Person re-identification (ReID) methods based on metric learning can adaptively learn metric matrixs with small sets to further improve the retrieval accuracy of existing models. However, inputs of metric learning are often single-scale discriminative features from deep networks via forward propagation in existing metric learning, which cannot adapt to varies poses and scales. To alleviate above issue, we proposes a metric learning method based on deep aggregate feature representation. Especially, we designs a hierarchical feature extraction module (HFE) that employs multiple large kernel convolutions to enhance the discriminative ability of features. Furthermore, an adaptive feature aggregation module (AFA) is proposed, which can utilize the complementary information of features at different layers more effectively and improve the robustness of feature representation. Experiments on four public datasets demonstrate that the proposed method can significantly improve the generalization performance of existing models.