我的H2O群集版本是否与h2oEnsemble 1.9不兼容，或者在不使用贝叶斯函数的情况下导致此“朴素贝叶斯”错误的原因？

Question

The h2oEnsemble library's author, Dr. LeDell, suggested checking out v1.9 from the h2o-3 GitHub repo to work around a problem I and others encountered, which is described in this thread.

This led to a different error in my case, where only the unmodified wrappers can be used with h2oEnsemble::h2o.ensemble() .

For instance, learners like this will work:

learner <- c("h2o.deeplearning.wrapper", "h2o.randomForest.wrapper", "h2o.glm.wrapper") 

But anything custom like this:

h2o.glm.2 <- function(..., alpha = 0.5) h2o.glm.wrapper(..., alpha = alpha)
learner <- c("h2o.glm.2") 

ens2 <- h2o.ensemble(x             = setdiff(colnames(train), "y"),
                    y              = "y",
                    training_frame = trainHex,
                    metalearner    = "h2o.deeplearning.wrapper",
                    learner        = learner)

throws this error:

 Error in h2o.ensemble(x = setdiff(colnames(train), "y"), y = "y", training_frame = trainHex, : The Naive Bayes function does not support regression, please remove this function from your set of base learners 

Dr. LeDell suggested to confirm compatibility through her compatibility table on GitHub and to move further questions to StackOverflow (because they are shutting down their forum).

It seems that v1.9 is not on that table though and my H2O cluster version is also newer than what was listed there. Does anyone have anecdotal knowledge of which H2O clusters will work with v1.9 of the R package?

> h2o.getVersion()

"3.10.0.2"

Update: I tried upgrading to 3.12 and 3.13, but encountered the same error with this. Though I did notice that h2o::h2o.stackedEnsemble() works in 3.13.

Update:

The package author suggested version 3.10.5.2 in the comments/updated question. I tried updating to that, but sadly the problem still occurred. Here's my cluster info after the update:

> h2o.init(nthreads = -1) 

  Connection successful! R is connected to the H2O cluster: H2O cluster uptime: 2 minutes 28 seconds H2O cluster version: 3.10.5.2 H2O cluster version age: 15 days H2O cluster name: hackr H2O cluster total nodes: 1 H2O cluster total memory: 16.98 GB H2O cluster total cores: 24 H2O cluster allowed cores: 24 H2O cluster healthy: TRUE H2O Connection ip: localhost H2O Connection port: 54321 H2O Connection proxy: NA H2O Internal Security: FALSE R Version: R version 3.4.0 Patched (2017-05-19 r72713) 

Doesn't work with custom learner:

h2o.glm.2 <- function(..., alpha = 0.5) h2o.glm.wrapper(..., alpha = alpha)
learner <- c("h2o.glm.2", "h2o.deeplearning.wrapper", "h2o.randomForest.wrapper", "h2o.glm.wrapper")

ens <- h2o.ensemble(x              = setdiff(colnames(train), "y"),
                    y              = "y",
                    training_frame = trainHex,
                    metalearner    ="h2o.randomForest.wrapper",
                    learner        = learner)
summary(ens)

Error in h2o.ensemble(x = setdiff(colnames(train), "y"), y = "y", training_frame = trainHex, : The Naive Bayes function does not support regression, please remove this function from your set of base learners.

but works with only default wrapper learners:

learner <- c("h2o.deeplearning.wrapper", "h2o.randomForest.wrapper", "h2o.glm.wrapper")

ens <- h2o.ensemble(x              = setdiff(colnames(train), "y"),
                    y              = "y",
                    training_frame = trainHex,
                    metalearner    ="h2o.randomForest.wrapper",
                    learner        = learner)
summary(ens)

R package version is h2o_3.10.5.2 .

Answer 1

h2oEnsemble v0.1.9 package has not been officially released yet, but it's on the "h2oEnsemble_v0.1.9" branch here . On the README for that branch, the updated table shows that it will be compatible with H2O >=3.10.1.*.

It looks like you are doing a regression problem and attempting to use the h2o.naiveBayes.wrapper() function? The error message tells you that can't use Naive Bayes on a regression problem (Naive Bayes is a classification algorithm and does not support regression).

Using the "h2oEnsemble_v0.1.9" branch version of h2oEnsemble and the latest stable release of H2O, 3.10.5.2, I can replicate your error using the code below, but all you need to do to fix it is remove the Naive Bayes model from your base learners:

library(h2oEnsemble)  # Install from h2oEnsemble_v0.1.9 branch
h2o.init()

# Import a sample binary outcome train/test set into R
train <- h2o.importFile("https://s3.amazonaws.com/erin-data/higgs/higgs_train_5k.csv")
test <- h2o.importFile("https://s3.amazonaws.com/erin-data/higgs/higgs_test_5k.csv")
y <- "response"
x <- setdiff(names(train), y)
family <- "gaussian"


# Specify the base learner library & the metalearner
learner <- c("h2o.glm.wrapper", "h2o.randomForest.wrapper", 
             "h2o.gbm.wrapper", "h2o.naiveBayes.wrapper")
metalearner <- "h2o.glm.wrapper"

# Train the ensemble 
fit <- h2o.ensemble(x = x, y = y, 
                    training_frame = train, 
                    family = family, 
                    learner = learner, 
                    metalearner = metalearner)