Cross-Domain Image Captioning with Discriminative Finetuning