'''Author: 365JHWZGoDescription:cnn--reviewDate: 2021/11/2 15:04FilePath: day1102-1.py'''#cnn 提取特征 卷积#导包import osimport torchimport torchvisionfrom torch.autograd import Variableimport torch.utils.data as Datatorch.manual_seed(1)#超参数设置BATCH_SIZE = 64LR = 0.01DOWNLOAD_MNIST = FalseEPOCH = 1#判断是否需要下载mnistif not os.path.exists('../mnist') or not os.listdir('../mnist'): DOWNLOAD_MNIST = True#train_datatrain_data = torchvision.datasets.MNIST( root='../mnist', transform=torchvision.transforms.ToTensor(), train=True, download=DOWNLOAD_MNIST)#train_loadertrain_loader = Data.DataLoader( dataset=train_data, shuffle=True, num_workers=2, batch_size=BATCH_SIZE)#test_datatest_data = torchvision.datasets.MNIST( root='../mnist', train=False)#创造数据# print(test_data.test_data[0].shape) #torch.Size([28, 28])# print(torch.unsqueeze(test_data.test_data[0],dim=1).shape) #torch.Size([28,1,28])x = torch.unsqueeze(test_data.test_data,dim=1).type(torch.FloatTensor)[:2000]/255.print(x.shape)y = test_data.test_labels[:2000]#cnnclass CNN(torch.nn.Module): def __init__(self): super(CNN,self).__init__() self.conv1 = torch.nn.Sequential( # 1*28*28 torch.nn.Conv2d( in_channels=1, out_channels=16, kernel_size=5, stride=1, padding=2 ), # 16*28*28 torch.nn.ReLU(), torch.nn.MaxPool2d( kernel_size=2 ) ) # 16*14*14 self.conv2 = torch.nn.Sequential( torch.nn.Conv2d( in_channels=16, out_channels=32, kernel_size=5, stride=1, padding=2 ), # 32*14*14 torch.nn.ReLU(), torch.nn.MaxPool2d( kernel_size=2 ) ) # 32*7*7 self.out = torch.nn.Linear(32*7*7,10) def forward(self,x): print('start x:',x.shape) x1 = self.conv1(x) #print('x1', x1.shape) x2 = self.conv2(x1) #print('x2', x2.shape) x3 = x2.view(x2.size(0), -1) #print('x3', x3.shape) output = self.out(x3) #print('out', output.shape) return output#实例化cnn = CNN()#创造优化器optimizer = torch.optim.Adam(cnn.parameters(),lr = LR)#损失函数loss_func = torch.nn.CrossEntropyLoss()#训练if __name__ == '__main__': for epoch in range(EPOCH): for step,(batch_x,batch_y) in enumerate(train_loader): batch_x = Variable(batch_x) batch_y = Variable(batch_y) output = cnn(batch_x) loss = loss_func(output, batch_y) optimizer.zero_grad() loss.backward() optimizer.step() if step%50==0: out_y = cnn(x) pred_y = torch.max(out_y,1)[1].data.numpy().squeeze() accuracy = sum((pred_y==y.data.numpy()).astype(int))/float(y.size(0)) print( 'EPOCH:', epoch, ' | loss:%.4f |' % loss, 'accuracy:%.3f' % accuracy )
今天再温习时，重新理解了里边的维度
datashapetrain_datatorch.Size([60000, 28, 28])test_datatorch.Size([10000, 28, 28])xtorch.Size([2000, 1, 28, 28])ytorch.Size([2000])batch_xtorch.Size([64, 1, 28, 28])batch_ytorch.Size([64])out_ytorch.Size([2000, 10])start xtorch.Size([64, 1, 28, 28])x1torch.Size([64, 16, 14, 14])x2torch.Size([64, 32, 7, 7])x3torch.Size([64, 1568])outtorch.Size([64, 10])pred_y(2000,)
从这可以看出，当执行cnn()后，进入forward()执行
也就是说传入cnn的维度是4
【batch_size，channel，width pixel，height pixel】
批处理单位，图片颜色通道，图片的宽的像素点，图片长的像素点
但是test_data中的维度是torch.Size([10000, 28, 28])，所以需要用torch.unsqueeze()添加了一个维度
而train_loader中本身出来的数据就是torch.Size([64, 1, 28, 28])，所以不需要添加维度
pred_y出来的维度是(2000,)其实也可以不用写squeeze