cnn。
将这两个文件拷贝到SD卡,按(🥒)照前面的博客对(👎)板子进行设置,将CNN的模型、CIFAR10数据也拷贝到SD卡中,板子上(💷)电,mount SD卡到(🌸)/mnt,执行cnn,得到输出如下:
<div class="blockcode"><blockquote>Please input the number of images(1~100):100Loading data...OK!Constructing CNN...OK!Begin calculation...Elapsed Time = 141.861 s.Real Label = 3(cat), Calc Label = 3(cat), error count = 0Real Label = 8(ship), Calc Label = 8(ship), error count = 0Real Label = 8(ship), Calc Label = 8(ship), error count = 0Real Label = 0(airplane), Calc Label = 0(airplane), error count = 0Real Label = 6(frog), Calc Label = 6(frog), error count = 0Real Label = 6(frog), Calc Label = 6(frog), error count = 0Real Label = 1(automobile), Calc Label = 1(automobile), error count = 0Real Label = 6(frog), Calc Label = 6(frog), error count = 0Real Label = 3(cat), Calc Label = 3(cat), error count = 0Real Label = 1(automobile), Calc Label = 1(automobile), error count = 0Real Label = 0(airplane), Calc Label = 0(airplane), error count = 0Real Label = 9(truck), Calc Label = 9(truck), error count = 0Real Label = 5(dog), Calc Label = 5(dog), error count = 0Real Label = 7(horse), Calc Label = 7(horse), error count = 0Real Label = 9(truck), Calc Label = 9(truck), error count = 0Real Label = 8(ship), Calc Label = 8(ship), error count = 0Real Label = 5(dog), Calc Label = 5(dog), error count = 0Real Label = 7(horse), Calc Label = 7(horse), error count = 0Real Label = 8(ship), Calc Label = 8(ship), error count = 0Real Label = 6(frog), Calc Label = 6(frog), error count = 0Real Label = 7(horse), Calc Label = 7(horse), error count = 0Real Label = 0(airplane), Calc Label = 2(bird), error count = 1Real Label = 4(deer), Calc Label = 4(deer), error count = 1Real Label = 9(truck), Calc Label = 9(truck), error count = 1Real Label = 5(dog), Calc Label = 4(deer), error count = 2Real Label = 2(bird), Calc Label = 3(cat), error count = 3Real Label = 4(deer), Calc Label = 4(deer), error count = 3Real Label = 0(airplane), Calc Label = 0(airplane), error count = 3Real Label = 9(truck), Calc Label = 9(truck), error count = 3Real Label = 6(frog), Calc Label = 6(frog), error count = 3Real Label = 6(frog), Calc Label = 6(frog), error count = 3Real Label = 5(dog), Calc Label = 5(dog), error count = 3Real Label = 4(deer), Calc Label = 4(deer), error count = 3Real Label = 5(dog), Calc Label = 5(dog), error count = 3Real Label = 9(truck), Calc Label = 9(truck), error count = 3Real Label = 2(bird), Calc Label = 3(cat), error count = 4Real Label = 4(deer), Calc Label = 7(horse), error count = 5Real Label = 1(automobile), Calc Label = 9(truck), error count = 6Real Label = 9(truck), Calc Label = 9(truck), error count = 6Real Label = 5(dog), Calc Label = 5(dog), error count = 6Real Label = 4(deer), Calc Label = 4(deer), error count = 6Real Label = 6(frog), Calc Label = 6(frog), error count = 6Real Label = 5(dog), Calc Label = 5(dog), error count = 6Real Label = 6(frog), Calc Label = 6(frog), error count = 6Real Label = 0(airplane), Calc Label = 0(airplane), error count = 6Real Label = 9(truck), Calc Label = 9(truck), error count = 6Real Label = 3(cat), Calc Label = 5(dog), error count = 7Real Label = 9(truck), Calc Label = 9(truck), error count = 7Real Label = 7(horse), Calc Label = 7(horse), error count = 7Real Label = 6(frog), Calc Label = 6(frog), error count = 7Real Label = 9(truck), Calc Label = 9(truck), error count = 7Real Label = 8(ship), Calc Label = 8(ship), error count = 7Real Label = 0(airplane), Calc Label = 2(bird), error count = 8Real Label = 3(cat), Calc Label = 3(cat), error count = 8Real Label = 8(ship), Calc Label = 8(ship), error count = 8Real Label = 8(ship), Calc Label = 8(ship), error count = 8Real Label = 7(horse), Calc Label = 7(horse), error count = 8Real Label = 7(horse), Calc Label = 7(horse), error count = 8Real Label = 4(deer), Calc Label = 3(cat), error count = 9Real Label = 6(frog), Calc Label = 3(cat), error count = 10Real Label = 7(horse), Calc Label = 7(horse), error count = 10Real Label = 3(cat), Calc Label = 5(dog), error count = 11Real Label = 6(frog), Calc Label = 6(frog), error count = 11Real Label = 3(cat), Calc Label = 3(cat), error count = 11Real Label = 6(frog), Calc Label = 6(frog), error count = 11Real Label = 2(bird), Calc Label = 2(bird), error count = 11Real Label = 1(automobile), Calc Label = 1(automobile), error count = 11Real Label = 2(bird), Calc Label = 2(bird), error count = 11Real Label = 3(cat), Calc Label = 3(cat), error count = 11Real Label = 7(horse), Calc Label = 9(truck), error count = 12Real Label = 2(bird), Calc Label = 2(bird), error count = 12Real Label = 6(frog), Calc Label = 6(frog), error count = 12Real Label = 8(ship), Calc Label = 8(ship), error count = 12Real Label = 8(ship), Calc Label = 8(ship), error count = 12Real Label = 0(airplane), Calc Label = 0(airplane), error count = 12Real Label = 2(bird), Calc Label = 2(bird), error count = 12Real Label = 9(truck), Calc Label = 0(airplane), error count = 13Real Label = 3(cat), Calc Label = 3(cat), error count = 13Real Label = 3(cat), Calc Label = 2(bird), error count = 14Real Label = 8(ship), Calc Label = 8(ship), error count = 14Real Label = 8(ship), Calc Label = 8(ship), error count = 14Real Label = 1(automobile), Calc Label = 1(automobile), error count = 14Real Label = 1(automobile), Calc Label = 1(automobile), error count = 14Real Label = 7(horse), Calc Label = 7(horse), error count = 14Real Label = 2(bird), Calc Label = 2(bird), error count = 14Real Label = 5(dog), Calc Label = 7(horse), error count = 15Real Label = 2(bird), Calc Label = 2(bird), error count = 15Real Label = 7(horse), Calc Label = 7(horse), error count = 15Real Label = 8(ship), Calc Label = 8(ship), error count = 15Real Label = 9(truck), Calc Label = 9(truck), error count = 15Real Label = 0(airplane), Calc Label = 0(airplane), error count = 15Real Label = 3(cat), Calc Label = 4(deer), error count = 16Real Label = 8(ship), Calc Label = 8(ship), error count = 16Real Label = 6(frog), Calc Label = 6(frog), error count = 16Real Label = 4(deer), Calc Label = 4(deer), error count = 16Real Label = 6(frog), Calc Label = 6(frog), error count = 16Real Label = 6(frog), Calc Label = 6(frog), error count = 16Real Label = 0(airplane), Calc Label = 2(bird), error count = 17Real Label = 0(airplane), Calc Label = 0(airplane), error count = 17Real Label = 7(horse), Calc Label = 7(horse), error count = 17Classify Score = 83 %.上面的执行(🥩)流程是这样的,首先(🚫)输入测试样本数目(1到100),由于DE1板子FPGA端SDRAM容量较小(🎗),难以加载全部测试数据(10000张图片),故每次最多装入100张图片。之后载(❤)入数据到HPS内存,然后开始构建CNN模型,构建过(🐇)程中也实现了Open CL的初始化。构建完毕,将输入图像依次通过CNN,得到一系列分类结果,与标签进行对比,统计错误分(⚡)类个数,计算分类准确率。
经过测试,分类准确率达到83%,与Caffe测试结(🐠)果一致。
经过以上测试,可以得到结论:
(1)使用Open CL可以很方便地移植高级语言编写的算法;
(2)CNN在移植过程中需要考虑实际(🎑)硬件,定制合适的模型和数据;
(3)Cyclone 5逻辑资源较少(85K,Open CL kernel占用了(🥈)83%),如果希望进一步提高(🈴)计(🤞)算速度,一方面可以选用高性能器件(如Stratix V、(🔑)Arria 10),另一方面可以使用RTL自己搭建计算系统。
以上图文内容均是EEWORLD论坛(📙)网友zhaoyongke原创(👠),在此(💴)感谢。
欢迎微博(🛋)@EEWORLD
如果你也(🆎)写过此类原创干货请关注微信订阅号(ID:eeworldbbs,将(🔳)你的原创发至:bbs_service@eeworld.com.cn,一经入(😒)选,我们(🆕)将帮你登上头条!
与更多行业内网友进行交流请登陆(🔏)EEWORLD论坛。
【久操Av在线的相关新闻】