Learning with few labeled data is a key challenge for visual recognition, as deep neural networks tend to overfit using a few samples only. One of the Few-shot learning methods called metric learning addresses this challenge by first learning a deep distance metric to determine whether a pair of images belong to the same category, then applying the trained metric to instances from other test set with limited labels. This method makes the most of the few samples and limits the overfitting effectively. However, extant metric networks usually employ Linear classifiers or Convolutional neural networks (CNN) that are not precise enough to globally capture the subtle differences between vectors. In this paper, we propose a novel approach named Bi-attention network to compare the instances, which can measure the similarity between embeddings of instances precisely, globally and efficiently. We verify the effectiveness of our model on two benchmarks. Experiments show that our approach achieved improved accuracy and convergence speed over baseline models.