These are configurations for ngme
optimization process.
Usage
control_opt(
seed = Sys.time(),
burnin = 100,
iterations = 500,
estimation = TRUE,
standardize_fixed = TRUE,
n_batch = 10,
iters_per_check = iterations/n_batch,
optimizer = adam(),
start_sd = 0.5,
n_parallel_chain = 4,
max_num_threads = n_parallel_chain,
print_check_info = FALSE,
max_relative_step = 0.5,
max_absolute_step = 0.5,
rao_blackwellization = FALSE,
n_trace_iter = 10,
sampling_strategy = "all",
solver_backend = if (Sys.info()["sysname"] == "Darwin") "accelerate" else "cholmod",
solver_type = "llt",
verbose = FALSE,
store_traj = TRUE,
robust = FALSE,
n_min_batch = 3,
n_slope_check = 3,
trend_std_conv_check = TRUE,
std_lim = 0.01,
trend_lim = 0.01,
R_hat_conv_check = TRUE,
max_R_hat = 1.1,
pflug_conv_check = TRUE,
pflug_alpha = 0.9
)Arguments
- seed
set the seed for pesudo random number generator
- burnin
interations for burn-in periods (before optimization)
- iterations
optimization iterations
- estimation
run the estimation process (call C++ in backend)
- standardize_fixed
whether or not standardize the fixed effect
- n_batch
number of checkpoints; optimization is split into
n_batchequal batches- iters_per_check
run how many iterations between each check point (or specify
n_batch)- optimizer
choose different sgd optimization method, currently support "sgd", "precond_sgd", "momentum", "adagrad", "rmsprop", "adam", "adamW" see ?sgd, ?precond_sgd, ?momentum, ?adagrad, ?rmsprop, ?adam, ?adamW
- start_sd
standard deviation of the initial parameter (1st chain fixed, other chains random), set 0 to be fixed for all chains
- n_parallel_chain
number of parallel chains
- max_num_threads
maximum number of threads used for parallel computing, by default will be set same as n_parallel_chain. If it is more than n_parallel_chain, the rest will be used to parallel different replicates of the model.
- print_check_info
print the convergence information
- max_relative_step
max relative step allowed in 1 iteration
- max_absolute_step
max absolute step allowed in 1 iteration
- rao_blackwellization
use rao_blackwellization
- n_trace_iter
use how many iterations to approximate the trace (Hutchinson’s trick)
- sampling_strategy
subsampling method of replicates of model, c("all", "is") "all" means using all replicates in each iteration, "ws" means weighted sampling (each iteration use 1 replicate to compute the gradient, the sample probability is proption to its number of observations)
- solver_backend
backend in ("eigen", "cholmod", "accelerate", "pardiso")
- solver_type
factorization type: "llt" or "ldlt"
- verbose
print estimation
- store_traj
store the optimizer trajectory for diagnostics (set FALSE to reduce memory)
- n_min_batch
minimum number of checkpoints before any convergence diagnostic is attempted
- n_slope_check
number of checkpoints used as the regression window for the trend test
- trend_std_conv_check
enable the trend/std diagnostic (uses
std_lim,trend_lim,n_slope_check)- std_lim
maximum allowed standard deviation
- trend_lim
maximum allowed slope
- max_R_hat
maximum allowed R_hat
- pflug_conv_check
use Pflug diagnostic for convergence check
- pflug_alpha
scaling factor (0-1] for Pflug criterion: require
pflug_sum < pflug_alpha * max_pflug_sum- max_R_hat_conv_check
use max_R_hat for convergence check
Details
Convergence diagnostics (multi-chain):
* R-hat: per-parameter Gelman–Rubin statistic; passes if R_hat <= max_R_hat.
* Trend/Std: uses the last n_slope_check checkpoints after at least n_min_batch batches.
Passes when both the relative std (sqrt(var)/|mean| <= std_lim) and linear trend
of the means (|slope| <= trend_lim) satisfy their thresholds.
* Pflug: per-chain criterion pflug_sum < pflug_alpha * max_pflug_sum in the latest batch;
if all chains satisfy it, overall convergence is declared.
Checks are evaluated every iters_per_check = iterations / n_batch. A parameter is marked
converged if any enabled parameter-level diagnostic (R-hat or Trend/Std) passes; the run stops
when all parameters converge or when the Pflug diagnostic triggers.