RAABBVI¶
- class viabel.RAABBVI(sgo, *, rho=0.5, iters0=1000, accuracy_threshold=0.1, inefficiency_threshold=1.0, init_rmsprop=False, **kwargs)[source]¶
A robust, automated, and accurate BBVI optimizer (RAABBVI)
This algorithm combines the FASO algorithm with a termination rule to determine the appropriate point where the algorithm could terminate. The termination rule is based on the trade-off between improved accuracy of the variational approximation if the current learning rate is reduced by an adaptation factor \(\rho \in (0,1)\) and the time required to reach that improved accuracy. If the improved accuracy level is large compared to the runtime then this algorithm adaptively decrease the learning rate and if not algorithm will be terminated.
See more details: https://arxiv.org/pdf/2203.15945.pdf
- Parameters:
- sgoStochasticGradientOptimizer object
Optimizer to use for computing descent direction.
- rhofloat, optional
Learning rate reducing factor. The default is 0.5
- iters0int, optional
Small iteration number. The default is 1000 for decreased learning rate.
- accuracy_thresholdfloat, optional
Absolute SKL accuracy threshold
- inefficiency_thresholdfloat, optional
Threshold for the inefficiency index.
- init_rmpspropBoolean, optional
Indicate whether to run using RMSProp optimization method for the initial learning rate
- **kwargs:
Keyword arguments for FASO
Methods
Detecting the relationship trend between learning rate and number of iterations to reach convergence
optimize(K_max, objective, init_param)- Parameters:
weighted_linear_regression(model, y, x[, s, ...])weighted regression with likelihood term having the weight Parameters ---------- model : pystan model Pystan model to conduct the sampling y : numpy_ndarray Response variable x : numpy_ndarray Predictor variable s : int, optional Observation weight parameter. The default is 9. a : int, optional Power of the weight parameter. The default is 1/4. n_chains: int, optional Number of sampling chains. The default is 4.
wls(x, y[, s, a])weighted least squares
- __init__(sgo, *, rho=0.5, iters0=1000, accuracy_threshold=0.1, inefficiency_threshold=1.0, init_rmsprop=False, **kwargs)[source]¶
- convg_iteration_trend_detection(slope)[source]¶
Detecting the relationship trend between learning rate and number of iterations to reach convergence
- Parameters:
- slopefloat
slope of the fitted regression model
- Returns:
- bool
Indicating having negative relationship or not
- optimize(K_max, objective, init_param)[source]¶
- Parameters:
- K_maxint
Number of iterations of the optimization
- objective: `function`
Function for constructing the objective and gradient function
- init_paramnumpy.ndarray, shape(var_param_dim,)
Initial values of the variational parameters
- weighted_linear_regression(model, y, x, s=9, a=0.25, n_chains=4)[source]¶
weighted regression with likelihood term having the weight Parameters ———- model : pystan model
Pystan model to conduct the sampling
- ynumpy_ndarray
Response variable
- xnumpy_ndarray
Predictor variable
- sint, optional
Observation weight parameter. The default is 9.
- aint, optional
Power of the weight parameter. The default is 1/4.
- n_chains: int, optional
Number of sampling chains. The default is 4.
- Returns:
- kappafloat
Power
- cfloat
Constant
- fitPystan object
Pystan fit object if results = True