fit_sgd: Title

In jtipton25/sgMRA: Multi-Resolution Gaussian Process Approximations with Stochastic Gradient Descent

Title

Description

Title

Usage

fit_sgd(
  y,
  locs,
  grid,
  y_obs = NULL,
  z = NULL,
  alpha = NULL,
  alpha_x1 = NULL,
  alpha_y1 = NULL,
  alpha_x2 = NULL,
  alpha_y2 = NULL,
  learn_rate = 0.001,
  rate_schedule = NULL,
  n_iter = 500,
  n_message = 50,
  penalized = FALSE,
  plot_during_fit = FALSE,
  use_spam = FALSE,
  adam_pars = NULL,
  use_adamw = TRUE,
  sparse_outer = TRUE,
  noisy = TRUE,
  ncores = 1L,
  nchunks = NULL
)

Arguments

y	The data
locs	An N x 2 matrix of spatial locations
grid	A grid object that is the output of make_grid
alpha	If specified, the top layer MRA parameters
alpha_x1	If specified, the first hidden layer MRA parameters for the x-axis variable
alpha_y1	If specified, the first hidden layer MRA parameters for the y-axis variable
alpha_x2	If specified, the second hidden layer MRA parameters for the x-axis variable
alpha_y2	If specified, the second hidden layer MRA parameters for the y-axis variable
learn_rate	The gradient descent learning rate
rate_schedule	If specified, the gradient descent learning rate schedule in decreasing values.
n_iter	The number of gradient descent iterations
n_message	The number of iterations between which to output a message
penalized	Fit using a penalty term
plot_during_fit	Plot the current parameter states every n_message iterations
use_spam	Whether to use the spam (use_spam = TRUE) or Matrix (use_spam = FALSE) package for sparse matrices
adam_pars	The adam parameter state to allow restarting the model
use_adamw	Use the adamW optimizer
sparse_outer	If TRUE, calculate the outer product in a sparse format. For all but the smallest models, this should be TRUE. I should make this automatic going forward
noisy	If TRUE, add random noise to the gradient.
ncores	The number of cores to use for parallelization
nchunks	The number of chunks to divide the distance calculation into. The default argument of NULL will use the same number of chunks as the number of cores.

jtipton25/sgMRA documentation built on Feb. 9, 2023, 4:53 a.m.