dynparam

Provides tools for describing parameters of algorithms in an abstract way. Description can include an id, a description, a domain (range or list of values), and a default value. ‘dynparam’ can also convert parameter sets to a ParamHelpers format, in order to be able to use dynparam in conjunction with mlrMBO. Check ?dynparam for an overview of all functionality provided by dynparam.

Examples

The main goal of dynparam is to be able to describe a set of parameters, be able to serialise the parameter sets, and also sample random settings from the parameter set.

library(tidyverse)
library(dynparam)
set.seed(1)

Define a parameter set as follows:

parameters <- parameter_set(
  integer_parameter(
    id = "num_iter",
    default = 100L,
    distribution = expuniform_distribution(lower = 1L, upper = 10000L),
    description = "Number of iterations"
  ),
  subset_parameter(
    id = "dimreds",
    default = c("pca", "mds"),
    values = c("pca", "mds", "tsne", "umap", "ica"),
    description = "Which dimensionality reduction methods to apply (can be multiple)"
  ),
  integer_range_parameter(
    id = "ks",
    default = c(3L, 15L),
    lower_distribution = uniform_distribution(1L, 5L),
    upper_distribution = uniform_distribution(10L, 20L),
    description = "The numbers of clusters to be evaluated"
  )
)

You can retrieve the default parameters as follows:

get_defaults(parameters)
## $num_iter
## [1] 100
## 
## $dimreds
## [1] "pca" "mds"
## 
## $ks
## [1]  3 15

Serialise a parameter set from/to json/yaml with the as.list() and as_parameter_set() functions.

li <- as.list(parameters)
pa <- as_parameter_set(li)

Sample a parameter set using with sip():

sip(pa, n = 3)
## Loading required namespace: ParamHelpers

## Loading required namespace: lhs

## # A tibble: 3 x 4
##   num_iter dimreds   ks        .object_class
##      <int> <list>    <list>    <list>       
## 1        2 <chr [2]> <dbl [2]> <chr [1]>    
## 2     2003 <chr [2]> <dbl [2]> <chr [1]>    
## 3        5 <chr [3]> <dbl [2]> <chr [1]>

Large parameter set

parameters <- parameter_set(
  integer_parameter(
    id = "num_iter", 
    default = 100L,
    distribution = expuniform_distribution(lower = 1L, upper = 10000L),
    description = "Number of iterations"
  ),
  numeric_parameter(
  id = "delta", 
  default = c(4.5, 2.4, 1.9), 
  distribution = normal_distribution(mean = 5, sd = 1),
  description = "Multiplying factors"
),
  character_parameter(
  id = "method", 
  default = "kendall",
  values = c("kendall", "spearman", "pearson"), 
  description = "Correlation method"
),
  logical_parameter(
  id = "inverse",
  default = TRUE, 
  description = "Inversion parameter"
),
  subset_parameter(
 id = "dimreds",
 default = c("pca", "mds"),
 values = c("pca", "mds", "tsne", "umap", "ica"),
 description = "Which dimensionality reduction methods to apply (can be multiple)"
),
  integer_range_parameter(
  id = "ks",
  default = c(3L, 15L),
  lower_distribution = uniform_distribution(1L, 5L),
  upper_distribution = uniform_distribution(10L, 20L),
  description = "The numbers of clusters to be evaluated."
),
  numeric_range_parameter(
  id = "quantiles",
  default = c(0.15, 0.90),
  lower_distribution = uniform_distribution(0, .4),
  upper_distribution = uniform_distribution(.6, 1),
  description = "Quantile cutoff range"
),
  forbidden = "inverse == (method == 'kendall')"
)

As yaml:

cat(yaml::as.yaml(as.list(parameters)))
- id: num_iter
  default: 100
  description: Number of iterations
  tuneable: yes
  distribution:
    lower: 1
    upper: 10000
    type: expuniform
  type: integer
- id: delta
  default:
  - 4.5
  - 2.4
  - 1.9
  description: Multiplying factors
  tuneable: yes
  distribution:
    lower: -.inf
    upper: .inf
    mean: 5.0
    sd: 1.0
    type: normal
  type: numeric
- id: method
  default: kendall
  description: Correlation method
  tuneable: yes
  values:
  - kendall
  - spearman
  - pearson
  type: character
- id: inverse
  default: yes
  description: Inversion parameter
  tuneable: yes
  type: logical
- id: dimreds
  default:
  - pca
  - mds
  description: Which dimensionality reduction methods to apply (can be multiple)
  tuneable: yes
  values:
  - pca
  - mds
  - tsne
  - umap
  - ica
  type: subset
- id: ks
  default:
  - 3
  - 15
  description: The numbers of clusters to be evaluated.
  tuneable: yes
  lower_distribution:
    lower: 1
    upper: 5
    type: uniform
  upper_distribution:
    lower: 10
    upper: 20
    type: uniform
  type: integer_range
- id: quantiles
  default:
  - 0.15
  - 0.9
  description: Quantile cutoff range
  tuneable: yes
  lower_distribution:
    lower: 0.0
    upper: 0.4
    type: uniform
  upper_distribution:
    lower: 0.6
    upper: 1.0
    type: uniform
  type: numeric_range
- forbidden: inverse == (method == 'kendall')

Generate a random parameter set:

sip(parameters, n = 2)
## # A tibble: 2 x 8
##   num_iter delta     method  inverse dimreds   ks        quantiles .object_class
##      <int> <list>    <chr>   <lgl>   <list>    <list>    <list>    <list>       
## 1     1823 <dbl [3]> kendall FALSE   <chr [3]> <dbl [2]> <dbl [2]> <chr [1]>    
## 2      125 <dbl [3]> pearson TRUE    <chr [2]> <dbl [2]> <dbl [2]> <chr [1]>

Convert paramhelper object:

as_paramhelper(parameters)
##                    Type len             Def                   Constr Req
## num_iter        numeric   -             0.5                   0 to 1   -
## delta     numericvector   3 0.309,0.0046...                   0 to 1   -
## method         discrete   -         kendall kendall,spearman,pearson   -
## inverse         logical   -            TRUE                        -   -
## dimreds   integervector   5       1,1,0,0,0                   0 to 1   -
## ks        numericvector   2         0.5,0.5                   0 to 1   -
## quantiles numericvector   2      0.375,0.75                   0 to 1   -
##           Tunable Trafo
## num_iter     TRUE     Y
## delta        TRUE     Y
## method       TRUE     -
## inverse      TRUE     -
## dimreds      TRUE     Y
## ks           TRUE     Y
## quantiles    TRUE     Y
## Forbidden region specified.

Latest changes

Check out news(package = "dynparam") or NEWS.md for a full list of changes.

Recent changes in dynparam 1.0.2 (04-01-2021)

Recent changes in dynparam 1.0.1 (24-06-2020)

Dynverse dependencies