Vgg16 model validation accuracy is stuck

Hi, I am working on a CNN model for MRI brain images classification (Alzheimer disease), I use transfer learning method for image classification - vgg16 model trained on ImageNet (1000 classes). I’ve modified last layer so the size of each output sample is set to 2.

I am using SGD optimizer, weighted CrossEntropyLoss as loss function because dataset is not balanced and learning rate scheduler - StepLR.

I encountered a problem with validation accuracy that stays (almost) the same or is decreasing (approximately 70% whereas state of the art in this topic is around 95-99%), I thought that the problem is overfitting but then I’ve tried to use L2 regularization, data augmentation and dropout but nothing helped. MRI images are normalized.

Here is a code of my training function

def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        for phase in ['train', 'val']:
            if phase == 'train':

            running_loss = 0.0
            running_corrects = 0

            for inputs, masks, labels in dataloaders[phase]:
                inputs =
                labels =


                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs.float())
                    _, preds = torch.max(outputs, 1)

                    loss = criterion(outputs, labels)

                    if phase == 'train':

                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds ==
            if phase == 'train':

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]

            print('{} Loss: {:.4f} Acc: {:.4f}'.format(
                phase, epoch_loss, epoch_acc))

            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())


    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(
        time_elapsed // 60, time_elapsed % 60))
    print('Best val Acc: {:4f}'.format(best_acc))

    return model

I am training the model for 100 epochs.
Learning rate is 0.001

I would appreciate any help. Thanks in advance