Initialize bias the same way as torch.nn.Linear.

Executes training. Args: run_or_experiment (function|class|str|Experiment): If function|class|str, this is the algorithm or

assigns "stream" to some global variable e.g. sys.stdout

twiddle_factor: (n // 2, 2, 2) stride: int Return: (n, n)

Product of block 2x2 diagonal matrices, with permutation, implemented in Numpy + Cython. Parameters: ABCDs: list of the ABCDs factors as u

In the constructor we instantiate two nn.Linear modules and assign them as member variables. This Feature extractor class ta

Parameters: size: size of diagonal matrix deg: degree of the polynomials diag1: initialization for the di

Constructor for the LabelSmoothing module. :param smoothing: label smoothing factor

Sets the learning rate to the initial LR decayed by 10 every few epochs

LR schedule that should yield 76% converged accuracy with batch size 256

Losses based on hard reconstruction. Only for evaluation. Doubly stochastic matrices are rounded with the matching function.

Has to be tuple and not pair, Pytorch doesn't like pair

twiddle: (log_n, n / 2, 2, 2) x: (n) Return: (n)

Combine product of two butterfly matrices into one Butterfly.

Naive multiplication P^T v where P is the matrix of coefficients of Chebyshev polynomials. Parameters: v: (batch_size, n) Return:

Multiply two complex matrices. Parameters: X: (..., n, m, 2) Y: (..., m, p, 2) Return: Z: (..., n, p, 2)

7x7 convolution with padding

Convert object of type OmegaConf to Munch so Wandb can log properly Support nested dictionary.