Utility method: register a tensor as a buffer or trainable parameter

Compute the Krylov matrix (b, Ab, A^2b, ...) using the squaring trick. If return_power=True, return A^{L-1} as well

Prints content of DictConfig using Rich library and its tree structure. Args: config (DictConfig): Configuration composed by Hydra.

Convert a torch float32 tensor to a numpy array, sharing the same memory.

logsumexp for complex

Arguments: p: (..., d) Return: A: (..., d, d)

Different loss functions depending on the dataset / task

y.shape is B x L x D

Same as triangular_toeplitz_multiply but inputs and output assume to be 0-padded already.

Transform data: f(x) = ax - b

Invert the causal convolution/polynomial/triangular Toeplitz matrix represented by u. This is easiest in the polynomial view: https://www.cs

Solve u * x = v. Initial attempt by inverting the multiplication algorithm, which I think doesn't work.

Compute the Krylov matrix (c^T b, c^T A b, c^T A^2 b, ...), where A = shift + p e_d^T. Arguments: p: (..., d), real b: (...,

Explicit construction of Krylov matrix [v A @ v A^2 @ v ... A^{n-1} @ v] where A = Z_f. This uses vectorized indexing and cumprod so it's much

Initialize weights for differencing kernel - Assume u is shape B x D x L

y = {'true': torch.stack(total_y_true) 'pred': torch.stack(total_y_pred), 'true_informer': total_y_true_informer 'pred