Migrating to H2O 3¶

Supervised¶

• Generalized Linear Model (GLM): Binomial classification, multinomial classification, regression (including logistic regression)
• Distributed Random Forest (DRF): Binomial classification, multinomial classification, regression
• Gradient Boosting Machine (GBM): Binomial classification, multinomial classification, regression
• Deep Learning (DL): Binomial classification, multinomial classification, regression

Unsupervised¶

• K-means
• Principal Component Analysis
• Autoencoder

There are a few algorithms that are still being refined to provide these same benefits and will be available in a future version of H2O.

Currently, the following algorithms and associated capabilities are still in development:

• Naïve Bayes

Check back for updates, as these algorithms will be re-introduced in an improved form in a future version of H2O.

Note: The SpeeDRF model has been removed, as it was originally intended as an optimization for small data only. This optimization will be added to the Distributed Random Forest model automatically for small data in a future version of H2O.

Porting R Scripts¶

This document outlines how to port R scripts written in previous versions of H2O (Nunes 2.8.6.2 or prior, also known as “H2O Classic”) for compatibility with the new H2O 3 API. When upgrading from H2O to H2O 3, most functions are the same. However, there are some differences that will need to be resolved when porting any scripts that were originally created using H2O to H2O 3.

The original R script for H2O is listed first, followed by the updated script for H2O 3.

Some of the parameters have been renamed for consistency. For each algorithm, a table that describes the differences is provided.

For additional assistance within R, enter a question mark before the command (for example, ?h2o.glm).

There is also a “shim” available that will review R scripts created with previous versions of H2O, identify deprecated or renamed parameters, and suggest replacements. For more information, refer to the repo here.

The simple way¶

This is the easiest way to change your local repo and is recommended for most users.

1. Enter git remote -v to view a list of your repositories.

2. Copy the address of your H2O classic repo (refer to the text in brackets below - your address will vary depending on your connection method):

   H2O_User-MBP:h2o H2O_User$ git remote -v
   origin    https://{H2O_User@github.com}/h2oai/h2o.git (fetch)
   origin    https://{H2O_User@github.com}/h2oai/h2o.git (push)
   

3. Enter git remote set-url origin {H2O_User@github.com}:h2oai/h2o-2.git, where {H2O_User@github.com} represents the address copied in the previous step.

The more complicated way¶

This method involves editing the Github config file and should only be attempted by users who are confident enough with their knowledge of Github to do so.

1. Enter vim .git/config.

2. Look for the [remote "origin"] section:

   [remote "origin"]
   url = https://H2O_User@github.com/h2oai/h2o.git
   fetch = +refs/heads/*:refs/remotes/origin/*
   

3. In the url = line, change h2o.git to h2o-2.git.

4. Save the changes.

The latest version of H2O is stored in the h2o-3 repository. All previous links to this repo will still work, but if you would like to manually update your Github configuration, follow the instructions above, replacing h2o-2 with h2o-3.

h2o.exec¶

The h2o.exec command is no longer supported. Any workflows using h2o.exec must be revised to remove this command. If the H2O 3 workflow contains any parameters or commands from H2O Classic, errors will result and the workflow will fail.

The purpose of h2o.exec was to wrap expressions so that they could be evaluated in a single \Exec2 call. For example, h2o.exec(fr[,1] + 2/fr[,3]) and fr[,1] + 2/fr[,3] produced the same results in H2O. However, the first example makes a single REST call and uses a single temp object, while the second makes several REST calls and uses several temp objects.

Due to the improved architecture in H2O 3, the need to use h2o.exec has been eliminated, as the expression can be processed by R as an “unwrapped” typical R expression.

Currently, the only known exception is when factor is used in conjunction with h2o.exec. For example, h2o.exec(fr$myIntCol <- factor(fr$myIntCol)) would become fr$myIntCol <- as.factor(fr$myIntCol)

Note also that an array is not inside a string:

An int array is [1, 2, 3], not “[1, 2, 3]”.

A String array is [“f00”, “b4r”], not “[“f00”, “b4r”]”

Only string values are enclosed in double quotation marks (").

h2o.performance¶

To access any exclusively binomial output, use h2o.performance, optionally with the corresponding accessor. The accessor can only use the model metrics object created by h2o.performance. Each accessor is named for its corresponding field (for example, h2o.AUC, h2o.gini, h2o.F1). h2o.performance supports all current algorithms except for K-Means.

If you specify a data frame as a second parameter, H2O will use the specified data frame for scoring. If you do not specify a second parameter, the training metrics for the model metrics object are used.

xval and validation slots¶

The xval slot has been removed, as nfolds is not currently supported.

The validation slot has been merged with the model slot.

Principal Components Regression (PCR)¶

Principal Components Regression (PCR) has also been deprecated. To obtain PCR values, create a Principal Components Analysis (PCA) model, then create a GLM model from the scored data from the PCA model.

Saving and Loading Models¶

Saving and loading a model from R is supported in version 3.0.0.18 and later. H2O 3 uses the same binary serialization method as previous versions of H2O, but saves the model and its dependencies into a directory, with each object as a separate file. The save_CV option for available in previous versions of H2O has been deprecated, as h2o.saveAll and h2o.loadAll are not currently supported. The following commands are now supported:

• h2o.saveModel
• h2o.loadModel

Algorithm Updates

• GBM
• GLM
• K-Means
• Deep Learning
• Distributed Random Forest

Renamed GBM Parameters¶

The following parameters have been renamed, but retain the same functions:

Deprecated GBM Parameters¶

The following parameters have been removed:

• group_split: Bit-set group splitting of categorical variables is now the default.
• importance: Variable importances are now computed automatically and displayed in the model output.
• holdout.fraction: The fraction of the training data to hold out for validation is no longer supported.
• grid.parallelism: Specifying the number of parallel threads to run during a grid search is no longer supported.

New GBM Parameters¶

The following parameters have been added:

• seed: A random number to control sampling and initialization when balance_classes is enabled.
• score_each_iteration: Display error rate information after each tree in the requested set is built.
• build_tree_one_node: Run on a single node to use fewer CPUs.

GBM Algorithm Comparison¶

Output¶

The following table provides the component name in H2O, the corresponding component name in H2O 3 (if supported), and the model type (binomial, multinomial, or all). Many components are now included in h2o.performance; for more information, refer to h2o.performance.

Renamed GLM Parameters¶

The following parameters have been renamed, but retain the same functions:

Deprecated GLM Parameters¶

The following parameters have been removed:

• return_all_lambda: A logical value indicating whether to return every model built during the lambda search. (may be re-added)
• higher_accuracy: For improved accuracy, adjust the beta_epsilon value.
• strong_rules: Discards predictors likely to have 0 coefficients prior to model building. (may be re-added as enabled by default)
• non_negative: Specify a non-negative response. (may be re-added)
• variable_importances: Variable importances are now computed automatically and displayed in the model output. They have been renamed to Normalized Coefficient Magnitudes.
• disable_line_search: This parameter has been deprecated, as it was mainly used for testing purposes.
• max_predictors: Stops training the algorithm if the number of predictors exceeds the specified value. (may be re-added)

New GLM Parameters¶

The following parameters have been added:

• validation_frame: Specify the validation dataset.
• solver: Select IRLSM or LBFGS.

GLM Algorithm Comparison¶

Output¶

The following table provides the component name in H2O, the corresponding component name in H2O 3 (if supported), and the model type (binomial, multinomial, or all). Many components are now included in h2o.performance; for more information, refer to h2o.performance.

Renamed K-Means Parameters¶

The following parameters have been renamed, but retain the same functions:

Note In H2O, the normalize parameter was disabled by default. The standardize parameter is enabled by default in H2O 3 to provide more accurate results for datasets containing columns with large values.

New K-Means Parameters¶

The following parameters have been added:

• user has been added as an additional option for the init parameter. Using this parameter forces the K-Means algorithm to start at the user-specified points.
• user_points: Specify starting points for the K-Means algorithm.

K-Means Algorithm Comparison¶

Output¶

The following table provides the component name in H2O and the corresponding component name in H2O 3 (if supported).

Renamed Deep Learning Parameters¶

The following parameters have been renamed, but retain the same functions:

Deprecated DL Parameters¶

The following parameters have been removed:

• classification: Classification is now inferred from the data type.
• holdout_fraction: Fraction of the training data to hold out for validation.
• dlmodel@model$best_cutoff: This output parameter has been removed.

New DL Parameters¶

The following parameters have been added:

• export_weights_and_biases: An additional option allowing users to export the raw weights and biases as H2O frames.

The following options for the loss parameter have been added:

• absolute: Provides strong penalties for mispredictions
• huber: Can improve results for regression

DL Algorithm Comparison¶

Output¶

The following table provides the component name in H2O, the corresponding component name in H2O 3 (if supported), and the model type (binomial, multinomial, or all). Many components are now included in h2o.performance; for more information, refer to h2o.performance.

Changes to DRF in H2O 3¶

Distributed Random Forest (DRF) was represented as h2o.randomForest(type="BigData", ...) in H2O Classic. In H2O Classic, SpeeDRF (type="fast") was not as accurate, especially for complex data with categoricals, and did not address regression problems. DRF (type="BigData") was at least as accurate as SpeeDRF (type="fast") and was the only algorithm that scaled to big data (data too large to fit on a single node). In H2O 3, we improved the performance of DRF so that the data fits on a single node (optimally, for all cases), which will make SpeeDRF obsolete. Ultimately, the goal is provide a single algorithm that provides the “best of both worlds” for all datasets and use cases. Please note that H2O does not currently support the ability to specify the number of trees when using h2o.predict for a DRF model.

Note: H2O 3 only supports DRF. SpeeDRF is no longer supported. The functionality of DRF in H2O 3 is similar to DRF functionality in H2O.

Renamed DRF Parameters¶

The following parameters have been renamed, but retain the same functions:

Deprecated DRF Parameters¶

The following parameters have been removed:

• classification: This is now automatically inferred from the response type. To achieve classification with a 0/1 response column, explicitly convert the response to a factor (as.factor()).
• importance: Variable importances are now computed automatically and displayed in the model output.
• holdout.fraction: Specifying the fraction of the training data to hold out for validation is no longer supported.
• doGrpSplit: The bit-set group splitting of categorical variables is now the default.
• verbose: Infonrmation about tree splits and extra statistics is now included automatically in the stdout.
• oobee: The out-of-bag error estimate is now computed automatically (if no validation set is specified).
• stat.type: This parameter was used for SpeeDRF, which is no longer supported.
• type: This parameter was used for SpeeDRF, which is no longer supported.

New DRF Parameters¶

The following parameter has been added:

• build_tree_one_node: Run on a single node to use fewer CPUs.

DRF Algorithm Comparison¶

Output¶

The following table provides the component name in H2O, the corresponding component name in H2O 3 (if supported), and the model type (binomial, multinomial, or all). Many components are now included in h2o.performance; for more information, refer to h2o.performance.