Description

DecisionTree with mae criterion does not free used memory.
As a result that kind of trees is unusable in cycle, forests and cross-validation.
(originally i encoutered this problem while trying use grid search with mae forest)

Steps/Code to Reproduce

import sklearn

params = dict(max_depth=range(1, 10))
params = sklearn.model_selection.ParameterGrid(params)
reg = sklearn.tree.DecisionTreeRegressor(criterion='mae')

# repeated several times in jupyter cell
for p in params:
    print p
    for i in range(100):
        reg.set_params(**p)
        reg.fit(xxx, yyy)

Expected Results

No additional memory usage after first completion of cycle.
Cell with 'for' cycle is runnable unlimited times.

Replacing 'mae' with 'mse' does give expected result.

Actual Results

Additional memory is used after each cell execution.
Deleteing tree with 'del' and enforcing 'gc.collect()' does not help.
When there is no memory left OS (windows) shuts down python process.

My guess is that there is memory leak in cython.

Versions

Windows-7-6.1.7601-SP1
('Python', '2.7.12 | packaged by conda-forge | (default, Aug 12 2016, 18:53:38) [MSC v.1500 64 bit (AMD64)]')
('NumPy', '1.11.2')
('SciPy', '0.18.0')
('Scikit-Learn', '0.18')