Multi-feature contrastive learning for unpaired image-to-image translation