Add N-fold CV, use GBM for MNIST.

R/examples/H2ODeepLearningNextML.R

@@ -60,6 +60,7 @@ adult_hex <- h2o.importFile(h2oServer, "adult.gz")
 dim(adult_hex)
 summary(adult_hex)
 
+
 ## Set column names manually
 colnames(adult_hex) <- c("age","workclass","fnlwgt","education","education-num","marital-status",
                          "occupation","relationship","race","sex","capital-gain","capital-loss",
@@ -82,11 +83,11 @@ best_model@model$params$activation
 best_model@model$params$l1
 best_model@model$params$hidden
 
-
-## Compare to GBM
-gbm_model <- h2o.gbm(data=adult_hex, x=pred, y=resp)
-gbm_model
-
+## Compute N-fold CV error for the best model
+cv_model <- h2o.deeplearning(data=adult_hex, x=pred, y="income",
+                             activation=c("Rectifier"), l1=best_model@model$params$l1,
+                             hidden=best_model@model$params$hidden, epochs=10, nfold=5)
+cv_model
 
 
 ## MNIST
@@ -97,8 +98,11 @@ mnist_test_hex <- h2o.importFile(h2oServer, "mnist.test.csv.gz")
 dim(mnist_train_hex)
 dim(mnist_test_hex)
 
-mnist_model <- h2o.deeplearning(data=mnist_train_hex, validation=mnist_test_hex, hidden=c(20,20,20), epochs=1, x=1:784, y=785)
-mnist_model
+dl_model <- h2o.deeplearning(data=mnist_train_hex, validation=mnist_test_hex, hidden=c(20,20,20), epochs=1, x=1:784, y=785)
+dl_model
 
+## Compare to GBM
+gbm_model <- h2o.gbm(data=mnist_train_hex, validation=mnist_test_hex, x=1:784, y=785, distribution="multinomial", n.trees=5)
+gbm_model
 
 ## For more examples, see http://learn.h2o.ai/content/