CNN Pytorch（FashionMNIST）中的训练损失和测试损失同样高

Question

The problem is the Training loss and test loss are the same and the loss and accuracy weren't changing, what's wrong with my CNN structure and training process?

Training results:

Epoch: 1/30.. Training Loss: 2.306.. Test Loss: 2.306.. Test Accuracy: 0.100

Epoch: 2/30.. Training Loss: 2.306.. Test Loss: 2.306.. Test Accuracy: 0.100

Class code:

class Model(nn.Module):

    def __init__(self):
        super(Model, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5)
        self.conv2 = nn.Conv2d(in_channels=6, out_channels=12, kernel_size=5)
        self.fc1 = nn.Linear(in_features=12 * 4 * 4, out_features=120)
        self.fc2 = nn.Linear(in_features=120, out_features=60)
        self.out = nn.Linear(in_features=60, out_features=10)
        #the output will be 0~9 (10)

Below is my CNN and training process:

def forward(self, t):
    # implement the forward pass 
    # (1)input layer
    t = t 
    # (2) hidden conv layer
    t = self.conv1(t)
    t = F.relu(t)
    t = F.max_pool2d(t, kernel_size=2, stride=2)

    # (3) hidden conv layer
    t = self.conv2(t)
    t = F.relu(t)
    t = F.max_pool2d(t, kernel_size=2, stride=2)
    
    # (4) hidden linear layer
    t = t.reshape(-1, 12 * 4 * 4)
    t = self.fc1(t)
    t = F.relu(t)

    # (5) hidden linear layer
    t = self.fc2(t)
    t = F.relu(t)
    # (6) output layer
    t = self.out(t)
    #t = F.softmax(t, dim=1)
    return t

epoch = 30

train_losses, test_losses = [], []

for e in range(epoch):
    train_loss = 0
    test_loss = 0
    accuracy = 0

    for images, labels in train_loader:

        optimizer.zero_grad()
        op = model(images) #output 
        loss = criterion(op, labels)
        train_loss += loss.item()
        loss.backward()
        optimizer.step()

    else:
        with torch.no_grad():
            model.eval()
            for images,labels in testloader:
                log_ps = model(images)
                prob = torch.exp(log_ps)
                top_probs, top_classes = prob.topk(1, dim=1)
                equals = labels == top_classes.view(labels.shape)
                accuracy += equals.type(torch.FloatTensor).mean()
                test_loss += criterion(log_ps, labels)
        model.train()
    print("Epoch: {}/{}.. ".format(e+1, epoch),
              "Training Loss: {:.3f}.. ".format(train_loss/len(train_loader)),
              "Test Loss: {:.3f}.. ".format(test_loss/len(testloader)),
              "Test Accuracy: {:.3f}".format(accuracy/len(testloader)))
    train_losses.append(train_loss/len(train_loader))
    test_losses.append(test_loss/len(testloader))

Answer 1

Be careful when using nn.CrossEntropyLoss and nn.NLLLoss without any confusion.

I don't think your code has problem, I tried to run it exactly the same way as you defined. Maybe you didn't give us other lines of code for initialization for other parts, and it might be a problem.

• log_ps is supposed to be log_softmax values but your network only produce logits values (As you said you used CrossEntropyLoss . These lines can be modified as below:

log_ps = model(images)
prob = torch.exp(log_ps)
top_probs, top_classes = prob.topk(1, dim=1)

# Change into simple code:
logits = model(images)
output = logits.argmax(dim=-1) # should give you the class of predicted label

• I just made a very similar version of your code and it works well:

  1. Define your model

import torch
import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):

    def __init__(self):
        super(Model, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5)
        self.conv2 = nn.Conv2d(in_channels=6, out_channels=12, kernel_size=5)
        self.fc1 = nn.Linear(in_features=12 * 4 * 4, out_features=120)
        self.fc2 = nn.Linear(in_features=120, out_features=60)
        self.out = nn.Linear(in_features=60, out_features=10)
        #the output will be 0~9 (10)

    def forward(self, t):
        # implement the forward pass 
        # (1)input layer
        t = t 
        # (2) hidden conv layer
        t = self.conv1(t)
        t = F.relu(t)
        t = F.max_pool2d(t, kernel_size=2, stride=2)

        # (3) hidden conv layer
        t = self.conv2(t)
        t = F.relu(t)
        t = F.max_pool2d(t, kernel_size=2, stride=2)

        # (4) hidden linear layer
        t = t.reshape(-1, 12 * 4 * 4)
        t = self.fc1(t)
        t = F.relu(t)

        # (5) hidden linear layer
        t = self.fc2(t)
        t = F.relu(t)
        # (6) output layer
        t = self.out(t)
        return t

2. Prepare your dataset

import torchvision
import torchvision.transforms as T

train_dataset = torchvision.datasets.FashionMNIST('./data', train=True, 
                                            transform=T.ToTensor(),
                                            download=True)

test_dataset = torchvision.datasets.FashionMNIST('./data', train=False, 
                                            transform=T.ToTensor(),
                                            download=True)

train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=64, shuffle=False)

3. Start training

epoch = 5
model = Model();
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())

train_losses, test_losses = [], []

for e in range(epoch):
    train_loss = 0
    test_loss = 0
    accuracy = 0

    for images, labels in train_loader:

        optimizer.zero_grad()
        logits = model(images) #output 
        loss = criterion(logits, labels)
        train_loss += loss.item()
        loss.backward()
        optimizer.step()

    else:
        with torch.no_grad():
            model.eval()
            for images,labels in test_loader:
                logits = model(images)
                output = logits.argmax(dim=-1)
                equals = (labels == output)
                accuracy += equals.to(torch.float).mean()
                test_loss += criterion(logits, labels)
        model.train()
    print("Epoch: {}/{}.. ".format(e+1, epoch),
              "Training Loss: {:.3f}.. ".format(train_loss/len(train_loader)),
              "Test Loss: {:.3f}.. ".format(test_loss/len(test_loader)),
              "Test Accuracy: {:.3f}".format(accuracy/len(test_loader)))
    train_losses.append(train_loss/len(train_loader))
    test_losses.append(test_loss/len(test_loader))

And here is the result, it converges at least:

Epoch: 1/5..  Training Loss: 0.721..  Test Loss: 0.525..  Test Accuracy: 0.809
Epoch: 2/5..  Training Loss: 0.473..  Test Loss: 0.464..  Test Accuracy: 0.829
Epoch: 3/5..  Training Loss: 0.408..  Test Loss: 0.391..  Test Accuracy: 0.858
Epoch: 4/5..  Training Loss: 0.370..  Test Loss: 0.396..  Test Accuracy: 0.858
Epoch: 5/5..  Training Loss: 0.348..  Test Loss: 0.376..  Test Accuracy: 0.858