Calculate accuracy of pretrained AlexNet with ImageFolder

vision
1

I have taken five classes of the ImageNet data set, feeded them into a pretrained AlexNet and want to calculate the accuracy of correctly classified images. With my code below, the loss is decreasing, but the accuracy is always between 20 and 30% (after 50 epochs). This behaviour confuses me extremely.
My five classes are in subfolders of root, as torchvision,datasets.ImageFolder wants it. Altogether are 6500 images.

batch_size = 32
num_epochs = 100

train_loader = data.DataLoader(dataset=train_dataset, batch_size=batch_size)

val_loader = data.DataLoader(dataset=val_dataset, batch_size=batch_size)

        for images, labels in val_loader:
            images = Variable(images)

            # Forward pass only to get logits/output
            outputs = model(images)

            # Get predictions from the maximum value
            _, predicted = torch.max(outputs.data, 1)

            # Total number of labels
            total += labels.size(0)

            correct += (predicted == labels).sum()

        accuracy = 100 * correct / total

        # Print Loss
        print('Iteration: {}. Loss: {}. Accuracy: {}'.format(iter, loss.data[0], accuracy))

Thanks in advance!

2

Can you please include code for your train_dataset and val_dataset as well.