search

mljar / mljar-supervised

Python package for AutoML on Tabular Data with Feature Engineering, Hyper-Parameters Tuning, Explanations and Automatic Documentation
https://mljar.com
MIT License
3.05k stars 406 forks source link

Questions: Is there the possibility to only save and load the best model? #548

Open PeteKinz opened 2 years ago

PeteKinz commented 2 years ago

I am training quite a lot of models and this library does help me a lot! Unfortunately the model folder gets quite big after going through some interactions. Is there a way to only save and load the best performing model?

pplonski commented 2 years ago

Hi @PeteKinz! I'm glad that you found the library useful :) Do you get good models? Do you like models README?

It is possible but needs to be implemented. Currently, all my efforts are focused on the https://github.com/mljar/mercury framework for converting notebooks to web apps.

I do provide commercial support for MLJAR AutoML. I can implement this feature if needed by a commercial client. Licenses start at $1000/year.

PeteKinz commented 2 years ago

Hey @pplonski Thanks for your answer!

Some models are still training but my first attempts had very promising results :)

Ah, I can fully understand that you have to focus on an other project. I am using you library more or less for personal use, so a Licenses fee is not really an option for me but thanks anyways!

pplonski commented 2 years ago

Let's not close the issue. Sometimes I got interns and one of them might code it.

Karlheinzniebuhr commented 2 years ago

+1 for this, ran in Perform mode and the resulting model weights 1.7 GB, not really practical at those dimensions. @pplonski if you can guide me on how to implement the "save only best model" function I will implement it

HazelPinto commented 1 year ago

Excuse me, I've tried to save my trained model but I don't find any function to it Is this possible to save in 'perform' mode? I ran 'compete' mode and even though I tried to save it manually (best: stacked ensemble) I can't do it. Re-train every time I use the tool is not an option for me. If there's an option to save it in pickle it would be amazing to know =)

pplonski commented 1 year ago

Hi @HazelPinto,

All models are automatically saved to results_path. To load them just use the same directory name. Here you have more details https://github.com/mljar/mljar-supervised#faq