Learning shape correspondence with anisotropic convolutional neural networks