MAF.options: Control of MAF design and training

In Infusion: Inference Using Simulation

Control of MAF design and training

Description

Masked Autoregressive Flows can be used by Infusion to infer various densities. This functionality requires the mafR package, and is requested through the using argument of infer.SLik.joint or refine (see Details).

config_mafR is a convenient way to reset the Python session (erasing all results stored in its memory), and in particular to enforce GPU usage.

MAF.options is a wrapper for Infusion.options which facilitates the specification of options that control the design of Masked Autoregressive Flows and their training.

Usage

config_mafR(torch_device, ...)
MAF.options(template = "zuko-like", ...)

Arguments

torch_device	character: "cpu", "cuda" (or "cuda:0", etc., useful if several GPUs are available) or "mps".
template	Defines a template list of options, subsequently modified by options specified through the ..., if any. Possible values of template are "zuko-like", "PSM19", NULL, or a list of options. The default value "zuko-like" reproduces Infusion's defaults.
...	For MAF.options: a named value, or several of them, that override or complete the template values. For config_mafR: for development purposes; not documented.

Details

Possible using values:

With using="c.mafR" four different MAFs are computed in every iteration: the density of statistics given parameters (providing the likelihood), the joint density, the instrumental density of parameters, and the “instrumental posterior” density of parameters given the data.

With using="MAFmix", MAFs are computed only for the joint density and the instrumental density. The likelihood is deduced from them and a multivariate Gaussian mixture model is used to infer the “instrumental posterior” density.

using="mafR" can be used to let Infusion select one of the above options. "MAFmix" is currently called as it is faster, but this is liable to change in the future, so do not use this if you want a repeatable workflow.

Possible template values for MAF.options:

"PSM19" defines a template list of control values recommended by Papamakarios et al. (2019, Section 5.1). "zuko-like" defines a template list of values inspired from the tutorials of the zuko Python package, with some extra twists. Concretely, their definition is

    if (template == "zuko-like") {
        optns <- list(design_hidden_layers = .design_hidden_layers_MGM_like, 
            MAF_patience = 30, MAF_auto_layers = 3L, Adam_learning_rate = 0.001)
    }
    else if (template == "PSM19") {
        optns <- list(design_hidden_layers = .design_hidden_layers_PSM19, 
            MAF_patience = 20, MAF_auto_layers = 5L, Adam_learning_rate = 1e-04)
    }

and any replacement value should match the types (function, numeric, integer...) of the shown values, and the formals of the internal functions, in order to avoid cryptic errors.

The internal .design_hidden_layers... functions return a vector of numbers H_i of hidden values per layer i of the neural network. THevector has an attribute giving the resulting approximate number of parameters P of the deep-learning model according to Supplementary Table 1 of Papamakarios et al. 2017. H_i=50L for "PSM19". For "zuko-like", a typically higher value will be used. It is defined as a power of two such that P is of the order of 8 to 16 times the default number of parameters of the multivariate gaussian mixture model that could be used instead of MAFs.

Other controls which can be modified through the ... are
* MAF_validasize, a function which returns the size of the validation set, whose default definition returns 5% of its input value nr which is the number of samples in the reference table (consistently with Papamakarios et al., 2019);
* MAF_batchsize, a function that returns the batch size for the Adam optimizer. Its default simply returns 100L, but non-default functions can be defined, with at least the ... as formal arguments (more elaborate formals are possible but not part of the API).

Value

config_mafR is used for its side effects. Returns NULL invisibly.

MAF.options returns the result of calling Infusion.options on the arguments defined by the template and the ... . Hence, it is a list of previous values of the affected options.

References

Papamakarios, G., T. Pavlakou, and I. Murray (2017) Masked Autoregressive Flow for Density Estimation. Pp. 2335–2344 in I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, eds. Advances in Neural Information Processing Systems 30. Curran Associates, Inc.
http://papers.nips.cc/paper/6828-masked-autoregressive-flow-for-density-estimation

Rozet, F., Divo, F., Schnake, S (2023) Zuko: Normalizing flows in PyTorch. https://doi.org/10.5281/zenodo.7625672

See Also

save_MAFs for saving and loading MAF objects.

Examples

  MAF.options(template = "zuko-like", 
              Adam_learning_rate=1e-4,
              MAF_batchsize = function(...) 100L)

## Not run: 
## These examples require the mafR package, 
## and a Python installation with cuda capabilities.

if (requireNamespace("mafR", quietly=TRUE)) {

config_mafR() # set default: "cpu", i.e. GPU not used
config_mafR("cuda") # sets cuda as GPU backend 
config_mafR() # reset python session, keeping latest backend


  config_mafR(torch_device="cuda")
  
  # function for sampling from N(.,sd=1)
  toyrnorm <- function(mu) {
    sam <- rnorm(n=40,mean=mu,sd=1)
    return(c(mean1=mean(sam)))
  } 

  # simulated data, standing for the actual data to be analyzed:  
  set.seed(123)
  Sobs <- toyrnorm(mu=4) 

  parsp <- init_reftable(lower=c(mu=2.8), 
                         upper=c(mu=5.2))
  simuls <- add_reftable(Simulate="toyrnorm", parsTable=parsp)

  MAF.options(template = "zuko-like")
  densv <- infer_SLik_joint(simuls,stat.obs=Sobs, using="mafR")

  # Usual workflow using inferred surface:
  slik_j <- MSL(densv, eval_RMSEs = FALSE) ## find the maximum of the log-likelihood surface
  # ETC.
  
  save_MAFs(slik_j, prefix = "toy_") # See its distinct documentation.
}

## End(Not run)

Infusion documentation built on Sept. 30, 2024, 9:16 a.m.