partial_fit for incremental training of HT or EFDT in river

[Piyushbalwani]

Hi Team,

As it is mentioned in https://riverml.xyz/latest/user-guide/mini-batching/ , river package learn with one sample at a time. But in skmultiflow, there was implementation of partial_fit, which can fit multiple instances. In river , I found only learn_one for HT andEFDT.
Is there any equivalent function or alternative way to fit multiple instances?

https://scikit-multiflow.readthedocs.io/en/stable/api/generated/skmultiflow.trees.HoeffdingTreeClassifier.html

Regards,
Piyush

[smastelini]

Hi @Piyushbalwani. The partial_fit method in skmultiflow simply did a loop over each input example and called a private method that trained in an instance-by-instance regimen. So there were no real mini-batching/vectorization speed-ups.

It is in our plans to support real mini-batching capabilities in the trees in the future. But for now, the way to go is to iterate over instances and call learn_one for each instance.

[Piyushbalwani]

Thanks @smastelini. I was unaware of skmultiflow partial_fit instance-by-instance implementation.
So, can we train the model with skmultiflow with initial large dataframe and save the model. Then load the model and do instance-by-instance incremental training?
Or is there any other workaround as dataframe is quite big?

Regards,
Piyush

[smastelini]

The best alternative will depend on your usage. If your running everything locally you have a few options:

• stream.iter_csv, stream.iter_arff, or other utilities from the stream submodule if your data is saved in disk. Note that these iterators do not load the whole data into memory.
• stream.iter_pandas if your data frame is already loaded in memory. You can also use iter_array when using numpy objects :)

In all the cases, the tree models will learn instance by instance. You can perform some sort of warm-start with a subset of the examples, for sure. Nonetheless, learn_one is currently the only way to update the trees.

We hope to leverage vectorized operations in the future and provide learn_many/predict_many methods.