前言
此博客主要介绍如何利用matlab一步一步训练caffe模型,类似使用caffe.exe 的train命令。
国际惯例,参考博客:
http://caffe.berkeleyvision.org/tutorial/interfaces.html
http://www.cnblogs.com/denny402/p/5110204.html
抱怨一下:matlab的教程是真少哇,大牛们都跑去玩Python了。。。o(╯□╰)o,开更。。。。。。。。。
【注】所有专业说法请参考caffe官网以及其它大牛博客,博主写博客可能有点白话文且没那么咬文嚼字。
一、读入模型
先去caffe主页瞄一眼。。。。。得到一个讯息:
solver = caffe.Solver('./models/bvlc_reference_caffenet/solver.prototxt');
这句话干什么的呢?读模型。
尝试一下,采用大家都有的mnist 中的solver,我采用了绝对路径,读者可采用相对路径,无影响
【注】我的solver可能修改了,前面有一篇博客介绍了修改内容和原因。贴一下下载地址:
lenet_solver1.prototxt:链接:http://pan.baidu.com/s/1qXWQrhy 密码:we0e
lenet_train_test1.prototxt:链接:http://pan.baidu.com/s/1miawrxQ 密码:ghxt
均值文件:链接:http://pan.baidu.com/s/1miFDNHe 密码:48az
下面用到的mnist_data:链接:http://pan.baidu.com/s/1bp62Enl 密码:royk
Google一下,感觉可能会有两个原因导致matlab未响应:一是dll没有链接到,就跟很多人出现caffe.set_mode_gpu()会直接未响应一样;二是prototxt内部错误。我不会说我折腾了一下午这个问题。
排除第一种情况,因为目前为止,使用caffe都是比较顺利的,dll问题可能性不大。那就是prototxt 路径问题了,去看prototxt是什么情况
net: "examples/mnist/lenet_train_test1.prototxt"
snapshot_prefix: "examples/mnist/lenet"
与路径有关的两句话,我们的matlab程序文件夹是E:\CaffeDev\caffe-master\matlab\demo,与这个路径相差十万八千里。保险起见,我的解决方法是把mnist训练需要的东西全都复制丢到matlab程序文件夹了。如下:
mnist_data文件夹存的是mnist数据集的lmdb文件以及lenet.prototxt,不想动手制作的去上面下载,想动手自己做的,前面有博客介绍。
移动完毕,那就得改改prototxt里面的路径了:
lenet_solver1.prototxt
# The train/test net protocol buffer definition
net: "lenet_train_test1.prototxt"
# test_iter specifies how many forward passes the test should carry out.
# In the case of MNIST, we have test batch size 100 and 100 test iterations,
# covering the full 10,000 testing images.
test_iter: 100
# Carry out testing every 500 training iterations.
test_interval: 500
# The base learning rate, momentum and the weight decay of the network.
base_lr: 0.01
momentum: 0.9
weight_decay: 0.0005
# The learning rate policy
lr_policy: "inv"
gamma: 0.0001
power: 0.75
# Display every 100 iterations
display: 1
# The maximum number of iterations
max_iter: 10000
# snapshot intermediate results
snapshot: 5000
snapshot_prefix: "mnist_data/lenet"
# solver mode: CPU or GPU
solver_mode: CPU
lenet_train_test1.prototxt 被修改部分
name: "LeNet"
layer {name: "mnist"type: "Data"top: "data"top: "label"include {phase: TRAIN}transform_param {mean_file: "mean.binaryproto"scale: 0.00390625}data_param {source: "mnist_data/mnist_train_lmdb"batch_size: 64backend: LMDB}
}
layer {name: "mnist"type: "Data"top: "data"top: "label"include {phase: TEST}transform_param {mean_file: "mean.binaryproto"scale: 0.00390625}data_param {source: "mnist_data/mnist_test_lmdb"batch_size: 100backend: LMDB}
}
再进行下一步操作之前,最好用bat测试一下是否能读取到这个prototxt并训练,排除这一步错误才能进行下步工作。
接下来再去读取模型:addpath('..')
caffe.reset_all
solver = caffe.Solver('lenet_solver1.prototxt');
显示一下:
>> solversolver = Solver with properties:net: [1x1 caffe.Net]test_nets: [1x1 caffe.Net]
二、训练模型
2.1、一次性训练模型
solver.solve();
这里会有一个幌子,你会发现运行以后matlab跟待机一样,啥输出都没。我还以为出错,分析了一下上面solver读取的两个net:
模型的输入竟然是empty的,难道我们的lmdb数据没有读进去,然后尝试了leveldb,以及各种改leveldb的路径,比如添加“./”之类的,都不行。这时候便想到了一种可能性,模型载入是不读取数据的,只有在运行时候读取数据,但是solver的solve方法是一次性训练模型,没有任何输出,matlab可能已经在训练模型了。为了验证此想法,run→吃饭→回来→观察,果然在下面这个路径中发现了训练好的model
snapshot_prefix: "mnist_data/lenet"
为了避免这些模型是用空数据训练的,使用mnist的classification_demo测试一下,竟然手写数字都分类正确,这样便验证了我们的想法:solver.solve()是一次性训练数据,不会附带任何输出,matlab表现会如死机了。
2.2训练模型step-by-step
依旧去官网找:
意思是我们可以用step命令设置每次训练多少次以后,可以干一下别的事情,然后再训练。
好,卡壳了,设置完毕step为1表示我们想在每一次迭代都取出loss和accuracy,但是然后呢?怎么继续训练?找了很多教程都是Python的,受次启发,以及反复看caffe的官网,发现:
net.blobs('data').set_data(data);
net.forward_prefilled();
prob = net.blobs('prob').get_data();
给出的解释简单翻译一下是:
net.forward
函数接受n维的输入,输出output的blob的数据
net.forward_prefilled使用的则是使用在模型中已经存在的数据继续训练,并不接受任何输入,以及提供任何输出。
看完这两个解释,思考一下,用step设置训练1次停一下,那么我们的数据是否依旧在blob中存着呢?
【最开始想法】那么就可以使用net.forward_prefilled做继续训练的工作,尝试一下:
%训练
clear
clc
addpath('..')
caffe.reset_all
solver=caffe.Solver('lenet_solver1.prototxt');
loss=[];
accuracy=[];for i=1:10000disp('.')solver.step(1);iter=solver.iter();solver.net.forward_prefilled
end
更新日志2016-10-21
【试验之后】按照上面的想法能训练,但是突然发现,为什么不要backward_prefilled呢?而且,去掉solver.net.forward_prefilled也能训练。应该是solver.step自动包含了forward和backward过程了,因此正式使用的训练代码是:
%训练
clear
clc
addpath('..')
caffe.reset_all
solver=caffe.Solver('lenet_solver1.prototxt');
loss=[];
accuracy=[];for i=1:10000disp('.')solver.step(1);iter=solver.iter();
end
接下来就是取出每次迭代的loss和accuracy了,想都不用想,用blob,为了训练快点,我切换到GPU版本的caffe-windows去了,代码如下:
<pre name="code" class="cpp"><pre name="code" class="cpp">%训练
clear
clc
close all
format long %设置精度,caffe的损失貌似精度在小数点后面好几位
addpath('..')
caffe.reset_all%重设网络,否则载入两个网络会卡住
solver=caffe.Solver('lenet_solver1.prototxt'); %载入网络
loss=[];%记录相邻两个loss
accuracy=[];%记录相邻两个accuracy
hold on%画图用的
accuracy_init=0;
loss_init=0;
for i=1:10000solver.step(1);%每迭代一次就取一次loss和accuracyiter=solver.iter();loss=solver.net.blobs('loss').get_data();%取训练集的lossaccuracy=solver.test_nets.blobs('accuracy').get_data();%取验证集的accuracy%画loss折线图x=[i-1,i];y=[loss_init loss];plot(x,y,'r-')drawnowloss_init=loss;
end
接下来我们便得到了实时的曲线图,每次迭代都有一个loss显示在折线图中。
为了避免训练错误,测试一下
E:\CaffeDev-GPU\caffe-master\Build\x64\Release\caffe.exe test --model=lenet_train_test1.prototxt -weights=mnist_data/lenet_iter_10000.caffemodel -gpu=0
pause
结果如下:
I1021 21:00:07.988450 8132 net.cpp:261] This network produces output accuracy
I1021 21:00:07.989449 8132 net.cpp:261] This network produces output loss
I1021 21:00:07.989449 8132 net.cpp:274] Network initialization done.
I1021 21:00:07.992449 8132 caffe.cpp:253] Running for 50 iterations.
I1021 21:00:07.999449 8132 caffe.cpp:276] Batch 0, accuracy = 0.96
I1021 21:00:07.999449 8132 caffe.cpp:276] Batch 0, loss = 0.168208
I1021 21:00:08.002449 8132 caffe.cpp:276] Batch 1, accuracy = 0.95
I1021 21:00:08.002449 8132 caffe.cpp:276] Batch 1, loss = 0.152652
I1021 21:00:08.005450 8132 caffe.cpp:276] Batch 2, accuracy = 0.88
I1021 21:00:08.005450 8132 caffe.cpp:276] Batch 2, loss = 0.320218
I1021 21:00:08.007450 8132 caffe.cpp:276] Batch 3, accuracy = 0.92
I1021 21:00:08.008450 8132 caffe.cpp:276] Batch 3, loss = 0.320782
I1021 21:00:08.010450 8132 caffe.cpp:276] Batch 4, accuracy = 0.88
I1021 21:00:08.011451 8132 caffe.cpp:276] Batch 4, loss = 0.354194
I1021 21:00:08.013450 8132 caffe.cpp:276] Batch 5, accuracy = 0.91
I1021 21:00:08.013450 8132 caffe.cpp:276] Batch 5, loss = 0.604682
I1021 21:00:08.015450 8132 caffe.cpp:276] Batch 6, accuracy = 0.88
I1021 21:00:08.015450 8132 caffe.cpp:276] Batch 6, loss = 0.310961
I1021 21:00:08.017451 8132 caffe.cpp:276] Batch 7, accuracy = 0.95
I1021 21:00:08.017451 8132 caffe.cpp:276] Batch 7, loss = 0.18691
I1021 21:00:08.019450 8132 caffe.cpp:276] Batch 8, accuracy = 0.93
I1021 21:00:08.019450 8132 caffe.cpp:276] Batch 8, loss = 0.302631
I1021 21:00:08.022451 8132 caffe.cpp:276] Batch 9, accuracy = 0.96
I1021 21:00:08.022451 8132 caffe.cpp:276] Batch 9, loss = 0.10867
I1021 21:00:08.024451 8132 caffe.cpp:276] Batch 10, accuracy = 0.94
I1021 21:00:08.024451 8132 caffe.cpp:276] Batch 10, loss = 0.283927
I1021 21:00:08.026451 8132 caffe.cpp:276] Batch 11, accuracy = 0.91
I1021 21:00:08.027451 8132 caffe.cpp:276] Batch 11, loss = 0.389279
I1021 21:00:08.029451 8132 caffe.cpp:276] Batch 12, accuracy = 0.87
I1021 21:00:08.029451 8132 caffe.cpp:276] Batch 12, loss = 0.618325
I1021 21:00:08.031451 8132 caffe.cpp:276] Batch 13, accuracy = 0.91
I1021 21:00:08.031451 8132 caffe.cpp:276] Batch 13, loss = 0.464931
I1021 21:00:08.033452 8132 caffe.cpp:276] Batch 14, accuracy = 0.91
I1021 21:00:08.033452 8132 caffe.cpp:276] Batch 14, loss = 0.348089
I1021 21:00:08.035451 8132 caffe.cpp:276] Batch 15, accuracy = 0.88
I1021 21:00:08.035451 8132 caffe.cpp:276] Batch 15, loss = 0.45388
I1021 21:00:08.037451 8132 caffe.cpp:276] Batch 16, accuracy = 0.93
I1021 21:00:08.038452 8132 caffe.cpp:276] Batch 16, loss = 0.277403
I1021 21:00:08.040452 8132 caffe.cpp:276] Batch 17, accuracy = 0.9
I1021 21:00:08.040452 8132 caffe.cpp:276] Batch 17, loss = 0.48363
I1021 21:00:08.042453 8132 caffe.cpp:276] Batch 18, accuracy = 0.91
I1021 21:00:08.042453 8132 caffe.cpp:276] Batch 18, loss = 0.519036
I1021 21:00:08.044452 8132 caffe.cpp:276] Batch 19, accuracy = 0.88
I1021 21:00:08.045452 8132 caffe.cpp:276] Batch 19, loss = 0.364235
I1021 21:00:08.047452 8132 caffe.cpp:276] Batch 20, accuracy = 0.9
I1021 21:00:08.047452 8132 caffe.cpp:276] Batch 20, loss = 0.414757
I1021 21:00:08.049453 8132 caffe.cpp:276] Batch 21, accuracy = 0.9
I1021 21:00:08.049453 8132 caffe.cpp:276] Batch 21, loss = 0.387713
I1021 21:00:08.051452 8132 caffe.cpp:276] Batch 22, accuracy = 0.93
I1021 21:00:08.051452 8132 caffe.cpp:276] Batch 22, loss = 0.308721
I1021 21:00:08.053452 8132 caffe.cpp:276] Batch 23, accuracy = 0.93
I1021 21:00:08.053452 8132 caffe.cpp:276] Batch 23, loss = 0.328804
I1021 21:00:08.055454 8132 caffe.cpp:276] Batch 24, accuracy = 0.92
I1021 21:00:08.055454 8132 caffe.cpp:276] Batch 24, loss = 0.385196
I1021 21:00:08.058454 8132 caffe.cpp:276] Batch 25, accuracy = 0.93
I1021 21:00:08.058454 8132 caffe.cpp:276] Batch 25, loss = 0.255955
I1021 21:00:08.061453 8132 caffe.cpp:276] Batch 26, accuracy = 0.92
I1021 21:00:08.061453 8132 caffe.cpp:276] Batch 26, loss = 0.49177
I1021 21:00:08.063453 8132 caffe.cpp:276] Batch 27, accuracy = 0.89
I1021 21:00:08.064453 8132 caffe.cpp:276] Batch 27, loss = 0.366904
I1021 21:00:08.066453 8132 caffe.cpp:276] Batch 28, accuracy = 0.93
I1021 21:00:08.066453 8132 caffe.cpp:276] Batch 28, loss = 0.309272
I1021 21:00:08.068454 8132 caffe.cpp:276] Batch 29, accuracy = 0.88
I1021 21:00:08.068454 8132 caffe.cpp:276] Batch 29, loss = 0.520516
I1021 21:00:08.070453 8132 caffe.cpp:276] Batch 30, accuracy = 0.92
I1021 21:00:08.070453 8132 caffe.cpp:276] Batch 30, loss = 0.358098
I1021 21:00:08.072453 8132 caffe.cpp:276] Batch 31, accuracy = 0.94
I1021 21:00:08.072453 8132 caffe.cpp:276] Batch 31, loss = 0.157759
I1021 21:00:08.074455 8132 caffe.cpp:276] Batch 32, accuracy = 0.91
I1021 21:00:08.075454 8132 caffe.cpp:276] Batch 32, loss = 0.336977
I1021 21:00:08.077455 8132 caffe.cpp:276] Batch 33, accuracy = 0.95
I1021 21:00:08.077455 8132 caffe.cpp:276] Batch 33, loss = 0.116172
I1021 21:00:08.079454 8132 caffe.cpp:276] Batch 34, accuracy = 0.93
I1021 21:00:08.079454 8132 caffe.cpp:276] Batch 34, loss = 0.136695
I1021 21:00:08.081454 8132 caffe.cpp:276] Batch 35, accuracy = 0.89
I1021 21:00:08.082454 8132 caffe.cpp:276] Batch 35, loss = 0.648639
I1021 21:00:08.084455 8132 caffe.cpp:276] Batch 36, accuracy = 0.91
I1021 21:00:08.084455 8132 caffe.cpp:276] Batch 36, loss = 0.256923
I1021 21:00:08.086454 8132 caffe.cpp:276] Batch 37, accuracy = 0.93
I1021 21:00:08.086454 8132 caffe.cpp:276] Batch 37, loss = 0.321325
I1021 21:00:08.088454 8132 caffe.cpp:276] Batch 38, accuracy = 0.92
I1021 21:00:08.088454 8132 caffe.cpp:276] Batch 38, loss = 0.28317
I1021 21:00:08.090456 8132 caffe.cpp:276] Batch 39, accuracy = 0.9
I1021 21:00:08.090456 8132 caffe.cpp:276] Batch 39, loss = 0.352922
I1021 21:00:08.093456 8132 caffe.cpp:276] Batch 40, accuracy = 0.93
I1021 21:00:08.093456 8132 caffe.cpp:276] Batch 40, loss = 0.298536
I1021 21:00:08.095455 8132 caffe.cpp:276] Batch 41, accuracy = 0.88
I1021 21:00:08.095455 8132 caffe.cpp:276] Batch 41, loss = 0.817203
I1021 21:00:08.097455 8132 caffe.cpp:276] Batch 42, accuracy = 0.89
I1021 21:00:08.097455 8132 caffe.cpp:276] Batch 42, loss = 0.324021
I1021 21:00:08.100455 8132 caffe.cpp:276] Batch 43, accuracy = 0.92
I1021 21:00:08.100455 8132 caffe.cpp:276] Batch 43, loss = 0.270256
I1021 21:00:08.102455 8132 caffe.cpp:276] Batch 44, accuracy = 0.89
I1021 21:00:08.102455 8132 caffe.cpp:276] Batch 44, loss = 0.443635
I1021 21:00:08.104455 8132 caffe.cpp:276] Batch 45, accuracy = 0.92
I1021 21:00:08.104455 8132 caffe.cpp:276] Batch 45, loss = 0.316793
I1021 21:00:08.106456 8132 caffe.cpp:276] Batch 46, accuracy = 0.9
I1021 21:00:08.106456 8132 caffe.cpp:276] Batch 46, loss = 0.353561
I1021 21:00:08.109457 8132 caffe.cpp:276] Batch 47, accuracy = 0.94
I1021 21:00:08.109457 8132 caffe.cpp:276] Batch 47, loss = 0.304726
I1021 21:00:08.111456 8132 caffe.cpp:276] Batch 48, accuracy = 0.88
I1021 21:00:08.111456 8132 caffe.cpp:276] Batch 48, loss = 0.643014
I1021 21:00:08.113456 8132 caffe.cpp:276] Batch 49, accuracy = 0.93
I1021 21:00:08.113456 8132 caffe.cpp:276] Batch 49, loss = 0.214009
I1021 21:00:08.113456 8132 caffe.cpp:281] Loss: 0.355134
I1021 21:00:08.113456 8132 caffe.cpp:293] accuracy = 0.9134
I1021 21:00:08.113456 8132 caffe.cpp:293] loss = 0.355134 (* 1 = 0.355134 loss)
附:读取日志文件的loss
如果使用的是dos窗口caffe -train命令训练,那么提取loss和accuracy就需要定向到caffe默认的日志文件去,找的方法很简单
按时间排序,找到最近的以caffe.exe开头的文件名称,用notepad++打开可以看到日志信息:
读文件的方法有很多,我用正则表达式去匹配loss信息:
先将这个记录log的文件拷贝出来,
%my loss
clear;
clc;
close all;train_log_file = 'caffe.exe.BINGO-PC.Bingo.log.INFO.20160924-193528.13464' ;
train_interval = 100 ;
test_interval = 500 ;[~, string_output] = dos(['type ' , train_log_file ]) ;
pat='1 = .*? loss';
o1=regexp(string_output,pat,'start');%用'start'参数指定输出o1为匹配正则表达式的子串的起始位置
o2=regexp(string_output,pat,'end');%用'start'参数指定输出o1为匹配正则表达式的子串的结束位置
o3=regexp(string_output,pat,'match');%用'match'参数指定输出o2为匹配正则表达式的子串 loss=zeros(1,size(o1,2));
for i=1:size(o1,2)loss(i)=str2num(string_output(o1(i)+4:o2(i)-5));
end
plot(loss)