手写数字识别Caffe实现

用lmdb训练的模型准确率很低，可能是lmdb出了问题，这里是直接使用图片进行训练的%99准确率
生成配置文件的时候特别要注意路径问题
下面是我的几个文件夹的位置关系和路径
wang是放着所有的数据和配置文件和脚本
wang/mnist是存放训练和测试数据两个文件夹
wang/mnist/train:训练数据
wang/mnist/test : 测试数据
配置文件和文件列表都在mnist文件夹下面
下面是脚本和数据和配置文件的目录结构图

　

转换数据图像为lmdb格式

生成图片文件列表txt

# 生成图片列表的txt文件，image_path:图像文件夹的位置
# save_path:保存图像列表txt文件的位置
import os
def create_flist(image_path, save_path):with open(save_path, 'a') as f:label_list = os.listdir(image_path)for label in label_list:filenames = os.listdir(image_path+"/"+label)for fname in filenames:f.write(label+'/'+fname+' '+label+'\n')

create_flist('mnist/train', 'mnist/train.txt')
create_flist('mnist/test', 'mnist/test.txt')

生成lmdb

caffe/bulid/tools/convert_imageset转换图像，需要四个参数

”’
-gray: 是否以灰度图的方式打开,调用opencv库中的imread()函数，默认为false
-shuffle: 是否随机打乱图片顺序。默认为false
-backend:需要转换成的db文件格式，可选为leveldb或lmdb,默认为lmdb
-resize_width/resize_height: 改变图片的大小。调用opencv库的resize（）
-check_size: 检查所有的数据是否有相同的尺寸。默认为false,不检查
-encoded: 是否将原图片编码放入最终的数据中，默认为false
-encode_type: 与前一个参数对应，将图片编码为哪一个格式：‘png’,’jpg’……
”’

import commands
# lmdb_name:生成的db文件的名字
# images_path:图片文件的路径（相对的）
# txt_save_path：图片列表文件的位置
def create_db(caffe_root, images_path, txt_save_path, lmdb_name):#生成的db文件的保存目录lmdb_save_path = '/home/jinghui/wang/mnist/' + lmdb_name#convert_imageset工具路径convert_imageset_path = caffe_root + 'build/tools/convert_imageset'cmd = """%s --shuffle --resize_height=256 --resize_width=256 %s %s %s"""status, output = commands.getstatusoutput(cmd % (convert_imageset_path, images_path, txt_save_path, lmdb_save_path))print outputif(status == 0):print "lmbd文件生成成功"

# create train.lmdb test.lmdb
caffe_root = '/home/jinghui/caffe/'# ipynb跟mnist实在同一文件夹下的，
# 文件路径和图片列表路径使用相对路径也可以
create_db(caffe_root, 'mnist/train/','mnist/train.txt','train.lmdb')

I0314 04:55:31.386880  2936 convert_imageset.cpp:86] Shuffling data
I0314 04:55:32.726089  2936 convert_imageset.cpp:89] A total of 60000 images.
I0314 04:55:32.726358  2936 db_lmdb.cpp:35] Opened lmdb /home/hao/wang/mnist/train.lmdb
I0314 04:55:36.342209  2936 convert_imageset.cpp:147] Processed 1000 files.
I0314 04:55:38.617094  2936 convert_imageset.cpp:147] Processed 2000 files.
I0314 04:55:40.841969  2936 convert_imageset.cpp:147] Processed 3000 files.
I0314 04:55:43.125208  2936 convert_imageset.cpp:147] Processed 4000 files.
I0314 04:55:45.316761  2936 convert_imageset.cpp:147] Processed 5000 files.
I0314 04:55:47.574946  2936 convert_imageset.cpp:147] Processed 6000 files.
I0314 04:55:49.799901  2936 convert_imageset.cpp:147] Processed 7000 files.
I0314 04:55:51.999716  2936 convert_imageset.cpp:147] Processed 8000 files.
I0314 04:55:54.232971  2936 convert_imageset.cpp:147] Processed 9000 files.
I0314 04:55:56.441159  2936 convert_imageset.cpp:147] Processed 10000 files.
I0314 04:55:58.674425  2936 convert_imageset.cpp:147] Processed 11000 files.
I0314 04:56:00.949287  2936 convert_imageset.cpp:147] Processed 12000 files.
I0314 04:56:03.115809  2936 convert_imageset.cpp:147] Processed 13000 files.
I0314 04:56:05.340706  2936 convert_imageset.cpp:147] Processed 14000 files.
I0314 04:56:07.615679  2936 convert_imageset.cpp:147] Processed 15000 files.
I0314 04:56:09.882153  2936 convert_imageset.cpp:147] Processed 16000 files.
I0314 04:56:12.082021  2936 convert_imageset.cpp:147] Processed 17000 files.
I0314 04:56:14.173527  2936 convert_imageset.cpp:147] Processed 18000 files.
I0314 04:56:16.390081  2936 convert_imageset.cpp:147] Processed 19000 files.
I0314 04:56:18.531627  2936 convert_imageset.cpp:147] Processed 20000 files.
I0314 04:56:20.773252  2936 convert_imageset.cpp:147] Processed 21000 files.
I0314 04:56:22.973078  2936 convert_imageset.cpp:147] Processed 22000 files.
I0314 04:56:25.123054  2936 convert_imageset.cpp:147] Processed 23000 files.
I0314 04:56:27.297933  2936 convert_imageset.cpp:147] Processed 24000 files.
I0314 04:56:29.564419  2936 convert_imageset.cpp:147] Processed 25000 files.
I0314 04:56:31.739270  2936 convert_imageset.cpp:147] Processed 26000 files.
I0314 04:56:33.914283  2936 convert_imageset.cpp:147] Processed 27000 files.
I0314 04:56:36.239109  2936 convert_imageset.cpp:147] Processed 28000 files.
I0314 04:56:38.555645  2936 convert_imageset.cpp:147] Processed 29000 files.
I0314 04:56:40.872205  2936 convert_imageset.cpp:147] Processed 30000 files.
I0314 04:56:43.088791  2936 convert_imageset.cpp:147] Processed 31000 files.
I0314 04:56:45.196931  2936 convert_imageset.cpp:147] Processed 32000 files.
I0314 04:56:47.371914  2936 convert_imageset.cpp:147] Processed 33000 files.
I0314 04:56:49.480028  2936 convert_imageset.cpp:147] Processed 34000 files.
I0314 04:56:51.580004  2936 convert_imageset.cpp:147] Processed 35000 files.
I0314 04:56:53.871465  2936 convert_imageset.cpp:147] Processed 36000 files.
I0314 04:56:56.154773  2936 convert_imageset.cpp:147] Processed 37000 files.
I0314 04:56:58.404657  2936 convert_imageset.cpp:147] Processed 38000 files.
I0314 04:57:00.712838  2936 convert_imageset.cpp:147] Processed 39000 files.
I0314 04:57:02.929388  2936 convert_imageset.cpp:147] Processed 40000 files.
I0314 04:57:05.179281  2936 convert_imageset.cpp:147] Processed 41000 files.
I0314 04:57:07.245843  2936 convert_imageset.cpp:147] Processed 42000 files.
I0314 04:57:09.520730  2936 convert_imageset.cpp:147] Processed 43000 files.
I0314 04:57:11.720535  2936 convert_imageset.cpp:147] Processed 44000 files.
I0314 04:57:13.895457  2936 convert_imageset.cpp:147] Processed 45000 files.
I0314 04:57:16.228660  2936 convert_imageset.cpp:147] Processed 46000 files.
I0314 04:57:18.436964  2936 convert_imageset.cpp:147] Processed 47000 files.
I0314 04:57:20.536757  2936 convert_imageset.cpp:147] Processed 48000 files.
I0314 04:57:22.695047  2936 convert_imageset.cpp:147] Processed 49000 files.
I0314 04:57:24.928246  2936 convert_imageset.cpp:147] Processed 50000 files.
I0314 04:57:27.028134  2936 convert_imageset.cpp:147] Processed 51000 files.
I0314 04:57:29.111331  2936 convert_imageset.cpp:147] Processed 52000 files.
I0314 04:57:31.302942  2936 convert_imageset.cpp:147] Processed 53000 files.
I0314 04:57:33.436180  2936 convert_imageset.cpp:147] Processed 54000 files.
I0314 04:57:35.660996  2936 convert_imageset.cpp:147] Processed 55000 files.
I0314 04:57:37.827560  2936 convert_imageset.cpp:147] Processed 56000 files.
I0314 04:57:40.077458  2936 convert_imageset.cpp:147] Processed 57000 files.
I0314 04:57:42.235694  2936 convert_imageset.cpp:147] Processed 58000 files.
I0314 04:57:44.560613  2936 convert_imageset.cpp:147] Processed 59000 files.
I0314 04:57:46.760392  2936 convert_imageset.cpp:147] Processed 60000 files.
lmbd文件生成成功

create_db(caffe_root , 'mnist/test/','mnist/test.txt','test.lmdb')

I0314 05:02:23.772184  2951 convert_imageset.cpp:86] Shuffling data
I0314 05:02:25.104369  2951 convert_imageset.cpp:89] A total of 10000 images.
I0314 05:02:25.104758  2951 db_lmdb.cpp:35] Opened lmdb /home/jinghui/wang/mnist/test.lmdb
I0314 05:02:28.693987  2951 convert_imageset.cpp:147] Processed 1000 files.
I0314 05:02:30.643874  2951 convert_imageset.cpp:147] Processed 2000 files.
I0314 05:02:32.718792  2951 convert_imageset.cpp:147] Processed 3000 files.
I0314 05:02:34.635351  2951 convert_imageset.cpp:147] Processed 4000 files.
I0314 05:02:36.643580  2951 convert_imageset.cpp:147] Processed 5000 files.
I0314 05:02:38.610074  2951 convert_imageset.cpp:147] Processed 6000 files.
I0314 05:02:40.660053  2951 convert_imageset.cpp:147] Processed 7000 files.
I0314 05:02:42.659940  2951 convert_imageset.cpp:147] Processed 8000 files.
I0314 05:02:44.843329  2951 convert_imageset.cpp:147] Processed 9000 files.
I0314 05:02:46.934710  2951 convert_imageset.cpp:147] Processed 10000 files.
lmbd文件生成成功

计算图片数据的均值文件.binaryproto

def create_meandb(caffe_tools, train_lmdb, save_meandb):compute_image_mean_path = caffe_tools+ 'compute_image_mean'cmd = """%s %s %s """status, output = commands.getstatusoutput(cmd%(compute_image_mean_path,train_lmdb, save_meandb))print outputif (status==0):print "train mean lmdb 生成完成"

create_meandb('/home/jinhui/caffe/build/tools/','mnist/train.lmdb','mnist/train.binaryproto')

I0314 06:19:05.025748  3129 db_lmdb.cpp:35] Opened lmdb mnist/train.lmdb
I0314 06:19:05.034680  3129 compute_image_mean.cpp:70] Starting iteration
I0314 06:19:16.149374  3129 compute_image_mean.cpp:95] Processed 10000 files.
I0314 06:19:27.005234  3129 compute_image_mean.cpp:95] Processed 20000 files.
I0314 06:19:37.893340  3129 compute_image_mean.cpp:95] Processed 30000 files.
I0314 06:19:41.742666  3129 compute_image_mean.cpp:95] Processed 40000 files.
I0314 06:19:43.716142  3129 compute_image_mean.cpp:95] Processed 50000 files.
I0314 06:19:45.689353  3129 compute_image_mean.cpp:95] Processed 60000 files.
I0314 06:19:45.689848  3129 compute_image_mean.cpp:108] Write to mnist/train.binaryproto
I0314 06:19:45.712528  3129 compute_image_mean.cpp:114] Number of channels: 3
I0314 06:19:45.712604  3129 compute_image_mean.cpp:119] mean_value channel [0]: 221.659
I0314 06:19:45.712698  3129 compute_image_mean.cpp:119] mean_value channel [1]: 221.659
I0314 06:19:45.712772  3129 compute_image_mean.cpp:119] mean_value channel [2]: 221.659
train mean lmdb 生成完成

生成训练/测试配置文件

from caffe import layers as L,params as P, to_proto

# 训练测试数据LMDB文件位置
train_lmdb = 'mnist/train.lmdb'
test_lmdb = 'mnist/test.lmdb'
# 训练测试的配置文件保存位置位置
train_proto = 'mnist/train.prototxt'
test_proto = 'mnist/test.prototxt'
# 训练文件均值文件的位置
train_mean = 'mnist/train.binaryproto'

# lmdb：lmdb文件的位置，batch_size:每次批训练图片的数量，include_acc:Fasle不计算准确率
def create_net(lmdb,batch_size, include_acc=False):# 输入层：同时传递，图片数据和标签，transform_param：对图像进行增强处理data, label = L.Data(source=lmdb, backend=P.Data.LMDB, batch_size=batch_size,ntop=2,transform_param=dict(crop_size=40,mean_file=train_mean,mirror=True))# 卷积层# data:图片的像素，kernelsize=卷积核的大小即：5x5,# siride:卷积核滑动的步长# num_output：输出特征图的数量（卷积核的数量)# pad : 对输入像素矩阵各边增加像素的数量# weight_filler:卷积核的初始化方式：xavierconv1 = L.Convolution(data, kernel_size=5, stride=1, num_output=16, pad=2,weight_filler=dict(type='xavier'))# 激活层# in_place:meaning the bottom and the top blob could be the same to preserve memory consumption.relu1 = L.ReLU(conv1, in_place=True)# 池化层pool1 = L.Pooling(relu1, pool=P.Pooling.MAX, kernel_size=3, stride=2)conv2 = L.Convolution(pool1, kernel_size=3, stride=1, num_output=32, pad=1,weight_filler=dict(type='xavier'))relu2 = L.ReLU(conv2, in_place=True)pool2 = L.Pooling(relu2, pool=P.Pooling.MAX, kernel_size=3, stride=2)# 全连接层fc1 = L.InnerProduct(pool2, num_output=1024, weight_filler=dict(type='xavier'))relu3 = L.ReLU(fc1, in_place=True)# dropout防止过拟合drop1 = L.Dropout(relu3, in_place=True)fc2 = L.InnerProduct(drop1, num_output=10, weight_filler=dict(type='xavier'))# softmaxloss = L.SoftmaxWithLoss(fc2, label)# 根据需要是否要计算accuracyif include_acc:acc = L.Accuracy(fc2, label)return to_proto(loss, acc)else:return to_proto(loss)

def create_prototxt(save_proto, lmdb,batch_size, acc=False):with open(save_proto, 'w') as f:f.write(str(create_net(lmdb, batch_size, acc)))

create_prototxt(train_proto, train_lmdb, 64)
create_prototxt(test_proto, test_lmdb,32, acc=True)

生成solver文件

from caffe.proto import caffe_pb2

s = caffe_pb2.SolverParameter()
solver_save_path = 'mnist/solver.prototxt'
# train_net :训练配置文件
# test_net :测试配置文件
# test_iter :迭代多少次后训练完一次所有的样本
# test_interval :测试间隔[训练多少个batch_size后测试一次]
# base_lr :初始学习效率
# display :屏幕显示间隔
# max_iter :最大迭代次数
# lr_policy :学习效率变化
# setpsize :每隔多少次迭代，调整一次lr,
# gamma : 学习效率的变化指数
# momentum: 影响lr的动量
# weight_decay: 权值衰减率
# snapshot : 保存模型的间隔
# snapshot_prefix: 保存模型的前缀
# solver_mode : 是否选择GPU
# type : 优化算法
s.train_net = 'mnist/train.prototxt'
s.test_net.append('mnist/test.prototxt')
s.test_iter.append(313)
s.test_interval = 782
s.base_lr = 0.001
s.display = 782
s.max_iter = 78200
s.lr_policy = 'step'
s.stepsize = 26067   # 调整３次lr,78200/3=26067
s.gamma = 0.1
s.momentum = 0.9
s.weight_decay = 5e-4
s.snapshot = 7820
s.snapshot_prefix = 'snapshot'
s.solver_mode = caffe_pb2.SolverParameter.GPU
s.type = 'SGD'
with open(solver_save_path, 'w') as f:f.write(str(s))

Training Model


import caffe
caffe.set_device(0)
caffe.set_mode_gpu()
solver = caffe.SGDSolver('mnist/solver.prototxt')
solver.solve()

有时间在补充

手写数字识别Caffe实现相关推荐

使用Caffe进行手写数字识别执行流程解析
之前在 http://blog.csdn.net/fengbingchun/article/details/50987185 中仿照Caffe中的examples实现对手写数字进行识别,这里详细介绍下 ...
Caffe MNIST 手写数字识别（全面流程）
目录 1.下载MNIST数据集 2.生成MNIST图片训练.验证.测试数据集 3.制作LMDB数据库文件 4.准备LeNet-5网络结构定义模型.prototxt文件 5.准备模型求解配置文件_sol ...
linux手写数字识别opencv,opencv实现KNN手写数字的识别
人工智能是当下很热门的话题,手写识别是一个典型的应用.为了进一步了解这个领域,我阅读了大量的论文,并借助opencv完成了对28x28的数字图片(预处理后的二值图像)的识别任务. 预处理一张图片: 首 ...
用MXnet实战深度学习之一：安装GPU版mxnet并跑一个MNIST手写数字识别（zz）
用MXnet实战深度学习之一:安装GPU版mxnet并跑一个MNIST手写数字识别我想写一系列深度学习的简单实战教程,用mxnet做实现平台的实例代码简单讲解深度学习常用的一些技术方向和实战样例.这 ...
深度学习数字仪表盘识别_【深度学习系列】手写数字识别实战
上周在搜索关于深度学习分布式运行方式的资料时,无意间搜到了paddlepaddle,发现这个框架的分布式训练方案做的还挺不错的,想跟大家分享一下.不过呢,这块内容太复杂了,所以就简单的介绍一下padd ...
手把手教你使用LabVIEW OpenCV DNN实现手写数字识别（含源码）
文章目录前言一.OpenCV DNN模块 1.OpenCV DNN简介 2.LabVIEW中DNN模块函数二.TensorFlow pb文件的生成和调用 1.TensorFlow2 Keras模 ...
基于深度学习的手写数字识别算法Python实现
摘要深度学习是传统机器学习下的一个分支,得益于近些年来计算机硬件计算能力质的飞跃,使得深度学习成为了当下热门之一.手写数字识别更是深度学习入门的经典案例,学习和理解其背后的原理对于深度学习的理解有 ...
深蓝学院第三章：基于卷积神经网络(CNN)的手写数字识别实践
参看之前篇章的用全连接神经网络去做手写识别:https://blog.csdn.net/m0_37957160/article/details/114105389?spm=1001.2014.3001 ...
深蓝学院第二章：基于全连接神经网络(FCNN)的手写数字识别
如何用全连接神经网络去做手写识别??? 使用的是jupyter notebook这个插件进行代码演示.(首先先装一个Anaconda的虚拟环境,然后自己构建一个自己的虚拟环境,然后在虚拟环境中安装ju ...
深度学习--TensorFlow（项目）Keras手写数字识别
目录效果展示基础理论 1.softmax激活函数 2.神经网络 3.隐藏层及神经元最佳数量一.数据准备 1.载入数据集 2.数据处理 2-1.归一化 2-2.独热编码二.神经网络拟合 1.搭建 ...

手写数字识别Caffe实现

手写数字识别Caffe实现相关推荐

最新文章

热门文章