AttributeError: 'LeastSquaresError' object has no attribute 'get_init_raw_predictions'

[mgb1028]

Description

Attribute error thrown when calling predict with GradientBoostedRegressor.

This is a model that has already been trained and that has been used before. I am trying to run some predictions on new data and this error arises. If I extract one of the estimators I can call predict.

Steps/Code to Reproduce

import pandas as pd
import numpy as np
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.externals import joblib

matrix = pd.read_feather(design_path)
matrix_oct_dic18 = matrix[(matrix.fechaAltaCliente >= '2018-10-01') & (matrix.fechaAltaCliente <= '2018-12-31')]
recommendation = matrix_oct_dic18[['pkprospecto'] + X_cols]
model = joblib.load(model_path)
y = model.predict(recommendation[X_cols])

Unfortunately I cannot share most of the code and data being used.

Here are the model specifications:
GradientBoostingRegressor(alpha=0.9, criterion='friedman_mse', init=None,
learning_rate=0.005, loss='ls', max_depth=3,
max_features=None, max_leaf_nodes=None,
min_impurity_decrease=0.0, min_impurity_split=None,
min_samples_leaf=1, min_samples_split=2,
min_weight_fraction_leaf=0.0, n_estimators=1000,
n_iter_no_change=None, presort='auto',
random_state=None, subsample=1.0, tol=0.0001,
validation_fraction=0.1, verbose=0, warm_start=False)

Expected Results

No error is thrown.

Actual Results

---

AttributeError Traceback (most recent call last)
in
----> 1 y = model.predict(recommendation[X_cols])

/opt/conda/lib/python3.7/site-packages/sklearn/ensemble/gradient_boosting.py in predict(self, X)
2530 X = check_array(X, dtype=DTYPE, order="C", accept_sparse='csr')
2531 # In regression we can directly return the raw value from the trees.
-> 2532 return self._raw_predict(X).ravel()
2533
2534 def staged_predict(self, X):

/opt/conda/lib/python3.7/site-packages/sklearn/ensemble/gradient_boosting.py in _raw_predict(self, X)
1669 def _raw_predict(self, X):
1670 """Return the sum of the trees raw predictions (+ init estimator)."""
-> 1671 raw_predictions = self.raw_predict_init(X)
1672 predict_stages(self.estimators, X, self.learning_rate,
1673 raw_predictions)

/opt/conda/lib/python3.7/site-packages/sklearn/ensemble/gradient_boosting.py in raw_predict_init(self, X)
1663 dtype=np.float64)
1664 else:
-> 1665 raw_predictions = self.loss.get_init_raw_predictions(
1666 X, self.init_).astype(np.float64)
1667 return raw_predictions

AttributeError: 'LeastSquaresError' object has no attribute 'get_init_raw_predictions'

Versions

System:
python: 3.7.3 | packaged by conda-forge | (default, Mar 27 2019, 23:01:00) [GCC 7.3.0]
executable: /opt/conda/bin/python
machine: Linux-4.15.0-51-generic-x86_64-with-debian-buster-sid

BLAS:
macros:
lib_dirs:
cblas_libs: cblas

Python deps:
pip: 19.1.1
setuptools: 41.0.1
sklearn: 0.21.2
numpy: 1.16.4
scipy: 1.2.1
Cython: None
pandas: 0.24.2

Thank you.

[mgb1028]

Downgrading the version of sklearn to 0.20.2 gets rid of this error.

[NicolasHug]

This happens because in #12983 we deprecated all the losses from gradient_boosting.py and created private losses in _gb_losses.py with a slightly different interface, that is now used in the GBDT code.

We broke pickle compatibility between version.

Is this something we shouldn't have done @jnothman @amueller ?

[thomasjpfan]

"While models saved using one version of scikit-learn might load in other versions, this is entirely unsupported and inadvisable. It should also be kept in mind that operations performed on such data could give different and unexpected results."

https://scikit-learn.org/stable/modules/model_persistence.html#security-maintainability-limitations

[NicolasHug]

Ah I knew we had that somewhere but couldn't find it, thanks.

So @mgb1028 I'm afraid the fix is to re-build the model with version 0.21

(While you're at it you might want to check the new GBDT implementation which is much faster!)

[amueller]

I think it's ok to break pickle compatibility here.

[amueller]

@mgb1028 you're getting a warning that the pickle might not work as expected, right?

[mgb1028]

Hi @amueller, sorry about the delayed response. You are correct. There is a warning that the pickle not might work as expected when loading the sklearn model with a different version.