Fast Neural Style Transfer

vision
1

I’m facing this error while training my model :
AttributeError: ‘list’ object has no attribute ‘relu2_2’

Code :
for epoch in range(1, epochs+1):
transformer.train()
print(‘No of Epochs:’, epoch)
for batch_i, (images, _) in enumerate(train_loader):
optimizer.zero_grad()
images_original = images.to(device)
generated = transformer(images_original)

        features_original = vgg(images_original)
        features_transformed = vgg(generated)

        content_loss = content_weight*l2_loss(features_transformed.relu2_2, features_original.relu2_2)

        style_loss = 0
        for ft_y, gm_s in zip(features_transformed, gram_style):
            gm_y = gram_matrix(ft_y)
            style_loss += l2_loss(gm_y, gm_s[: images.size(0), :, :])
        style_loss *= style_weight

        total_loss = content_loss + style_loss
        total_loss.backward(retain_graph=True)
        optimizer.step()

        train_metrics['content'] += [content_loss.item()]
        train_metrics['style'] += [style_loss.item()]
        train_metrics['total'] += [total_loss.item()]

        print('Train Metrics Total:',train_metrics['total'])

Can anyone figure out why this error is coming as when I checked with other github repos on fastnst they worked perfectly with somewhat similar code. Thanks in advance for help!

2

Based on the error message it seems:

features_transformed = vgg(generated)

returns a list which then raises the error in:

l2_loss(features_transformed.relu2_2, ...)

so check the forward method of your vgg definition and make sure an object is returned which contains the relu2_2 attribute.

1 Like
3

Thank you for the help !!