说明：
通过使用LeNet网络模型来训练mnist数据集，分别使用直接读入图片方式和通过lmdb读入两种方式来训练LeNet。训练结果准确率可达到99.2%。
步骤：
1.通过直接读入图片方式
vim train_mnist.py
# -*- coding: utf-8 -*- """ yuandanfei Editor 输入ImageData格式 """ import caffe from caffe import layers as L, params as P, proto, to_proto #设定文件路径 root = '/home/yuandanfei/work/caffe/mnist2/' #根目录路径 train_list = root + 'mnist/train/train.txt' #训练图片列表 test_list = root + 'mnist/test/test.txt' #测试图片列表 train_proto = root + 'train.prototxt' #训练网络文件 test_proto = root + 'test.prototxt' #测试网络文件 solver_proto = root + 'solver.prototxt' #参数配置文件 #Lenet网络模型 def Lenet(img_list, batch_size, include_acc=False): #数据层0: n*3*28*28; 输入ImageData格式 data, label = L.ImageData(source=img_list, batch_size=batch_size, ntop=2, root_folder=root, transform_param=dict(scale=0.00390625) ) #卷积层1: n*20*24*24; c1=num_output;w1/h1=(w0/h0+2*pad-kernel_size)/stride+1 conv1 = L.Convolution(data, kernel_size=5, stride=1, pad=0, num_output=20, weight_filler=dict(type='xavier')) #池化层1: n*20*12*12; c1=c0;w1/h1=(w0/h0+2*pad-kernel_size)/stride+1 pool1 = L.Pooling(conv1, pool=P.Pooling.MAX, kernel_size=2, stride=2) #卷积层2: n*50*8*8 conv2 = L.Convolution(pool1, kernel_size=5, stride=1, pad=0, num_output=50, weight_filler=dict(type='xavier')) #池化层2: n*50*4*4 pool2 = L.Pooling(conv2, pool=P.Pooling.MAX, kernel_size=2, stride=2) #全连接层3: n*500*1*1 fc3 = L.InnerProduct(pool2, num_output=500, weight_filler=dict(type='xavier')) #激活层3: n*500*1*1 relu3 = L.ReLU(fc3, in_place=True) #全连接层4: n*10*1*1 fc4 = L.InnerProduct(relu3, num_output=10, weight_filler=dict(type='xavier')) #损失率层5 loss = L.SoftmaxWithLoss(fc4, label) #准确率层5 if include_acc: #测试网络 acc = L.Accuracy(fc4, label) return to_proto(loss, acc) else: #训练网络 return to_proto(loss) #写入网络模型文件 def write_Lenet(): #写入训练网络文件 with open(train_proto, 'w') as f: f.write( str(Lenet(train_list, batch_size=64)) ) #写入测试网络文件 with open(test_proto, 'w') as f: f.write( str(Lenet(test_list, batch_size=100, include_acc=True)) ) #写入参数配置文件 def write_solver(solver_file, train_net, test_net): s = proto.caffe_pb2.SolverParameter() s.train_net = train_net #训练网络路径 s.test_net.append(test_net) #测试网络路径 s.test_interval = 938 #测试间隔次数: 938 = 60000/64 s.test_iter.append(100) #测试迭代次数: 100 = 10000/100 s.max_iter = 9380 #最大迭代次数: 9380 = 938*10 s.lr_policy = 'step' #学习率变化规则: learning_rate = base_lr*gamma^(floor(iter/stepsize)) s.base_lr = 0.01 #基础学习率 s.gamma = 0.1 #学习率变化指数 s.stepsize = 3127 #学习率变化频率: 3127 = 9380/3 s.momentum = 0.9 #学习率动量 s.weight_decay = 5e-4 #权重衰减率 s.display = 938 #屏幕显示间隔: 938 = 9380/10 s.snapshot = 4690 #保存权重间隔: 4690 = 9380/2 s.snapshot_prefix = root + 'mnist/lenet' #权重路径前缀 s.type = 'SGD' #优化算法 s.solver_mode = proto.caffe_pb2.SolverParameter.GPU #计算方式 #写入参数配置文件 with open(solver_file, 'w') as f: f.write( str(s) ) def train_model(): caffe.set_device(0) caffe.set_mode_gpu() solver = caffe.SGDSolver(solver_proto) solver.solve() if __name__ == '__main__': write_Lenet() write_solver(solver_proto, train_proto, test_proto) train_model()
2.通过lmdb读入图片方式
vim train_mnist.py
# -*- coding: utf-8 -*- """ yuandanfei Editor 输入LMDB格式 """ import caffe from caffe import layers as L, params as P, proto, to_proto #设定文件路径 root = '/home/yuandanfei/work/caffe/mnist2/' #根目录路径 train_lmdb = root + 'mnist/train_lmdb' #训练图片路径 test_lmdb = root + 'mnist/test_lmdb' #测试图片路径 train_proto = root + 'train.prototxt' #训练网络文件 test_proto = root + 'test.prototxt' #测试网络文件 solver_proto = root + 'solver.prototxt' #参数配置文件 #Lenet网络模型 def Lenet(lmdb, batch_size, include_acc=False): #数据层0: n*3*28*28; 输入LMDB格式 data, label = L.Data(source=lmdb, backend=P.Data.LMDB, batch_size=batch_size, ntop=2, transform_param=dict(scale=0.00390625) ) #卷积层1: n*20*24*24; c1=num_output;w1/h1=(w0/h0+2*pad-kernel_size)/stride+1 conv1 = L.Convolution(data, kernel_size=5, stride=1, pad=0, num_output=20, weight_filler=dict(type='xavier')) #池化层1: n*20*12*12; c1=c0;w1/h1=(w0/h0+2*pad-kernel_size)/stride+1 pool1 = L.Pooling(conv1, pool=P.Pooling.MAX, kernel_size=2, stride=2) #卷积层2: n*50*8*8 conv2 = L.Convolution(pool1, kernel_size=5, stride=1, pad=0, num_output=50, weight_filler=dict(type='xavier')) #池化层2: n*50*4*4 pool2 = L.Pooling(conv2, pool=P.Pooling.MAX, kernel_size=2, stride=2) #全连接层3: n*500*1*1 fc3 = L.InnerProduct(pool2, num_output=500, weight_filler=dict(type='xavier')) #激活层3: n*500*1*1 relu3 = L.ReLU(fc3, in_place=True) #全连接层4: n*10*1*1 fc4 = L.InnerProduct(relu3, num_output=10, weight_filler=dict(type='xavier')) #损失率层5 loss = L.SoftmaxWithLoss(fc4, label) #准确率层5 if include_acc: #测试网络 acc = L.Accuracy(fc4, label) return to_proto(loss, acc) else: #训练网络 return to_proto(loss) #写入网络模型文件 def write_Lenet(): #写入训练网络文件 with open(train_proto, 'w') as f: f.write( str(Lenet(train_lmdb, batch_size=64)) ) #写入测试网络文件 with open(test_proto, 'w') as f: f.write( str(Lenet(test_lmdb, batch_size=100, include_acc=True)) ) #写入参数配置文件 def write_solver(solver_file, train_net, test_net): s = proto.caffe_pb2.SolverParameter() s.train_net = train_net #训练网络路径 s.test_net.append(test_net) #测试网络路径 s.test_interval = 938 #测试间隔次数: 938 = 60000/64 s.test_iter.append(100) #测试迭代次数: 100 = 10000/100 s.max_iter = 9380 #最大迭代次数: 9380 = 938*10 s.lr_policy = 'step' #学习率变化规则: learning_rate = base_lr*gamma^(floor(iter/stepsize)) s.base_lr = 0.01 #基础学习率 s.gamma = 0.1 #学习率变化指数 s.stepsize = 3127 #学习率变化频率: 3127 = 9380/3 s.momentum = 0.9 #学习率动量 s.weight_decay = 5e-4 #权重衰减率 s.display = 938 #屏幕显示间隔: 938 = 9380/10 s.snapshot = 4690 #保存权重间隔: 4690 = 9380/2 s.snapshot_prefix = root + 'mnist/lenet' #权重模型路径 s.type = 'SGD' #优化算法 s.solver_mode = proto.caffe_pb2.SolverParameter.GPU #计算方式 #写入参数配置文件 with open(solver_file, 'w') as f: f.write( str(s) ) def train_model(): caffe.set_device(0) caffe.set_mode_gpu() solver = caffe.SGDSolver(solver_proto) solver.solve() if __name__ == '__main__': write_Lenet() write_solver(solver_proto, train_proto, test_proto) train_model()
参考资料：
https://www.cnblogs.com/denny402/p/5684431.html