pytorch 实现alexNet + imagenet 部分数据

文章：《ImageNet Classification with Deep Convolutional Neural Networks》

        个人认为alexNet 真正的将深度学习带热了，当时的有非常大数据库 imagenet 1000类，硬件性能也大幅提升 如gpu。简直可以成为天时地利人和。 alexNet 的效果也是不负众望，当时的imageNet 分类比赛中遥遥领先第二名。 

        alexNet 在网络结构中只要引入了relu 层 和dropout层，同时进行了Data augmentation ,都是针对大数据过拟合，从而使百万级大数据的训练成为了可能。数据的增加主要是包括 随机crop，resize ，旋转，亮度，饱和度等调节，尽可能的增加数据量，深度学习中有句话叫 “数据为王”，没有数据再好的算法也没用。

        关于relu  函数为 f(x)=max(0,x) ， 在此之前主要是通过sigmod来计算节点，该方法存在梯度消失和梯度爆炸的风险。relu能很大程度上解决这个问题，现在对relu 有好多改进， 如prelu ，lrelu。主要是针对x<0 的情况。

      关于dropout 网络已百分比丢弃节点，赋值0，目的是使节点变得稀疏，防止过拟合现象。

     alexnet 网络如下：

    

再说pytorch ，最近比较火热，一些大神都在力推。大致学习了下网络代码如下：

import torch
import torch.backends.cudnn as cudnn
import cv2
from torch.autograd import Variable
from torchvision import transforms
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import numpy as np
root="../alexnet/"

# -----------------ready the dataset--------------------------
def opencvLoad(imgPath,resizeH,resizeW):
    image = cv2.imread(imgPath)
    image = cv2.resize(image, (resizeH, resizeW), interpolation=cv2.INTER_CUBIC)
    image = image.astype(np.float32)
    image = np.transpose(image, (2, 1, 0))  
    image = torch.from_numpy(image)
    return image
    
class LoadPartDataset(Dataset):
    def __init__(self, txt):
        fh = open(txt, 'r')
        fh = open(txt, 'r')
        imgs = []
        for line in fh:
            line = line.strip('\n')
            line = line.rstrip()
            words = line.split()
            labelList = int(words[1])
            imageList = words[0]
            imgs.append((imageList, labelList))
        self.imgs = imgs
            
    def __getitem__(self, item):
        image, label = self.imgs[item]
        img = opencvLoad(image,227,227)
        return img,label
    def __len__(self):
        return len(self.imgs)
        
def loadTrainData(txt=None):
    fh = open(txt, 'r')
    imgs = []
    for line in fh:
        line = line.strip('\n')
        line = line.rstrip()
        words = line.split()
        label = int(words[1])
        image = cv2.imread(words[0])
        image = cv2.resize(image, (227, 227), interpolation=cv2.INTER_CUBIC)
        image = image.astype(np.float32)
        image = np.transpose(image, (2, 1, 0))  
        image = torch.from_numpy(image)
        imgs.append((image, label))
    return imgs

# trainSet=loadTrainData(txt=root+'train.txt')
# test_data=loadTrainData(txt=root+'train.txt')
trainSet =LoadPartDataset(txt=root+'train.txt')
test_data=LoadPartDataset(txt=root+'train.txt')
train_loader = DataLoader(dataset=trainSet, batch_size=64, shuffle=True)
test_loader = DataLoader(dataset=test_data, batch_size=64)


#-----------------create the Net and training------------------------

class Net(torch.nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = torch.nn.Sequential(
            torch.nn.Conv2d(3, 96, 11, 4, 0),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(3,2)
        )
        self.conv2 = torch.nn.Sequential(
            torch.nn.Conv2d(96, 256, 5, 1, 2),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(3,2)
        )
        self.conv3 = torch.nn.Sequential(
            torch.nn.Conv2d(256,384, 3, 1, 1),
            torch.nn.ReLU(),
        )
        self.conv4 = torch.nn.Sequential(
            torch.nn.Conv2d(384,384, 3, 1, 1),
            torch.nn.ReLU(),
        )
        self.conv5 = torch.nn.Sequential(
            torch.nn.Conv2d(384,256, 3, 1, 1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(3,2)
        )
        self.dense = torch.nn.Sequential(
            torch.nn.Linear(9216, 4096),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.5),
            torch.nn.Linear(4096, 4096),
            torch.nn.ReLU(),
            torch.nn.Dropout(0.5),
            torch.nn.Linear(4096, 50)
        )

    def forward(self, x):
        conv1_out = self.conv1(x)
        conv2_out = self.conv2(conv1_out)
        conv3_out = self.conv3(conv2_out)
        conv4_out = self.conv4(conv3_out)
        conv5_out = self.conv5(conv4_out)
        res = conv5_out.view(conv5_out.size(0), -1)
        out = self.dense(res)
        return out


model = Net()

finetune = None
finetune = r'./model/_iter_99.pth'

if finetune is not None:
    print '[0] Load Model {}'.format(finetune)

    pretrained_dict = model.state_dict()
    finetune_dict = torch.load(finetune)

    # model_dict = torch.load(finetune)
    # pretrained_dict = net.state_dict()

    model_dict = {k: v for k, v in finetune_dict.items() if k in pretrained_dict}
    pretrained_dict.update(model_dict)

    model.load_state_dict(pretrained_dict)

model = torch.nn.DataParallel(model, device_ids=[0, 1])
model.cuda()
cudnn.benchmark = True
print(model)


#optimizer = torch.optim.Adam(model.parameters())
#loss_func = torch.nn.CrossEntropyLoss()

# updata net
lr = 1e-5
loss_func = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(list(model.parameters())[:], lr=lr, momentum=0.9)

for epoch in range(10000):
    print('epoch {}'.format(epoch + 1))
    # training-----------------------------
    train_loss = 0.
    train_acc = 0.
    for trainData, trainLabel in train_loader:
        trainData, trainLabel = Variable(trainData.cuda()), Variable(trainLabel.cuda())
        out = model(trainData)
        loss = loss_func(out, trainLabel)
        train_loss += loss.data[0]
        pred = torch.max(out, 1)[1]
        train_correct = (pred == trainLabel).sum()
        train_acc += train_correct.data[0]
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
  #  if epoch % 100 == 0:
    print('Train Loss: {:.6f}, Acc: {:.6f}'.format(train_loss / (len(
        trainSet)), train_acc / (len(trainSet))))

    if (epoch + 1) % 10 == 0:
        sodir = './model/_iter_{}.pth'.format(epoch)
        print '[5] Model save {}'.format(sodir)
        torch.save(model.module.state_dict(), sodir)

    # adjust
    if (epoch + 1)% 100 == 0:
        lr = lr / 10
        optimizer = torch.optim.SGD(model.parameters(), lr=lr)

    # evaluation--------------------------------
    # model.eval()
    # eval_loss = 0.
    # eval_acc = 0.
    # for trainData, trainLabel in test_loader:
    #     trainData, trainLabel = Variable(trainData, volatile=True), Variable(trainLabel, volatile=True)
    #     out = model(trainData)
    #     loss = loss_func(out, trainData)
    #     eval_loss += loss.data[0]
    #     pred = torch.max(out, 1)[1]
    #     num_correct = (pred == trainData).sum()
    #     eval_acc += num_correct.data[0]
    # print('Test Loss: {:.6f}, Acc: {:.6f}'.format(eval_loss / (len(
    #     test_data)), eval_acc / (len(test_data))))
       

图片集太大就不上传了。

网络下载地址：https://download.youkuaiyun.com/download/eatapples/10394160