GLM java.lang.ArrayIndexOutOfBoundsException for model with all Factor Variables

[awil...@gmail.com]

Hello,

Using H2O 3.10.4.8 on Ubuntu 14.04.4 with Python 3.6.0 (I also experience this on Python 3.5.2), if I try to fit a large GLM with only factor variables, it fails with a java.lang.ArrayIndexOutOfBoundsException. If I add a single continuous predictor to the model that fails, it runs fine.

I am able to recreate the problem with the example below:

import pandas as pd
import numpy as np
import h2o
h2o.init()

np.random.seed(1234)
x1 = np.random.randint(1,193,33000*5)
x2 = np.random.randint(1,193,33000*5)
x3 = np.random.randint(1,8,33000*5)
x4 = np.random.randint(1,8,33000*5)
x5 = np.random.random(33000*5)*10
n = np.random.randint(1,100000,33000*5)
y = np.random.random(33000*5)*n
X = pd.DataFrame({'y' : y, 'n' : n, 'x1' : x1, 'x2' : x2, 'x3' : x3, 'x4' : x4, 'x5' : x5, 'offset':np.log(n)})

X_h2o = h2o.H2OFrame(X)
X_h2o['x1'] = X_h2o['x1'].asfactor()
X_h2o['x2'] = X_h2o['x2'].asfactor()
X_h2o['x3'] = X_h2o['x3'].asfactor()
X_h2o['x4'] = X_h2o['x4'].asfactor()

## This fails with java.lang.ArrayIndexOutOfBoundsException
predictors = ['x1', 'x2', 'x3', 'x4']
ix_list = ['x1','x2']
from h2o.estimators.glm import H2OGeneralizedLinearEstimator
poisson = H2OGeneralizedLinearEstimator(family="poisson", alpha=0.025, lambda_=0.0001, model_id='sem_poisson',\
standardize=False, interactions = ix_list, intercept=False)
poisson.train(training_frame=X_h2o, y='y', x=predictors, offset_column="offset")

## This runs without error
predictors = ['x1', 'x2', 'x3', 'x4', 'x5']
ix_list = ['x1','x2']
from h2o.estimators.glm import H2OGeneralizedLinearEstimator
poisson = H2OGeneralizedLinearEstimator(family="poisson", alpha=0.025, lambda_=0.0001, model_id='sem_poisson',\
standardize=False, interactions = ix_list, intercept=False)
poisson.train(training_frame=X_h2o, y='y', x=predictors, offset_column="offset")

This is the error message:
---------------------------------------------------------------------------
OSError Traceback (most recent call last)
<ipython-input-74-3b9ec76c023b> in <module>()
9 from h2o.estimators.glm import H2OGeneralizedLinearEstimator
10 poisson = H2OGeneralizedLinearEstimator(family="poisson", alpha=0.025, lambda_=0.0001, model_id='sem_poisson', standardize=False, interactions = ix_list, intercept=False)
---> 11 poisson.train(training_frame=X_h2o, y='y', x=predictors, offset_column="offset")

/home/acorreia/anaconda3/lib/python3.6/site-packages/h2o/estimators/estimator_base.py in train(self, x, y, training_frame, offset_column, fold_column, weights_column, validation_frame, max_runtime_secs, ignored_columns, model_id)
202 return
203
--> 204 model.poll()
205 model_json = h2o.api("GET /%d/Models/%s" % (rest_ver, model.dest_key))["models"][0]
206 self._resolve_model(model.dest_key, model_json)

/home/acorreia/anaconda3/lib/python3.6/site-packages/h2o/job.py in poll(self)
71 if (isinstance(self.job, dict)) and ("stacktrace" in list(self.job)):
72 raise EnvironmentError("Job with key {} failed with an exception: {}\nstacktrace: "
---> 73 "\n{}".format(self.job_key, self.exception, self.job["stacktrace"]))
74 else:
75 raise EnvironmentError("Job with key %s failed with an exception: %s" % (self.job_key, self.exception))

OSError: Job with key $03017f00000132d4ffffffff$_ac26dc0f4f46e174f756718a6ed744a4 failed with an exception: java.lang.ArrayIndexOutOfBoundsException
stacktrace:
java.lang.ArrayIndexOutOfBoundsException

For the problem I have when I'm using my real data set (which I'm unable to share), this is the tail end of the slightly different stacktrace that I sometimes see, in addition to the one above:
---------------------------------------------------------------------------
/home/acorreia/anaconda3/lib/python3.6/site-packages/h2o/estimators/estimator_base.py in train(self, x, y, training_frame, offset_column, fold_column, weights_column, validation_frame, max_runtime_secs, ignored_columns, model_id)
202 return
203
--> 204 model.poll()
205 model_json = h2o.api("GET /%d/Models/%s" % (rest_ver, model.dest_key))["models"][0]
206 self._resolve_model(model.dest_key, model_json)

/home/acorreia/anaconda3/lib/python3.6/site-packages/h2o/job.py in poll(self)
71 if (isinstance(self.job, dict)) and ("stacktrace" in list(self.job)):
72 raise EnvironmentError("Job with key {} failed with an exception: {}\nstacktrace: "
---> 73 "\n{}".format(self.job_key, self.exception, self.job["stacktrace"]))
74 else:
75 raise EnvironmentError("Job with key %s failed with an exception: %s" % (self.job_key, self.exception))

OSError: Job with key $03017f00000132d4ffffffff$_a243f5bd946a68975109a5e8c1fc4b4c failed with an exception: java.lang.ArrayIndexOutOfBoundsException: 1
stacktrace:
java.lang.ArrayIndexOutOfBoundsException: 1
at hex.DataInfo.filterExpandedColumns(DataInfo.java:485)
at hex.glm.ComputationState.applyStrongRules(ComputationState.java:175)
at hex.glm.ComputationState.setLambda(ComputationState.java:74)
at hex.glm.GLM$GLMDriver.computeSubmodel(GLM.java:1040)
at hex.glm.GLM$GLMDriver.computeImpl(GLM.java:1117)
at hex.ModelBuilder$Driver.compute2(ModelBuilder.java:169)
at hex.glm.GLM$GLMDriver.compute2(GLM.java:542)
at water.H2O$H2OCountedCompleter.compute(H2O.java:1315)
at jsr166y.CountedCompleter.exec(CountedCompleter.java:468)
at jsr166y.ForkJoinTask.doExec(ForkJoinTask.java:263)
at jsr166y.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:974)
at jsr166y.ForkJoinPool.runWorker(ForkJoinPool.java:1477)
at jsr166y.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:104)

[awil...@gmail.com]

I'm realizing a possible a solution (although perhaps inefficient?) is to explicitly code the interaction and include it as another variable in the data set. But, if this works, I'd expect the interaction to also work. Appreciate any thoughts on this.

import pandas as pd
import numpy as np
import h2o
h2o.init()

np.random.seed(1234)
x1 = np.random.randint(1,193,33000*5)
x2 = np.random.randint(1,193,33000*5)
x3 = np.random.randint(1,8,33000*5)
x4 = np.random.randint(1,8,33000*5)
x5 = np.random.random(33000*5)*10
n = np.random.randint(1,100000,33000*5)
y = np.random.random(33000*5)*n
X = pd.DataFrame({'y' : y, 'n' : n, 'x1' : x1, 'x2' : x2, 'x3' : x3, 'x4' : x4, 'x5' : x5, 'offset':np.log(n)})

u_ixs = X[['x1', 'x2']].drop_duplicates()
u_ixs['x1x2_combo'] = np.arange(u_ixs.shape[0])+1
X = X.merge(u_ixs)

X_h2o = h2o.H2OFrame(X)
X_h2o['x1'] = X_h2o['x1'].asfactor()
X_h2o['x2'] = X_h2o['x2'].asfactor()
X_h2o['x3'] = X_h2o['x3'].asfactor()
X_h2o['x4'] = X_h2o['x4'].asfactor()

X_h2o['x1x2_combo'] = X_h2o['x1x2_combo'].asfactor()

## No errors
predictors = ['x1', 'x2', 'x3', 'x4', 'x1x2_combo']

from h2o.estimators.glm import H2OGeneralizedLinearEstimator
poisson = H2OGeneralizedLinearEstimator(family="poisson", alpha=0.025, lambda_=0.0001, model_id='sem_poisson',\

standardize=False, intercept=False, max_iterations=500)

[carlo....@gmail.com]

I am having the same problem.

[Darren Cook]

> I am having the same problem.

Is it the exact same error? What version of H2O are you using?

The first code the OP posted is clearly showing a bug: an out of bounds
exception with ['x1', 'x2', 'x3', 'x4'] but it works okay with ['x1',
'x2', 'x3', 'x4', 'x5'].

If the exact same error, does the workaround (in the code you quoted),
of hand-coding the interactions, work for you too?

Darren

--
Darren Cook, Software Researcher/Developer
My New Book: Practical Machine Learning with H2O:
http://shop.oreilly.com/product/0636920053170.do

[carlo....@gmail.com]

I am running h2o 3.16.0.2 on python 2.7.13. I replicated Awil error example on my machine. Originally I was stepping in the same ArrayIndexOutOfBoundsException with the following setting:

model = H2OGeneralizedLinearEstimator(interactions=interactions_list, compute_p_values=True, lambda_=0,
solver='IRLSM',
remove_collinear_columns=True)

In the latter case adding a continuous variable did not help.