Saving uboost BDT with tf/keras base estimators about hep_ml

[arogozhnikov]

Poor picklability of keras is a long-known issue (you can google keras with the same mistake error).

You may be fortunate to have some of variables being passed e.g. in lambda not through calls,
https://stackoverflow.com/questions/55280201/keras-typeerror-cant-pickle-thread-rlock-objects

But otherwise I'm not sure there will be a simple solution

[srishtibhasin]

Yes, but of course keras has its own save_model functionality. So do you think it is simply not possible to save a uBoost model which is based on a keras model?

[arogozhnikov]

it is simply not possible to save a uBoost model which is based on a keras model?

It all goes down do pickle-ability of items.
hep_ml's uBoost is completely pickle-able, but keras model is not.

Option 1. Ask keras maintainers why your model is not pickle-able
Option 2. Store keras models separately from uBoost.

estimators = clf.estimators_ # list of keras models
# TODO save estimators somehow using keras tools
# delete estimators 
clf.estimators_ = None
with open('uboost.pkl', 'wb') as f:
   joblib.dump(clf, f)

# loading
with open('uboost.pkl', 'rb') as f:
   clf = joblib.load(f)

estimators = .... # load  estimators
clf.estimators_ = estimators

[arogozhnikov]

there is option 3 as well - find truly sklearn-compatible NN package =)

[srishtibhasin]

Ok I will try to implement option 2, thanks!