pmxNODE (version 0.1.0)

NN: Neural Network ODE language in nlmixr2 langauage

Description

Neural Network ODE language in nlmixr2 langauage

Usage

NN(
  number = 1,
  state = "t",
  min_init = 0.5,
  max_init = 10,
  n_hidden = 5,
  act = c("ReLU", "Softplus"),
  time_nn = FALSE,
  beta = 20,
  pop = getOption("pmxNODE.pop", TRUE),
  eta_model = c("prop", "add"),
  theta_scale = 0.1,
  eta_scale = 0.1,
  pre_fixef = getOption("pmxNODE.pre_fixef", NULL),
  iniDf = NULL
)

Value

A list with the before and replace elements and iniDf to allow integration in the rxode2/nlmixr2 language directly.

Arguments

number

The neural network number

state

The state to be used in the neural network. For time, use t

min_init

The minimum value of activation point for the neural network, (i.e., minimal expected state value)

max_init

The maximum value of activation point for the NN (i.e. maximum expected state value)

n_hidden

The number of hidden layers in the neural network (default 5)

act

activation in the hidden layer, ReLU and Softplus implemented. Default is ReLU.

time_nn

defines if the neural network is time dependent and consequently all weights from inputs to hidden layer should be strictly negative (default is FALSE)

beta

(numeric) Beta value for the Softplus activation function, only applicable if act="Softplus"; Default to 20.

pop

(boolean) If the generated nlmixr model function should be a fit without (TRUE) or with (FALSE). Default is FALSE.

eta_model

(string)

• “prop” is of form W = lW * EXP(etaW)

• “add” is of form W = lW + etaW

theta_scale

(numeric) Scale in which typical NN parameter values are initialized, default is 0.1, i.e., weights are initialized between -0.3 and 0.3

eta_scale

(numeric) Initial standard deviation of random effects on NN parameters, default is 0.1

pre_fixef

(named vector) Specific initial values for typical parameters, can be obtained from a run nlmixr model (e.g., run_1) through run_1$fixef

iniDf

iniDf

Author

Matthew L Fidler (uses the same functions `nn_generator_nlmixr`, written by Dominic Bräm)

Examples

if (FALSE) {
 if (requireNamespace("rxode2", quietly = TRUE)) {

# Called directly, this isn't that interesting, but can show what
# is produced for rxode2 integration

library(rxode2)

NN(1, state="t", min_init=0.1, max_init=24, pop=TRUE)

# This can be used in the rxode2 language as follows:

f_ode_pop <- function(){
  ini({
    lV <- 1
    prop.err <- 0.1
  })
  model({
    V <- lV
    d/dt(centr)  =  NN(1, state=centr,min_init=0,max_init=300)
    cp = centr / V
    cp ~ prop(prop.err)
  })
}

# but it expands to the complete model:

f_ode_pop()

# This is because pmxNODE uses the extensible user model interface
# in rxode2.  This only works if you load rxode2/nlmixr2 and pmxNODE

}
}