Recent advances in person re-identification have demonstrated enhanced\ndiscriminability, especially with supervised learning or transfer learning.\nHowever, since the data requirements---including the degree of data\ncurations---are becoming increasingly complex and laborious, there is a\ncritical need for unsupervised methods that are robust to large intra-class\nvariations, such as changes in perspective, illumination, articulated motion,\nresolution, etc. Therefore, we propose an unsupervised framework for person\nre-identification which is trained in an end-to-end manner without any\npre-training. Our proposed framework leverages a new attention mechanism that\ncombines group convolutions to (1) enhance spatial attention at multiple scales\nand (2) reduce the number of trainable parameters by 59.6%. Additionally, our\nframework jointly optimizes the network with agglomerative clustering and\ninstance learning to tackle hard samples. We perform extensive analysis using\nthe Market1501 and DukeMTMC-reID datasets to demonstrate that our method\nconsistently outperforms the state-of-the-art methods (with and without\npre-trained weights).\n