caffe 进行手写数字训练

标签：训练 solver proto train caffe test 手写 lmdb

案例数据准备

下载
链接：https://pan.baidu.com/s/10CmpZUdEVmma4A0mziu9dw
提取码：dmjr
复制这段内容后打开百度网盘手机App，操作更方便哦
解压后放到data/mnist

进入C:\Windows\System32\WindowsPowerShell\v1.0

管理员运行PowerShell

PS F:\caffe-windows> examples\mnist\create_mnist.ps1

生成两个目录

caffe 进行手写数字训练_.net

之后将mnist拷贝到自己的工程目录备用

运行

#coding='utf-8'
import lmdb
import caffe
from caffe.proto import caffe_pb2
from caffe import layers as L
from caffe import params as P
from matplotlib import pyplot as plt
import numpy as np


     
solver_file = './mnist/lenet_auto_solver.prototxt'
train_proto = "./mnist/lenet_auto_train.prototxt"
train_lmdb = "./mnist/mnist_train_lmdb"
test_proto = "./mnist/lenet_auto_test.prototxt"
test_lmdb = "./mnist/mnist_test_lmdb"

def lenet(lmdb,batch_size):
  n = caffe.NetSpec()
  n.data,n.label = L.Data(batch_size=batch_size,backend=P.Data.LMDB,source=lmdb,transform_param=dict(scale=1./255),ntop=2)

  n.conv1 = L.Convolution(n.data,num_output=20,kernel_size=5,weight_filler=dict(type='xavier'))
  n.pool1 = L.Pooling(n.conv1,pool=P.Pooling.MAX,kernel_size=2,stride=2)

  n.conv2 = L.Convolution(n.pool1,num_output=50,kernel_size=5,weight_filler=dict(type='xavier'))
  n.pool2 = L.Pooling(n.conv2,pool=P.Pooling.MAX,kernel_size=2,stride=2)

  n.fc1 = L.InnerProduct(n.pool2, num_output=500, weight_filler=dict(type='xavier'))
  n.relu1 = L.ReLU(n.fc1,in_place=True)

  n.score = L.InnerProduct(n.relu1, num_output=10, weight_filler=dict(type='xavier'))
  n.loss = L.SoftmaxWithLoss(n.score,n.label)

  return n.to_proto()

def gen_solver(solver_file, train_proto, test_net_file=None):
  s = caffe_pb2.SolverParameter()
  s.train_net = train_proto
  if not test_proto:
    s.test_net.append(train_proto)
  else:
    s.test_net.append(test_proto)
  s.test_interval = 500  
  s.test_iter.append(100) 
  s.display = 500
  s.max_iter = 10000   
  s.base_lr = 0.001    # 基础学习率
  s.momentum = 0.9    # momentum系数
  s.weight_decay = 5e-4    # 正则化权值衰减因子，防止过拟合

  s.lr_policy = 'step'    # 学习率衰减方法
  s.stepsize=1000    # 只对step方法有效， base_lr*gamma^floor(iter/stepsize)
  s.gamma = 0.1
  s.display = 500    # 输出日志间隔迭代次数
  s.snapshot = 5000   # 在指定迭代次数时保存模型
  s.snapshot_prefix = 'mnist/lenet'
  s.type = 'SGD'  # 迭代算法类型, ADADELTA, ADAM, ADAGRAD, RMSPROP, NESTEROV
  s.solver_mode = caffe_pb2.SolverParameter.GPU

  with open(solver_file, 'w') as f:
    f.write(str(s))




def write_data(train_proto,train_lmdb,test_proto,test_lmdb):
  with open(train_proto,'w') as f:
    f.write(str(lenet(train_lmdb,64)))

  with open(test_proto,'w') as f:
    f.write(str(lenet(test_lmdb,100)))


def main():
  write_data(train_proto,train_lmdb,test_proto,test_lmdb)
  #caffe.set_device(0)#使用GPU，我的caffe GPU配置失败，暂时使用CPU
  #caffe.set_mode_gpu()
  gen_solver(solver_file, train_proto, test_net_file=None)

  solver = None
  solver = caffe.SGDSolver(solver_file)

  #for k,v in solver.net.blobs.items():
  # print(k,v.data.shape)

  test_interval = 20
  niter = 200

  train_loss = np.zeros(niter)
  test_acc = np.zeros(int(np.ceil(niter/test_interval)))

  output = np.zeros((niter,8,10))

  for it in range(niter):
    solver.step(1)
    train_loss[it] = solver.net.blobs['loss'].data
    solver.test_nets[0].forward(start='conv1')
    output[it]  = solver.test_nets[0].blobs["score"].data[:8]

    if it%test_interval == 0:
      print("run test ing...")
      correct = 0

      for test_it in range(100):
        solver.test_nets[0].forward()
        correct+=sum(solver.test_nets[0].blobs['score'].data.argmax(1) == solver.test_nets[0].blobs['label'].data)

      test_acc[it//test_interval] = correct/1e4

  _, ax1 = plt.subplots()
  ax2 = ax1.twinx()
  ax1.plot(np.arange(niter), train_loss)
  ax2.plot(test_interval*np.arange(len(test_acc)), test_acc, 'r')
  ax1.set_xlabel('iteration')
  ax1.set_ylabel('train loss')
  ax2.set_ylabel('test accuracy')
  ax2.set_title('Test accuracy:{:.2f}'.format(test_acc[-1]))
  _.savefig('result1.png')

main()