SPOTMisc (version 1.19.52)

funKerasMnist: funKerasMnist

Description

Hyperparameter Tuning: Keras MNIST Classification Test Function.

Usage

funKerasMnist(x, kerasConf, data)

Value

1-column matrix with resulting function values (test loss)

Arguments

x

matrix of hyperparameter values to evaluate with the function. Rows for points and columns for dimension.

kerasConf

List of additional parameters passed to keras as described in getKerasConf. Default: kerasConf = getKerasConf().

data

mnist data set. Default: getMnistData.

Details

Trains a simple deep NN on the MNIST dataset. Provides a template that can be used for other networks as well. Standard Code from https://tensorflow.rstudio.com/ Modified by T. Bartz-Beielstein ([e-mail])

See Also

getKerasConf

evalKerasMnist

fit

Examples

# \donttest{
### These examples require an activated Python environment as described in
### Bartz-Beielstein, T., Rehbach, F., Sen, A., and Zaefferer, M.:
### Surrogate Model Based Hyperparameter Tuning for Deep Learning with SPOT,
### June 2021. http://arxiv.org/abs/2105.14625.
PYTHON_RETICULATE <- FALSE
if(PYTHON_RETICULATE){

library("SPOTMisc")
library("SPOT")
kerasConf <- getKerasConf()
## The following two settings are default:
kerasConf$encoding = "oneHot"
kerasConf$model = "dl"
## get the data with the correct encoding
mnist <- getMnistData(kerasConf)
## get the model
cfg <-  getModelConf(kerasConf)

### First example: simple function call:
x <- matrix(cfg$default, nrow=1)
if (length(cfg$transformations) > 0) {  x <- transformX(xNat=x, fn=cfg$transformations)}
funKerasMnist(x, kerasConf = kerasConf, data = mnist)
### Use convnet:
kerasConf <- getKerasConf()
kerasConf$encoding <- "tensor"
kerasConf$model <- "cnn"
mnist <- getMnistData(kerasConf)
cfg <-  getModelConf(kerasConf)
x <- matrix(cfg$default, nrow=1)
if (length(cfg$transformations) > 0) {  x <- transformX(xNat=x, fn=cfg$transformations)}
funKerasMnist(x, kerasConf = kerasConf, data=mnist)

### Second example: evaluation of several (three) hyperparameter settings:
x <- matrix(cfg$default, nrow=1)
if (length(cfg$transformations) > 0) {  x <- transformX(xNat=x, fn=cfg$transformations)}
xxx <- rbind(x,x,x)
funKerasMnist(xxx, kerasConf = kerasConf, data=mnist)

### Third example: spot call (dense network):
kerasConf <- getKerasConf()
kerasConf$verbose <- 0
kerasConf$encoding = "oneHot"
kerasConf$model = "dl"
## get the data with the correct encoding
mnist <- getMnistData(kerasConf)
## get the model
cfg <-  getModelConf(kerasConf)
## max 32 training epochs
cfg$upper[6] <- 5
resDl <- spot(x = NULL,
            fun = funKerasMnist,
            lower = cfg$lower,
            upper = cfg$upper,
            control = list(funEvals=15,
                         transformFun = cfg$transformations,
                         types = cfg$type,
                         noise = TRUE,
                         plots = TRUE,
                         progress = TRUE,
                         seedFun = 1,
                         seedSPOT = 1),
                         kerasConf = kerasConf,
                         data = mnist)

### Fourth example: spot call (convnet):
kerasConf <- getKerasConf()
kerasConf$verbose <- 1
kerasConf$encoding <- "tensor"
kerasConf$model <- "cnn"
## get the data with the correct encoding
mnist <- getMnistData(kerasConf)
## get the model
cfg <-  getModelConf(kerasConf)
## max 32 training epochs
cfg$upper[6] <- 5
resCnn <- spot(x = NULL,
            fun = funKerasMnist,
            lower = cfg$lower,
            upper = cfg$upper,
            control = list(funEvals=15,
                         transformFun = cfg$transformations,
                         types = cfg$type,
                         noise = TRUE,
                         plots = TRUE,
                         progress = TRUE,
                         seedFun = 1,
                         seedSPOT = 1),
                         kerasConf = kerasConf,
                         data = mnist)
}
# }