进入下一个epoch时准确率下降是否正常？

Question

我正在训练 CNN 以使用 MNIST 数据库预测数字。我正在做数据增强，由于某种原因，在进入下一个 epoch 时准确度急剧下降（图像中的迭代 60）

它与数据增强（代码中的转换 = my_transforms）有关，因为当我停用增强（转换 = 无）时，前进到下一个时代时准确性不会降低。但我无法解释为什么。有谁知道为什么会发生这种情况？

my_transforms = transforms.Compose([
    transforms.ToPILImage(),
    transforms.RandomCrop((25,25)),
    transforms.Resize((28,28)),
    transforms.RandomRotation(degrees=45, fill=255),
    transforms.RandomVerticalFlip(p=0.1),
    transforms.RandomHorizontalFlip(p=0.5),
    transforms.ToTensor()
    ])

dataset = MNISTDataset(transform = my_transforms)
train_loader = DataLoader(dataset = dataset, batch_size = 1000, shuffle=True)

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(1,10,kernel_size=5)
        self.pool = nn.MaxPool2d(kernel_size=2,stride=2)
        self.conv2 = nn.Conv2d(10,20,kernel_size=5)
        self.fc1 = nn.Linear(20*4*4, 64)
        self.fc2 = nn.Linear(64, 10)

def forward(self,x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1,20*4*4)
        x = F.relu(self.fc1(x))
        x = F.softmax(self.fc2(x), dim=1)

        return x

net = Net()

loss_function=nn.NLLLoss()
optimizer=optim.Adam(net.parameters())

EPOCHS=2
iteracion = 0

for epoch in range(EPOCHS):
    for data in train_loader:
        inputs, labels = data
        inputs = inputs.view(-1,1,28,28)

        net.zero_grad()

        probabilities=net(inputs)

        matches=[torch.argmax(i)==int(j) for i,j in zip(probabilities,labels)]

        in_batch_acc=matches.count(True)/len(matches)

        loss=loss_function(torch.log(probabilities), labels)

        print('Loss:', round(float(loss), 3))
        print('In-batch acc:', round(in_batch_acc, 2))

        iteracion += 1

        loss.backward()
        optimizer.step()

Answer 1

我通过数据增强复制了您的模型，并尝试绘制准确性和损失，似乎问题在于您绘制的方式。

在以下几行中，我附上了我的代码以及损失和准确度图：

代码：

# -- Imports -- #
import torch

from torch import nn, optim
from torch.utils.data import DataLoader
from torchvision import transforms, datasets
import torch.nn.functional as F
import matplotlib.pyplot as plt

# -- Data Loader -- #
my_transforms = transforms.Compose([
    transforms.ToPILImage(),
    transforms.RandomCrop((25,25)),
    transforms.Resize((28,28)),
    transforms.RandomRotation(degrees=45, fill=255),
    transforms.RandomVerticalFlip(p=0.1),
    transforms.RandomHorizontalFlip(p=0.5),
    transforms.ToTensor()
    ])
dataset = datasets.MNIST('../data', train=True, download=True,
                         transform = transforms.Compose([
                                            transforms.RandomCrop((25,25)),
                                            transforms.Resize((28,28)),
                                            transforms.RandomRotation(degrees=45, fill=255),
                                            transforms.RandomVerticalFlip(p=0.1),
                                            transforms.RandomHorizontalFlip(p=0.5),
                                            transforms.ToTensor()
                                            ]))
train_loader = DataLoader(dataset = dataset, batch_size = 1000, shuffle=True)

# -- Define Model -- #
class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(1,10,kernel_size=5)
        self.pool = nn.MaxPool2d(kernel_size=2,stride=2)
        self.conv2 = nn.Conv2d(10,20,kernel_size=5)
        self.fc1 = nn.Linear(20*4*4, 64)
        self.fc2 = nn.Linear(64, 10)

    def forward(self,x):
            x = self.pool(F.relu(self.conv1(x)))
            x = self.pool(F.relu(self.conv2(x)))
            x = x.view(-1,20*4*4)
            x = F.relu(self.fc1(x))
            x = F.softmax(self.fc2(x), dim=1)

            return x

device = 'cuda'
net = Net()
net.to(device)

loss_function=nn.NLLLoss()
optimizer=optim.Adam(net.parameters())

EPOCHS=2
iteracion = 0
accuracy = []
loss_record = []
for epoch in range(EPOCHS):
    for data in train_loader:
        inputs, labels = data
        inputs = inputs.view(-1,1,28,28)
        inputs, labels = inputs.to(device), labels.to(device)

        # -- Forward -- #
        net.zero_grad()
        probabilities=net(inputs)
        matches=[torch.argmax(i)==int(j) for i,j in zip(probabilities,labels)]
        in_batch_acc=matches.count(True)/len(matches)
        loss=loss_function(torch.log(probabilities), labels)

        # -- Statistics -- #
        accuracy.append(in_batch_acc)
        loss_record.append(loss)
        print('Loss:', round(float(loss), 3))
        print('In-batch acc:', round(in_batch_acc, 2))

        iteracion += 1

        loss.backward()
        optimizer.step()

# -- Accuracy plot -- #
iterations = range(0,120)
plt.plot(iterations, accuracy, 'g', label='Accuracy')
plt.title('Accuracy')
plt.xlabel('Iterations')
plt.ylabel('Accuracy')
plt.legend()
plt.show()

# -- Loss plot -- #
plt.plot(iterations, loss_record, label='Loss')
plt.title('Loss')
plt.xlabel('Iterations')
plt.ylabel('Loss')
plt.legend()
plt.show()

情节：

Accuracy plot

Loss plot

如你所见，(iteration = 61) -> Next epoch 时没有跳转