CIFAR10 classification accuracy is not improved

1

I am having trouble with loss_value not going down when I am learning using CIFAR10 with VGG16 that I built myself. What am I doing wrong?

class VGG16(nn.Module):
    def __init__(self):
        super(VGG16, self).__init__()
        self.features = nn.Sequential(
                        nn.Conv2d(3, 64, kernel_size=(3,3),stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
                        nn.Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
                        nn.Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
                        nn.Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
                        nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU6(inplace=True),
                        nn.Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
                        nn.ReLU(inplace=True),
                        nn.MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False),
                        )
        self.avgpool = nn.AdaptiveAvgPool2d((7, 7))
        self.classifier = nn.Sequential(
                        nn.Linear(512*7*7,4096),
                        nn.ReLU(inplace=True),
                        nn.Dropout(0.5,inplace=False),
                        nn.Linear(4096,4096),
                        nn.ReLU(inplace=True),
                        nn.Dropout(0.5,inplace=False),
                        nn.Linear(4096,10),
                        nn.LogSoftmax(dim = 0),
        )
    def forward(self,x):
        x = self.features(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x

def train(model,train_loader,device):
# 交差エントロピー
    criterion = nn.CrossEntropyLoss()
    # 確率的勾配降下法
    optimizer = optim.AdamW(model.parameters(), lr=0.004)
    loss_values = []
    for epoch in range(100):
        running_loss = 0.0
        for i, data in enumerate(train_loader, 0):
            #print(i)
            inputs, labels = data
            optimizer.zero_grad()
            outputs = model(inputs.to(device))
            loss = criterion(outputs, labels.to(device))
            # 誤差逆伝播
            loss.backward()
            optimizer.step()
            train_loss = loss.item()
            running_loss += loss.item()
            if i % 500 == 499:
                print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 500))
                loss_values.append(running_loss/500)
                running_loss = 0.0
    plt.plot(loss_values)
    plt.show()
    #print('Finished Training')
    return model
def val(model,val_dataloader,device):
    model = model.eval()
    correct = 0
    total = 0
    with torch.no_grad():
        for data in val_dataloader:
            images, labels = data
            images = images.to(device)
            labels = labels.to(device)
            outputs = model(images)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            #print(outputs.data)
            #print((predicted == labels).sum())
            correct += (predicted == labels).sum().item()
    print('Accuracy of the network on images: %d %%' % (100 * correct / total))
if __name__ == '__main__':
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    model = (v.VGG16()).to(device)
    # ToTensor：画像のグレースケール化（RGBの0~255を0~1の範囲に正規化）、Normalize：Z値化（RGBの平均と標準偏差を0.5で決め打ちして正規化）
    transform = transforms.Compose(
    [transforms.ToTensor(),
     transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

# トレーニングデータをダウンロード
    trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform)
    train_loader = torch.utils.data.DataLoader(trainset, batch_size=100, shuffle=True, num_workers=2)
# テストデータをダウンロード
    testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform)
    val_dataloader = torch.utils.data.DataLoader(testset, batch_size=10, shuffle=True, num_workers=2)
    #train_loader = torch.load("D:\Imagetrain.pt")
    #train_dataloader = data.DataLoader(train_loader, batch_size = 100, shuffle = True)
    #val_loader = torch.load("D:\Imagevalid.pt")
    #val_dataloader = data.DataLoader(val_loader, batch_size = 100, shuffle = True)
    print("VGG16")
    model = train(model,train_loader,device)
    
    val(model,val_dataloader,device)

[1, 500] loss: 2.712
[2, 500] loss: 2.303
[3, 500] loss: 2.303
[4, 500] loss: 2.303
[5, 500] loss: 2.303
[6, 500] loss: 2.303
[7, 500] loss: 2.303
[8, 500] loss: 2.303
[9, 500] loss: 2.303
[10, 500] loss: 2.303

2

The last activation as nn.LogSoftmax(dim = 0) looks wrong since you are calculating the log probabilities in the batch dimension instead of the class dimension.
Also, you can remove this layer completely as nn.CrossEntropyLoss expects raw logits.

3

Should I just set nn.LogSoftmax(dim = 0) to nn.Softmax? Or do I just delete nn.LogSoftmax(dim = 0)?

4

No, the usage of softmax would be wrong so remove it completely and pass the outputs of the linear layer to the criterion directly.

5

I removed nn.LogSoftmax(dim = 0) from the model but the loss_value does not drop. What am I doing wrong?
[1, 500] loss: 5.94170756626129126232
[2, 500] loss: 2.30296494007110608138
[3, 500] loss: 2.30301657056808473101
[4, 500] loss: 2.30306150436401368964
[5, 500] loss: 2.30292595243453979137
[6, 500] loss: 2.30289011669158938389
[7, 500] loss: 2.30299579811096188919
[8, 500] loss: 2.30294379997253439996
[9, 500] loss: 2.30301538562774643637
self.classifier = nn.Sequential(
nn.Linear(512, 512), #512 * 7 * 7, 4096),
nn.ReLU(True),
nn.Dropout(),
nn.Linear(512, 32 ), #4096, 4096),
nn.ReLU(True),
nn.Dropout(),
nn.Linear(32, num_classes), #4096
)
Would self.classifier be more accurate if I did it this way?

6

ohhhh,i found my error that my pca1.weight shape is [8, 10,10]
I want to use eight convolutions to check 100 matrices for dimensionality reduction, and finally turn them into eight. Is it possible?

7

Where and how should I make changes?