Package 'mafR'

Utility to manage torch tensors

Description

(Currently not used nor exported) utility converting an R object to a torch tensor.

Usage

.r_to_torch(x, py_handle, device)

Arguments

x	An R object suitable for use in reticulate::r_to_py(x) (this being as indefinite as the r_to_py documentation in this respect.)
py_handle	The return value of get_py_MAF_handle, or possibly more generally an environment with (at least) elements torch and device defined as in such a return value.
device	Character: "cpu"; or a GPU backend, either "cuda" (or "cuda:0", etc.) or "mps" depending on system capabilities.

Value

r_to_torch returns a 32-bit floating-point torch tensor allocated on the given device.

Examples

my_env <- list2env(list(is_set=FALSE),parent = emptyenv())
my_env <- get_py_MAF_handle(my_env, reset=FALSE, torch_device="cpu")

---

Python controls

Description

Interface to control variables in a Python environment possibly used by Infusion. Currently the only implemented control is that of the torch random seed.

Usage

control_py_env(py_handle, seed = NULL)

Arguments

py_handle	An R environment that provides access to a Python evaluation environment, as produced by get_py_MAF_handle
seed	Numeric: passed (as integer value) to torch.random.manual_seed.

Value

Returns NULL invisibly.

Examples

## Initialization of Python session:
my_env <- list2env(list(is_set=FALSE),parent = emptyenv())
py_handle <- get_py_MAF_handle(my_env, reset=FALSE, torch_device="cpu")

if (inherits(py_handle,"environment"))  control_py_env(py_handle, seed=0L)

---

Utilities to manage Python environment and torch tensors

Description

Utility initializing a Python environment for running zuko.flows.MAF and retrieving it.

Usage

get_py_MAF_handle(envir, reset=FALSE, torch_device="cpu", GPU_mem=NULL,
                    verbose = TRUE)

Arguments

envir	An environment (in the R sense) initialized as shown in the Examples.
reset	Boolean: Whether to reinitialize the Python session or not.
torch_device	Character: "cpu"; or a GPU backend, either "cuda" (or "cuda:0", etc.) or "mps" depending on system capabilities.
GPU_mem	For development purposes (effect is complicated). An amount of (dedicated) GPU memory, in bytes.
verbose	Boolean. Whether to print some messages or not.

Value

If successful, get_py_MAF_handle returns the modified input environment. If sourcing the Python code provided by mafR failed (presumably from trying to use an improperly set-up Python environment), the error condition message is returned.

Examples

# Initialization of Python session:
my_env <- list2env(list(is_set=FALSE),parent = emptyenv())
my_env <- get_py_MAF_handle(my_env, reset=FALSE, torch_device="cpu")
    
if (inherits(my_env,"environment"))  {
    # => provides access to:
    my_env$torch # Imported Python package (result of reticulate::import("torch"))
    my_env$device # the torch_device
    # and to internal definitions for MAF training 
}

---

Interface for masked autoregressive flows

Description

This wraps Python procedures to train Masked Autoregressive Flows (MAFs, Paramakarios et al. 2017) using the Python package zuko. It has been tested with version 1.1.0 and 1.2.0 of that package. Note that objects created by its version 1.2.0 cannot be read with its version 1.1.0 (i.e., when saved in and read from pickle files).

The simplest portable way to get mafR working may be to install it in a conda environment. Below is a complete installation recipe. More information about alternative installation procedure may be found on the Git repository for mafR, https://github.com/f-rousset/mafR.

mkdir -p ~/miniconda3
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.sh
bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
rm ~/miniconda3/miniconda.sh

~/miniconda3/bin/conda init bash
conda create --name maf-conda python==3.10
conda activate maf-conda

pip install zuko

conda install R                                       
conda install conda-forge::r-gmp
conda install conda-forge::gsl

and, in an R session within the maf-conda environment:

install.packages("reticulate") 
library(reticulate)
use_condaenv(condaenv="maf-conda", conda="~/miniconda3/bin/conda")
install.packages("mafR")

# 'mafR' was first designed for use with 'Infusion':
install.packages("Infusion")
install.packages("Rmixmod") # only a Suggested dependency of Infusion, but needed.

References

Papamakarios, G., D. Sterratt, and I. Murray. 2019. Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows. Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, PMLR 89:837-848, 2019. https://doi.org/10.48550/arXiv.1705.07057 ; https://proceedings.mlr.press/v89/papamakarios19a.html

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

---