learn_rate_annealing

  • Available in: GBM

  • Hyperparameter: yes

Description

Use this option to reduce the learning rate by this factor after every tree. When used, then for N trees, you would start with learn_rate and end with learn_rate * learn_rate_annealing^N.

The following provides some reference factors. (Refer to Taylor series for more information.):

  • 0.99^100 = 0.366

  • 0.99^1000 = 4.3e-5

  • 0.999^1000 = 0.368

  • 0.999^10000 = 4.5e-5

With this option, then instead of learn_rate=0.01, you can try (for example) learn_rate=0.05 along with learn_rate_annealing=0.99. The result should converge much faster with almost the same accuracy. Note, however, that this can also result in overfitting, so use caution when specifying this option.

The value range for this option is between 0 and 1. This option defaults to 1.0, which disables the learning rate annealing.

Example

R

Python

library(h2o)
h2o.init()
# import the titanic dataset:
# This dataset is used to classify whether a passenger will survive '1' or not '0'
# original dataset can be found at https://stat.ethz.ch/R-manual/R-devel/library/datasets/html/Titanic.html
titanic <-  h2o.importFile("https://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/titanic.csv")

# convert response column to a factor
titanic['survived'] <- as.factor(titanic['survived'])

# set the predictor names and the response column name
# predictors include all columns except 'name' and the response column ("survived")
predictors <- setdiff(colnames(titanic), colnames(titanic)[2:3])
response <- "survived"

# split into train and validation
titanic.splits <- h2o.splitFrame(data =  titanic, ratios = .8, seed = 1234)
train <- titanic.splits[[1]]
valid <- titanic.splits[[2]]

# try using the `learn_rate_annealing` parameter:
# combine learn_rate with learn_rate_annealing
# since we have learning_rate_annealing, we can afford to start with a bigger learning rate (.05)
# learning rate annealing = .99 means learning_rate shrinks by 1% after every tree
# early stopping makes it okay to use 'more than enough' trees
titanic.gbm <- h2o.gbm(x = predictors, y = response, training_frame = train, validation_frame = valid,
                       learn_rate = .05, learn_rate_annealing =.99,
                       # use early stopping once the validation AUC doesn't improve by at least 0.01%
                       # for 5 consecutive scoring events
                       stopping_rounds = 5,
                       stopping_tolerance = 1e-4,
                       stopping_metric = "AUC", seed = 1234)

# print the auc for the validation data
print(h2o.auc(titanic.gbm, valid = TRUE))

import h2o
from h2o.estimators.gbm import H2OGradientBoostingEstimator
h2o.init()

# import the titanic dataset:
# This dataset is used to classify whether a passenger will survive '1' or not '0'
# original dataset can be found at https://stat.ethz.ch/R-manual/R-devel/library/datasets/html/Titanic.html
titanic = h2o.import_file("https://s3.amazonaws.com/h2o-public-test-data/smalldata/gbm_test/titanic.csv")

# convert response column to a factor
titanic['survived'] = titanic['survived'].asfactor()

# set the predictor names and the response column name
# predictors include all columns except 'name' and the response column ("survived")
predictors = titanic.columns
del predictors[1:3]
response = 'survived'

# split into train and validation sets
train, valid = titanic.split_frame(ratios = [.8], seed = 1234)

# try using the `learn_rate_annealing` parameter:
# combine learn_rate with learn_rate_annealing
# since we have learning_rate_annealing, we can afford to start with a bigger learning rate (.05)
# learning rate annealing = .99 means learning_rate shrinks by 1% after every tree
# early stopping makes it okay to use 'more than enough' trees
# initialize your estimator
titanic_gbm = H2OGradientBoostingEstimator(ntrees = 10000, learn_rate = 0.05, learn_rate_annealing = .99,
                                           # use early stopping once the validation AUC doesn't improve
                                           # by at least 0.01% for 5 consecutive scoring events
                                           stopping_rounds = 5, stopping_metric = "auc",
                                           stopping_tolerance = 1e-4, seed = 1234)

# then train the model
titanic_gbm.train(x = predictors, y = response, training_frame = train, validation_frame = valid)

# print the auc for the validation data
print(titanic_gbm.auc(valid=True))