[深度应用]·实战掌握PyTorch图片分类简明教程
个人网站--> http://www.yansongsong.cn/
项目GitHub地址--> https://github.com/xiaosongshine/image_classifier_PyTorch/
1.引文
深度学习的比赛中,图片分类是很常见的比赛,同时也是很难取得特别高名次的比赛,因为图片分类已经被大家研究的很透彻,一些开源的网络很容易取得高分。如果大家还掌握不了使用开源的网络进行训练,再慢慢去模型调优,很难取得较好的成绩。
我们在[PyTorch小试牛刀]实战六·准备自己的数据集用于训练讲解了如何制作自己的数据集用于训练,这个教程在此基础上,进行训练与应用。
2.数据介绍
数据 下载地址
这次的实战使用的数据是交通标志数据集,共有62类交通标志。其中训练集数据有4572张照片(每个类别大概七十个),测试数据集有2520张照片(每个类别大概40个)。数据包含两个子目录分别train与test:
为什么还需要测试数据集呢?这个测试数据集不会拿来训练,是用来进行模型的评估与调优。
train与test每个文件夹里又有62个子文件夹,每个类别在同一个文件夹内:
我从中打开一个文件间,把里面图片展示出来:
其中每张照片都类似下面的例子,100*100*3的大小。100是照片的照片的长和宽,3是什么呢?这其实是照片的色彩通道数目,RGB。彩色照片存储在计算机里就是以三维数组的形式。我们送入网络的也是这些数组。
3.网络构建
1.导入Python包,定义一些参数
import torch as t import torchvision as tv import os import time import numpy as np from tqdm import tqdm class DefaultConfigs(object): data_dir = "./traffic-sign/" data_list = ["train","test"] lr = 0.001 epochs = 10 num_classes = 62 image_size = 224 batch_size = 40 channels = 3 gpu = "0" train_len = 4572 test_len = 2520 use_gpu = t.cuda.is_available() config = DefaultConfigs()
2.数据准备,采用PyTorch提供的读取方式(具体内容参考[PyTorch小试牛刀]实战六·准备自己的数据集用于训练)
注意一点Train数据需要进行随机裁剪,Test数据不要进行裁剪了
normalize = tv.transforms.Normalize(mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225] ) transform = { config.data_list[0]:tv.transforms.Compose( [tv.transforms.Resize([224,224]),tv.transforms.CenterCrop([224,224]), tv.transforms.ToTensor(),normalize]#tv.transforms.Resize 用于重设图片大小 ) , config.data_list[1]:tv.transforms.Compose( [tv.transforms.Resize([224,224]),tv.transforms.ToTensor(),normalize] ) } datasets = { x:tv.datasets.ImageFolder(root = os.path.join(config.data_dir,x),transform=transform[x]) for x in config.data_list } dataloader = { x:t.utils.data.DataLoader(dataset= datasets[x], batch_size=config.batch_size, shuffle=True ) for x in config.data_list }
3.构建网络模型(使用resnet18进行迁移学习,训练参数为最后一个全连接层 t.nn.Linear(512,num_classes))
def get_model(num_classes): model = tv.models.resnet18(pretrained=True) for parma in model.parameters(): parma.requires_grad = False model.fc = t.nn.Sequential( t.nn.Dropout(p=0.3), t.nn.Linear(512,num_classes) ) return(model)
如果电脑硬件支持,可以把下述代码屏蔽,则训练整个网络,最终准确率会上升,训练数据会变慢。
for parma in model.parameters(): parma.requires_grad = False
模型输出
ResNet( (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) (layer1): Sequential( (0): BasicBlock( (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) (1): BasicBlock( (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (layer2): Sequential( (0): BasicBlock( (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (downsample): Sequential( (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False) (1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (1): BasicBlock( (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (layer3): Sequential( (0): BasicBlock( (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (downsample): Sequential( (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False) (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (1): BasicBlock( (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (layer4): Sequential( (0): BasicBlock( (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (downsample): Sequential( (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False) (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (1): BasicBlock( (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) (relu): ReLU(inplace) (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True) ) ) (avgpool): AvgPool2d(kernel_size=7, stride=1, padding=0) (fc): Sequential( (0): Dropout(p=0.3) (1): Linear(in_features=512, out_features=62, bias=True) ) )
4.训练模型(支持自动GPU加速,GPU使用教程参考:[开发技巧]·PyTorch如何使用GPU加速)
def train(epochs): model = get_model(config.num_classes) print(model) loss_f = t.nn.CrossEntropyLoss() if(config.use_gpu): model = model.cuda() loss_f = loss_f.cuda() opt = t.optim.Adam(model.fc.parameters(),lr = config.lr) time_start = time.time() for epoch in range(epochs): train_loss = [] train_acc = [] test_loss = [] test_acc = [] model.train(True) print("Epoch {}/{}".format(epoch+1,epochs)) for batch, datas in tqdm(enumerate(iter(dataloader["train"]))): x,y = datas if (config.use_gpu): x,y = x.cuda(),y.cuda() y_ = model(x) #print(x.shape,y.shape,y_.shape) _, pre_y_ = t.max(y_,1) pre_y = y #print(y_.shape) loss = loss_f(y_,pre_y) #print(y_.shape) acc = t.sum(pre_y_ == pre_y) loss.backward() opt.step() opt.zero_grad() if(config.use_gpu): loss = loss.cpu() acc = acc.cpu() train_loss.append(loss.data) train_acc.append(acc) #if((batch+1)%5 ==0): time_end = time.time() print("Batch {}, Train loss:{:.4f}, Train acc:{:.4f}, Time: {}" .format(batch+1,np.mean(train_loss)/config.batch_size,np.mean(train_acc)/config.batch_size,(time_end-time_start))) time_start = time.time() model.train(False) for batch, datas in tqdm(enumerate(iter(dataloader["test"]))): x,y = datas if (config.use_gpu): x,y = x.cuda(),y.cuda() y_ = model(x) #print(x.shape,y.shape,y_.shape) _, pre_y_ = t.max(y_,1) pre_y = y #print(y_.shape) loss = loss_f(y_,pre_y) acc = t.sum(pre_y_ == pre_y) if(config.use_gpu): loss = loss.cpu() acc = acc.cpu() test_loss.append(loss.data) test_acc.append(acc) print("Batch {}, Test loss:{:.4f}, Test acc:{:.4f}".format(batch+1,np.mean(test_loss)/config.batch_size,np.mean(test_acc)/config.batch_size)) t.save(model,str(epoch+1)+"ttmodel.pkl") if __name__ == "__main__": train(config.epochs)
训练结果如下:
def train(epochs): model = get_model(config.num_classes) print(model) loss_f = t.nn.CrossEntropyLoss() if(config.use_gpu): model = model.cuda() loss_f = loss_f.cuda() opt = t.optim.Adam(model.fc.parameters(),lr = config.lr) time_start = time.time() for epoch in range(epochs): train_loss = [] train_acc = [] test_loss = [] test_acc = [] model.train(True) print("Epoch {}/{}".format(epoch+1,epochs)) for batch, datas in tqdm(enumerate(iter(dataloader["train"]))): x,y = datas if (config.use_gpu): x,y = x.cuda(),y.cuda() y_ = model(x) #print(x.shape,y.shape,y_.shape) _, pre_y_ = t.max(y_,1) pre_y = y #print(y_.shape) loss = loss_f(y_,pre_y) #print(y_.shape) acc = t.sum(pre_y_ == pre_y) loss.backward() opt.step() opt.zero_grad() if(config.use_gpu): loss = loss.cpu() acc = acc.cpu() train_loss.append(loss.data) train_acc.append(acc) #if((batch+1)%5 ==0): time_end = time.time() print("Batch {}, Train loss:{:.4f}, Train acc:{:.4f}, Time: {}" .format(batch+1,np.mean(train_loss)/config.batch_size,np.mean(train_acc)/config.batch_size,(time_end-time_start))) time_start = time.time() model.train(False) for batch, datas in tqdm(enumerate(iter(dataloader["test"]))): x,y = datas if (config.use_gpu): x,y = x.cuda(),y.cuda() y_ = model(x) #print(x.shape,y.shape,y_.shape) _, pre_y_ = t.max(y_,1) pre_y = y #print(y_.shape) loss = loss_f(y_,pre_y) acc = t.sum(pre_y_ == pre_y) if(config.use_gpu): loss = loss.cpu() acc = acc.cpu() test_loss.append(loss.data) test_acc.append(acc) print("Batch {}, Test loss:{:.4f}, Test acc:{:.4f}".format(batch+1,np.mean(test_loss)/config.batch_size,np.mean(test_acc)/config.batch_size)) t.save(model,str(epoch+1)+"ttmodel.pkl") if __name__ == "__main__": train(config.epochs)
训练10个Epoch,测试集准确率可以到达0.86,已经达到不错效果。通过修改参数,增加训练,可以达到更高的准确率。