PyTorch has torch.inverse, but given the nature of noisy data-driven problems it would be useful to have a torch.pinverse option which is calculated using the SVD.

As a side note -- it would also be great if these two functions (.inverse and the proposed .pinverse) were more optimized for a batch operation. As in, passing a tensor of shape (32, 10, 10) would perform 32 instances of inverse on 10x10 matricies, or passing in a tensor of shape (32, 8, 4, 4) would perform 32 * 8 inverses.