Can we use h2o.train_segments for autoencoder?

13 views
Skip to first unread message

Fanwei Zeng

unread,
Feb 17, 2021, 8:11:51 AM2/17/21
to H2O Open Source Scalable Machine Learning - h2ostream
I have been using h2o's autoencoder (deeplearning with autoencoder=TRUE) and h2o.anomaly for anomaly detection. So far these have been performed on the entire dataset. However, now I need to build a separate model for each segment (subpopulation) of the dataset. I found h2o.train_segments so I tried it, but it gave me this error message:

Error in is.numeric(y) : argument "y" is missing, with no default

Does this mean that h2o.train_segments doesn't work for h2o's autoencoder?

Here is my code:

mod <- h2o.train_segments(algorithm = "deeplearning",
                                                  segment_columns = "some_col",
                                                  parallelism = 1,
                                                  training_frame = data.hex,
                                                  activation = "tanh",
                                                  standardize = TRUE,
                                                  hidden = c(10, 1, 10),
                                                  epochs = 5,
                                                  autoencoder = TRUE,
                                                  loss = "CrossEntropy",
                                                  seed = 42)

Thank you for your insights!
 
Reply all
Reply to author
Forward
0 new messages