caffe手写数字识别-训练模型
为什么写第二篇呢以为第一篇有坑用lmdb格式的数据训练，可能生成的lmdb数据有问题，验证准确率很低，所以这次直接用原始数据训练，准确率%99左右，只训练了１０epoch，这个结果不算太好，我之前用Keras+TensorFlow准确率在%99.6左右
下面是一些训练过程中生成文件的截图：
train文件夹下的子文件夹，test文件夹类似
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
# type：mnist/train/ or mnist/test/
import os
def create_flist(image_path, save_path, Type):
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(Type +label+'/'+fname+' '+label+'\n')此时生成的图片文件路径txt过于整齐不利于模型学习
trianlist.txt类似
create_flist('mnist/train', 'mnist/trainlist.txt','mnist/train/')
create_flist('mnist/test', 'mnist/testlist.txt','mnist/test/')生成随机的图像路径列表
import random
def create_randomList(listfile, savepath):
with open(listfile, 'r') as f:
all_text = f.readlines()
with open(savepath, 'w') as f:
for i in xrange(len(all_text)):
index = random.randint(0,len(all_text)-1)
f.write(all_text[index])create_randomList('mnist/trainlist.txt','mnist/train.txt')
create_randomList('mnist/testlist.txt','mnist/test.txt')
生成训练/测试配置文件
from caffe import layers as L,params as P, to_proto# 训练测试数据路径文件位置
train_txt = 'mnist/train.txt'
test_txt = 'mnist/test.txt'
# 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'
# img＿list：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(img_list,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.ImageData(source=img_list,batch_size=batch_size,ntop=2,
transform_param=dict(scale=1./255))
# 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, num_output=32, 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=2, stride=2)
conv2 = L.Convolution(pool1, kernel_size=5, num_output=64, pad=1,
weight_filler=dict(type='xavier'))
relu2 = L.ReLU(conv2, in_place=True)
pool2 = L.Pooling(relu2, pool=P.Pooling.MAX, kernel_size=2, stride=2)
# 全连接层
fc1 = L.InnerProduct(pool2, num_output=500, 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,img_list,batch_size, acc=False):
with open(save_proto, 'w') as f:
f.write(str(create_net(img_list, batch_size, acc)))
create_prototxt(train_proto, 'mnist/train.txt', 100)
create_prototxt(test_proto,'mnist/test.txt',100, acc=True)生成solver文件
from caffe.proto import caffe_pb2s = 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 :迭代多少次测试完一次所有的样本
# 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(100)
s.test_interval = 600
s.base_lr = 0.01
s.display = 20
s.max_iter = 6000
s.lr_policy = 'step'
s.stepsize = 2000
s.gamma = 0.1
s.momentum = 0.9
s.weight_decay = 5e-4
s.snapshot = 2000
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()