Need to a workable example about mleap HOT 9 CLOSED

Hello,

Glad to hear about your interest in MLeap. We are currently making a big push to get to version 0.3.0, which will get us much closer to a stable API to work with. We have 3 pull requests to get through and then we will release it to maven central. Over the next week, we will be building out documentation and examples about how to use MLeap.

Thank you for your patience, we'll check back in on this ticket as we build out the documentation.

from mleap.

@lei-clari ,

We had to write our own OneHotEncoder because Spark’s uses metadata to perform the transformation. This is unusual for Spark models, so instead, we use a custom estimator/model which stores the size of the output vectors so we don’t need to rely on metadata to do one hot encoding.

Re bundle.ml: You can choose whether you want to serialize to json or to protobuf - your choice should depend on which transformers you’re using and which models you’re training (i.e. for tree-based models we recommend protofub). I’ll add documentation and an example of serialization to protobuf.

As far as TC, they were using it when we left, but have not stayed up to date on the latest usage there. If you’d like to talk with someone who’s running it in production we can make an intro.

Best
Mikhail

from mleap.

@ltang We have added in pretty complete documentation here:
http://mleap-docs.combust.ml/

Please let us know in a separate issue if you would like us to add more specific documentation.

Cheers,
Hollin

from mleap.

Hi @ltang ,

We checked in a few examples with both scala and pyspark - can you give them a try?
https://github.com/combust-ml/mleap-demo/tree/master/notebooks

We're working on getting a public model server up in the next few weeks so you can test .deploy() as well.

Best
M

from mleap.

Thanks. I'll have a try and get back to you in few days.

from mleap.

@seme0021 Thanks for the demos.

I just walked through the airbnb example (except the deploy part). works out of the box!

I could not get toree to work due to some dependency conflict. May have to dig deeper. But the example works fine as long as I import all mleap dependencies.

Several minor questions:

1. why is the demo using MLeap's OneHotEncoder rather than Spark's default one?
2. Is bundle.ml format simply a bundle of json files (node.json and model.json)?
3. Is TrueCar already using MLeap?

Our own use case requires some additional transformers. Will check your source code to see whether we can add customized transformers.

from mleap.

Hey, just giving an update here. We are building out comprehensive documentation for MLeap 0.5.0. You can see it here as it develops:

http://mleap-docs.combust.ml/

from mleap.

from mleap.

@geoHeil I responded to this question in another thread, but I'll post here as well.

Here is a list of supported transformers: http://mleap-docs.combust.ml/core-concepts/transformers/support.html

We are coming out with Sklearn modules on PyPi with the next release of MLeap, 0.6.0, which should be a fairly quick release after 0.5.0.

from mleap.