Skip to content

Multilayer perceptron via h2o

Source: R/mlp_h2o.R
details_mlp_h2o.Rd

h2o::h2o.deeplearning() fits a feed-forward neural network.

Details

For this engine, there are multiple modes: classification and regression

Tuning Parameters

This model has 6 tuning parameters:

  • hidden_units: # Hidden Units (type: integer, default: 200L)

  • penalty: Amount of Regularization (type: double, default: 0.0)

  • dropout: Dropout Rate (type: double, default: 0.5)

  • epochs: # Epochs (type: integer, default: 10)

  • activation: Activation function (type: character, default: ‘see below’)

  • learn_rate: Learning Rate (type: double, default: 0.005)

The naming of activation functions in h2o::h2o.deeplearning() differs from parsnip’s conventions. Currently, only “relu” and “tanh” are supported and will be converted internally to “Rectifier” and “Tanh” passed to the fitting function.

penalty corresponds to l2 penalty. h2o::h2o.deeplearning() also supports specifying the l1 penalty directly with the engine argument l1.

Other engine arguments of interest:

  • stopping_rounds controls early stopping rounds based on the convergence of another engine parameter stopping_metric. By default, h2o::h2o.deeplearning stops training if simple moving average of length 5 of the stopping_metric does not improve for 5 scoring events. This is mostly useful when used alongside the engine parameter validation, which is the proportion of train-validation split, parsnip will split and pass the two data frames to h2o. Then h2o::h2o.deeplearning will evaluate the metric and early stopping criteria on the validation set.

  • h2o uses a 50% dropout ratio controlled by dropout for hidden layers by default. h2o::h2o.deeplearning() provides an engine argument input_dropout_ratio for dropout ratios in the input layer, which defaults to 0.

Translation from parsnip to the original package (regression)

agua::h2o_train_mlp is a wrapper around h2o::h2o.deeplearning().

mlp(
  hidden_units = integer(1),
  penalty = double(1),
  dropout = double(1),
  epochs = integer(1),
  learn_rate = double(1),
  activation = character(1)
) %>%
  set_engine("h2o") %>%
  set_mode("regression") %>%
  translate()

## Single Layer Neural Network Model Specification (regression)
##
## Main Arguments:
##   hidden_units = integer(1)
##   penalty = double(1)
##   dropout = double(1)
##   epochs = integer(1)
##   activation = character(1)
##   learn_rate = double(1)
##
## Computational engine: h2o
##
## Model fit template:
## agua::h2o_train_mlp(x = missing_arg(), y = missing_arg(), weights = missing_arg(),
##     validation_frame = missing_arg(), hidden = integer(1), l2 = double(1),
##     hidden_dropout_ratios = double(1), epochs = integer(1), activation = character(1),
##     rate = double(1))

Translation from parsnip to the original package (classification)

mlp(
  hidden_units = integer(1),
  penalty = double(1),
  dropout = double(1),
  epochs = integer(1),
  learn_rate = double(1),
  activation = character(1)
) %>%
  set_engine("h2o") %>%
  set_mode("classification") %>%
  translate()

## Single Layer Neural Network Model Specification (classification)
##
## Main Arguments:
##   hidden_units = integer(1)
##   penalty = double(1)
##   dropout = double(1)
##   epochs = integer(1)
##   activation = character(1)
##   learn_rate = double(1)
##
## Computational engine: h2o
##
## Model fit template:
## agua::h2o_train_mlp(x = missing_arg(), y = missing_arg(), weights = missing_arg(),
##     validation_frame = missing_arg(), hidden = integer(1), l2 = double(1),
##     hidden_dropout_ratios = double(1), epochs = integer(1), activation = character(1),
##     rate = double(1))

Preprocessing requirements

Factor/categorical predictors need to be converted to numeric values (e.g., dummy or indicator variables) for this engine. When using the formula method via fit(), parsnip will convert factor columns to indicators.

Predictors should have the same scale. One way to achieve this is to center and scale each so that each predictor has mean zero and a variance of one.

By default, h2o::h2o.deeplearning() uses the argument standardize = TRUE to center and scale all numeric columns.

Initializing h2o

To use the h2o engine with tidymodels, please run h2o::h2o.init() first. By default, This connects R to the local h2o server. This needs to be done in every new R session. You can also connect to a remote h2o server with an IP address, for more details see h2o::h2o.init().

You can control the number of threads in the thread pool used by h2o with the nthreads argument. By default, it uses all CPUs on the host. This is different from the usual parallel processing mechanism in tidymodels for tuning, while tidymodels parallelizes over resamples, h2o parallelizes over hyperparameter combinations for a given resample.

h2o will automatically shut down the local h2o instance started by R when R is terminated. To manually stop the h2o server, run h2o::h2o.shutdown().

Saving fitted model objects

Models fitted with this engine may require native serialization methods to be properly saved and/or passed between R sessions. To learn more about preparing fitted models for serialization, see the bundle package.