Is DatasetInfo a necessary parameter for Hoeffding Tree?

[xlindo]

In hoeffding_tree.hpp, there is a default constructor without parameters, and also a Train method does not need DatasetInfo.
But when using these two to avoid passing DatasetInfo, like,

 std::shared_ptr<mlpack::tree::HoeffdingTree<>> mlpack_ht(
                                                  new mlpack::tree::HoeffdingTree<>());
// ... Loading data
try
{
    mlpack_ht->Train(mlpack_texture_train_X, tmp_labels);
}
catch (std::exception& e)
{
    std::cerr << e.what() << std::endl << std::endl;
}

an exception is thrown.

requested type of dimension 0, but dataset only has 0 dimensions

So is DatasetInfo a necessary parameter for Hoeffding Tree? Or is it possible to train a Hoeffding tree without passing in DatasetInfo explicitly?

[rcurtin]

Hey @xlindo, thanks for reporting this! I played with the example and found that this is indeed a bug in the handling of new datasets. I opened a PR, #2964, which should fix the issue. If you'd like to try your code on that branch and see if it works, that would be great!

In any case, to answer the original question, it should be possible to train a Hoeffding tree without a data::DatasetInfo. But, due to a bug in the code, that does not currently work. It will, though, once #2964 is merged. 😄

[xlindo]

Hey @xlindo, thanks for reporting this! I played with the example and found that this is indeed a bug in the handling of new datasets. I opened a PR, #2964, which should fix the issue. If you'd like to try your code on that branch and see if it works, that would be great!

In any case, to answer the original question, it should be possible to train a Hoeffding tree without a data::DatasetInfo. But, due to a bug in the code, that does not currently work. It will, though, once #2964 is merged. 😄

Thanks a lot!
I will try it now.

[xlindo]

Hey @xlindo, thanks for reporting this! I played with the example and found that this is indeed a bug in the handling of new datasets. I opened a PR, #2964, which should fix the issue. If you'd like to try your code on that branch and see if it works, that would be great!

In any case, to answer the original question, it should be possible to train a Hoeffding tree without a data::DatasetInfo. But, due to a bug in the code, that does not currently work. It will, though, once #2964 is merged. 😄

For consistency with other classification algorithms, I thought there can be an optional parameter num_classes=0 of Train(...).

[rcurtin]

Hey @xlindo, thanks for reporting this! I played with the example and found that this is indeed a bug in the handling of new datasets. I opened a PR, #2964, which should fix the issue. If you'd like to try your code on that branch and see if it works, that would be great!
In any case, to answer the original question, it should be possible to train a Hoeffding tree without a data::DatasetInfo. But, due to a bug in the code, that does not currently work. It will, though, once #2964 is merged. smile

For consistency with other classification algorithms, I thought there can be an optional parameter num_classes=0 of Train(...).

Nice point! Let me add that to the changes in #2964. 👍

[xlindo]

Hey @xlindo, thanks for reporting this! I played with the example and found that this is indeed a bug in the handling of new datasets. I opened a PR, #2964, which should fix the issue. If you'd like to try your code on that branch and see if it works, that would be great!
In any case, to answer the original question, it should be possible to train a Hoeffding tree without a data::DatasetInfo. But, due to a bug in the code, that does not currently work. It will, though, once #2964 is merged. smile

For consistency with other classification algorithms, I thought there can be an optional parameter num_classes=0 of Train(...).

Nice point! Let me add that to the changes in #2964. 👍

Many thanks.
This issue could be closed now.