deepregression_init: Initializing Deep Distributional Regression Models
In davidruegamer/deepregression: Fitting Deep Distributional Regression

deepregression_init

R Documentation

Initializing Deep Distributional Regression Models

Description

Initializing Deep Distributional Regression Models

Usage

deepregression_init(
  n_obs,
  ncol_structured,
  ncol_deep,
  list_structured,
  list_deep,
  use_bias_in_structured = FALSE,
  nr_params = 2,
  train_together = NULL,
  lambda_lasso = NULL,
  lambda_ridge = NULL,
  family,
  dist_fun = NULL,
  variational = TRUE,
  weights = NULL,
  learning_rate = 0.01,
  optimizer = optimizer_adam(lr = learning_rate),
  monitor_metric = list(),
  posterior = posterior_mean_field,
  prior = prior_trainable,
  orthog_fun = orthog,
  orthogX = NULL,
  kl_weight = 1/n_obs,
  output_dim = 1,
  mixture_dist = FALSE,
  split_fun = split_model,
  ind_fun = function(x) x,
  extend_output_dim = 0,
  offset = NULL,
  additional_penalty = NULL,
  constraint_fun = NULL,
  compile_model = TRUE
)

Arguments

`n_obs`	number of observations
`ncol_structured`	a vector of length #parameters defining the number of variables used for each of the parameters. If any of the parameters is not modelled using a structured part the corresponding entry must be zero.
`ncol_deep`	a vector of length #parameters defining the number of variables used for each of the parameters. If any of the parameters is not modelled using a deep part the corresponding entry must be zero. If all parameters are estimated by the same deep model, the first entry must be non-zero while the others must be zero.
`list_structured`	list of (non-linear) structured layers (list length between 0 and number of parameters)
`list_deep`	list of deep models to be used (list length between 0 and number of parameters)
`use_bias_in_structured`	logical, whether or not to use a bias in structured layer
`nr_params`	number of distribution parameters
`train_together`	see `?deepregression`
`lambda_lasso`	penalty parameter for l1 penalty of structured layers
`lambda_ridge`	penalty parameter for l2 penalty of structured layers
`family`	family specifying the distribution that is modelled
`dist_fun`	a custom distribution applied to the last layer, see `make_tfd_dist` for more details on how to construct a custom distribution function.
`variational`	logical value specifying whether or not to use variational inference. If `TRUE`, details must be passed to the via the ellipsis to the initialization function
`weights`	observation weights used in the likelihood
`learning_rate`	learning rate for optimizer
`optimizer`	optimizer used (defaults to adam)
`monitor_metric`	see `?deepregression`
`posterior`	function defining the posterior
`prior`	function defining the prior
`orthog_fun`	function defining the orthogonalization
`orthogX`	vector of columns defining the orthgonalization layer
`kl_weight`	KL weights for variational networks
`output_dim`	dimension of the output (> 1 for multivariate outcomes)
`mixture_dist`	see `?deepregression`
`split_fun`	see `?deepregression`
`ind_fun`	see `?deepregression`
`extend_output_dim`	see `?deepregression`
`offset`	list of logicals corresponding to the paramters; defines per parameter if an offset should be added to the predictor
`additional_penalty`	to specify any additional penalty, provide a function that takes the `model$trainable_weights` as input and applies the additional penalty. In order to get the correct index for the trainable weights, you can run the model once and check its structure.
`constraint_fun`	function; a constraint for the linear layers
`compile_model`	logical; whether to compile the model (default is TRUE)
`fsbatch_optimizer`	logical; see `?deepregression`