CitrineInformatics/lolo

How to save the fitted RandomForestRegressor model?

Opened this issue · 2 comments

How can I save a lolopy model?"

I tried to train a model like this:

from lolopy.learners import RandomForestRegressor
model = RandomForestRegressor()
model.fit(X, Y)

After that, I attempted to save the model using:

joblib.dump(model, "./model.pkl")

But, it didn't work with the following error:

AttributeError: 'RandomForestRegressor' object has no attribute 'gateway'

Thank you for sharing your great program.

I have the same issue. Here are some additional details that might be helpful for debugging.

Version

3.0.0 (current pip version)

Issue

Lolopy fails to pickle itself when trying to save a trained model with joblib

Minimal code to reproduce

from lolopy.learners import RandomForestRegressor as LoloRandomForestRegressor
import joblib
import numpy as np

X = np.random.rand(20,5)
y = np.random.rand(20,1)

model = LoloRandomForestRegressor()
model.fit(X, y)
joblib.dump(model, 'model.joblib')

Error Message

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[5], line 10
      8 model = LoloRandomForestRegressor()
      9 model.fit(X, y)
---> 10 joblib.dump(model, 'model.joblib')

File ~/Applications/miniforge3/envs/pmg/lib/python3.12/site-packages/joblib/numpy_pickle.py:553, in dump(value, filename, compress, protocol, cache_size)
    551 elif is_filename:
    552     with open(filename, 'wb') as f:
--> 553         NumpyPickler(f, protocol=protocol).dump(value)
    554 else:
    555     NumpyPickler(filename, protocol=protocol).dump(value)

File ~/Applications/miniforge3/envs/pmg/lib/python3.12/pickle.py:481, in _Pickler.dump(self, obj)
    479 if self.proto >= 4:
    480     self.framer.start_framing()
--> 481 self.save(obj)
    482 self.write(STOP)
    483 self.framer.end_framing()

File ~/Applications/miniforge3/envs/pmg/lib/python3.12/site-packages/joblib/numpy_pickle.py:355, in NumpyPickler.save(self, obj)
    352     wrapper.write_array(obj, self)
    353     return
--> 355 return Pickler.save(self, obj)

File ~/Applications/miniforge3/envs/pmg/lib/python3.12/pickle.py:572, in _Pickler.save(self, obj, save_persistent_id)
    570 reduce = getattr(obj, "__reduce_ex__", None)
    571 if reduce is not None:
--> 572     rv = reduce(self.proto)
    573 else:
    574     reduce = getattr(obj, "__reduce__", None)

File ~/Applications/miniforge3/envs/pmg/lib/python3.12/site-packages/lolopy/learners.py:59, in BaseLoloLearner.__getstate__(self)
     57 # If there is a model set, replace it with the JVM copy
     58 if self.model_ is not None:
---> 59     state['model_'] = self.gateway.jvm.io.citrine.lolo.util.LoloPyDataLoader.serializeObject(self.model_,
     60                                                                                              self._compress_level)
     61 return state

AttributeError: 'RandomForestRegressor' object has no attribute 'gateway'

Solved by #318. Use the new save, load methods instead of joblib dump for now.

Updated Example Code -- working

from lolopy.learners import RandomForestRegressor as LoloRandomForestRegressor
import numpy as np

X = np.random.rand(20,5)
y = np.random.rand(20,1)

model = LoloRandomForestRegressor()
model.fit(X, y)
print(model.predict(X))

### Save
model.save('model.lolopy.rfr')

### Load
model = LoloRandomForestRegressor.load('model.lolopy.rfr')
print(model.predict(X))