Trained generative moment matching networks (GMMNs); see also
the demo GMMN_QMC or the vignette GMMN_QMC.
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_C_tau_0.25")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_C_tau_0.5")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_C_tau_0.75")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_G_tau_0.25")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_G_tau_0.5")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_G_tau_0.75")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_eqmix_C_tau_0.5_rot90_t4_tau_0.5")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_eqmix_G_tau_0.5_rot90_t4_tau_0.5")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_eqmix_MO_0.75_0.6_rot90_t4_tau_0.5")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_MO_0.75_0.6")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_t4_tau_0.25")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_t4_tau_0.5")
data("GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_t4_tau_0.75")
data("GMMN_dim_3_300_3_ntrn_60000_nbat_5000_nepo_300_NC21_tau_0.25_0.5")
data("GMMN_dim_3_300_3_ntrn_60000_nbat_5000_nepo_300_NG21_tau_0.25_0.5")
data("GMMN_dim_5_300_5_ntrn_60000_nbat_5000_nepo_300_C_tau_0.5")
data("GMMN_dim_5_300_5_ntrn_60000_nbat_5000_nepo_300_G_tau_0.5")
data("GMMN_dim_5_300_5_ntrn_60000_nbat_5000_nepo_300_NC23_tau_0.25_0.5_0.75")
data("GMMN_dim_5_300_5_ntrn_60000_nbat_5000_nepo_300_NG23_tau_0.25_0.5_0.75")
data("GMMN_dim_5_300_5_ntrn_60000_nbat_5000_nepo_300_t4_tau_0.5")
data("GMMN_dim_10_300_10_ntrn_60000_nbat_5000_nepo_300_C_tau_0.5")
data("GMMN_dim_10_300_10_ntrn_60000_nbat_5000_nepo_300_G_tau_0.5")
data("GMMN_dim_10_300_10_ntrn_60000_nbat_5000_nepo_300_NC55_tau_0.25_0.5_0.75")
data("GMMN_dim_10_300_10_ntrn_60000_nbat_5000_nepo_300_NG55_tau_0.25_0.5_0.75")
data("GMMN_dim_10_300_10_ntrn_60000_nbat_5000_nepo_300_t4_tau_0.5")raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate Clayton
copula (with parameter chosen such that Kendall's tau equals 0.25).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate Clayton
copula (with parameter chosen such that Kendall's tau equals 0.5).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate Clayton
copula (with parameter chosen such that Kendall's tau equals 0.75).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate Gumbel
copula (with parameter chosen such that Kendall's tau equals 0.25).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate Gumbel
copula (with parameter chosen such that Kendall's tau equals 0.5).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate Gumbel
copula (with parameter chosen such that Kendall's tau equals 0.75).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate half-half mixture of a Clayton
copula (with parameter chosen such that Kendall's tau equals 0.5)
and a rotated (by 90 degree) $t$ copula (with 4 degrees of freedom
and correlation parameter chosen such that Kendall's tau equals
0.5); see vignette("GMMN_QRNG", package = "gnn").
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate half-half mixture of a Gumbel
copula (with parameter chosen such that Kendall's tau equals 0.5)
and a rotated (by 90 degree) $t$ copula (with 4 degrees of freedom
and correlation parameter chosen such that Kendall's tau equals
0.5).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a bivariate half-half mixture of a Marshall--Olkin
copula (with \(\alpha_{1}=0.75\) and
\(\alpha_{2}=0.60\))
and a rotated (by 90 degree) $t$ copula (with 4 degrees of freedom
and correlation parameter chosen such that Kendall's tau equals
0.5).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a Marshall--Olkin
copula (with \(\alpha_{1}=0.75\) and
\(\alpha_{2}=0.60\)).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a two-dimensional $t$ copula (with 4 degrees of freedom
and equi-correlation parameter chosen such that Kendall's tau equals
0.25).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a two-dimensional $t$ copula (with 4 degrees of freedom
and equi-correlation parameter chosen such that Kendall's tau equals
0.5).
raw R object representing a GMMN (input and output
layer are two-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a two-dimensional $t$ copula (with 4 degrees of freedom
and equi-correlation parameter chosen such that Kendall's tau equals
0.75).
raw R object representing a GMMN (input and output
layer are three-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a three-dimensional nested Clayton copula
(with sector dimensions 2 and 1, corresponding Kendall's tau 0.5
within the first sector and Kendall's tau 0.25 between the two sectors).
raw R object representing a GMMN (input and output
layer are three-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a three-dimensional nested Gumbel copula
(with sector dimensions 2 and 1, corresponding Kendall's tau 0.5
within the first sector and Kendall's tau 0.25 between the two sectors).
raw R object representing a GMMN (input and output
layer are five-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a five-dimensional Clayton
copula (with parameter chosen such that Kendall's tau equals 0.5).
raw R object representing a GMMN (input and output
layer are five-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a five-dimensional Gumbel
copula (with parameter chosen such that Kendall's tau equals 0.5).
raw R object representing a GMMN (input and output
layer are five-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a five-dimensional nested Clayton copula
(with sector dimensions 2 and 3, corresponding Kendall's tau 0.5
and 0.75, and Kendall's tau 0.25 between the two sectors).
raw R object representing a GMMN (input and output
layer are five-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a five-dimensional nested Gumbel copula
(with sector dimensions 2 and 3, corresponding Kendall's tau 0.5
and 0.75, and Kendall's tau 0.25 between the two sectors);
see vignette("GMMN_QRNG", package = "gnn").
raw R object representing a GMMN (input and output
layer are five-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a five-dimensional $t$ copula (with 4 degrees of freedom
and equi-correlation parameter chosen such that Kendall's tau equals
0.5); see vignette("GMMN_QRNG", package = "gnn").
raw R object representing a GMMN (input and output
layer are 10-dimensional, the single hiddenlayer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a 10-dimensional Clayton
copula (with parameter chosen such that Kendall's tau equals 0.5).
raw R object representing a GMMN (input and output
layer are 10-dimensional, the single hiddenlayer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a 10-dimensional Gumbel
copula (with parameter chosen such that Kendall's tau equals 0.5).
raw R object representing a GMMN (input and output
layer are 10-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a 10-dimensional nested Clayton copula
(with sector dimensions 5 and 5, corresponding Kendall's tau 0.5
and 0.75, and Kendall's tau 0.25 between the two sectors).
raw R object representing a GMMN (input and output
layer are 10-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a 10-dimensional nested Gumbel copula
(with sector dimensions 5 and 5, corresponding Kendall's tau 0.5
and 0.75, and Kendall's tau 0.25 between the two sectors).
raw R object representing a GMMN (input and output
layer are 10-dimensional, the single hidden layer is
300-dimensional) trained on 60000 pseudo-samples (with batch size 5000 and
300 epochs) from a 10-dimensional $t$ copula (with 4 degrees of freedom
and equi-correlation parameter chosen such that Kendall's tau equals
0.5).
Hofert, M., Prasad, A. and Zhu, M. (2018). Quasi-Monte Carlo for multivariate distributions via generative neural networks. (See https://arxiv.org/abs/1811.00683 for an early version)
GMMN_model(), to_callable()
# NOT RUN {
# to avoid win-builder error "Error: Installation of TensorFlow not found"
## Load a trained GMMN (see train_once())
NNname <- "GMMN_dim_2_300_2_ntrn_60000_nbat_5000_nepo_300_eqmix_C_tau_0.5_rot90_t4_tau_0.5"
NN <- read_rda(NNname, package = "gnn")
GMMN1 <- to_callable(NN)
str(GMMN1)
## Alternative
NNnm <- data(list = NNname)
GMMN2 <- to_callable(get(NNnm))
str(GMMN2)
## Check (the check-able components)
stopifnot(identical(GMMN1[names(GMMN1) != "model"],
GMMN2[names(GMMN2) != "model"]))
## Evaluate
set.seed(271)
N.prior <- matrix(rnorm(2000 * 2), ncol = 2)
X <- predict(GMMN1[["model"]], x = N.prior)
plot(X, xlab = expression(X[1]), ylab = expression(X[2]))
# }
Run the code above in your browser using DataLab