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 output
    if(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 output
    if (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：xavier
    conv1 = 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'))
    # softmax
    loss = L.SoftmaxWithLoss(fc2, label)
    # 根据需要是否要计算accuracy
    if 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