python
/
nyaggle
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

add method option to averaging_opt

feature/enhance-averaging
nyanp 2022-03-15 20:02:06 +09:00
parent 36566e748f
commit c0fac30372
2 changed files with 37 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
nyaggle/ensemble/averaging.py
View File

@ -72,12 +72,13 @@ def averaging(test_predictions: List[np.ndarray],
def averaging_opt(test_predictions: List[np.ndarray],
oof_predictions: Optional[List[np.ndarray]],
y: Optional[pd.Series],
eval_func: Optional[Callable],
eval_func: Optional[Callable[[np.ndarray, np.ndarray], float]],
higher_is_better: bool,
weight_bounds: Tuple = (0, 1),
rank_averaging: bool = False) -> EnsembleResult:
weight_bounds: Tuple[float, float] = (0.0, 1.0),
rank_averaging: bool = False,
method: Optional[str] = None) -> EnsembleResult:
"""
Perform averaging with optimal weights using scipy.optimize
Perform averaging with optimal weights using scipy.optimize.
Args:
test_predictions:
@ -87,13 +88,16 @@ def averaging_opt(test_predictions: List[np.ndarray],
y:
Target value
eval_func:
Evaluation metric used for calculating result score. Used only if ``oof_predictions`` and ``y`` are given.
Evaluation metric f(y_true, y_pred) used for calculating result score.
Used only if ``oof_predictions`` and ``y`` are given.
higher_is_better:
Determine the direction of optimize ``eval_func``.
weight_bounds:
Specify lower/upper bounds of each weight.
rank_averaging:
If ``True``, predictions will be converted to rank before averaging.
method:
Type of solver. If ``None``, SLSQP will be used.
Returns:
Namedtuple with following members
@ -118,11 +122,16 @@ def averaging_opt(test_predictions: List[np.ndarray],
if rank_averaging:
test_predictions, oof_predictions = _to_rank(test_predictions, oof_predictions)
cons = ({'type': 'eq', 'fun': lambda w: 1 - sum(w)})
method = method or 'SLSQP'
if method in ['COBYLA', 'SLSQP', 'trust-constr']:
cons = ({'type': 'eq', 'fun': lambda w: 1 - sum(w)})
else:
cons = None
bounds = [weight_bounds] * len(test_predictions)
result = minimize(_minimize, weights, method='SLSQP', constraints=cons, bounds=bounds)
result = minimize(_minimize, weights, method=method, constraints=cons, bounds=bounds)
return averaging(test_predictions, oof_predictions, y, result['x'], eval_func)

21
tests/ensemble/test_averaging.py
View File

@ -175,6 +175,27 @@ def test_averaging_opt_minimize():
assert result.score <= result_simple_avg.score
def test_averaging_opt_minimize_with_method():
X, y = make_regression_df(n_samples=1024)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)
oof, test = _make_1st_stage_preds(X_train, y_train, X_test)
best_single_model = min(mean_squared_error(y_train, oof[0]),
mean_squared_error(y_train, oof[1]),
mean_squared_error(y_train, oof[2]))
result1 = averaging_opt(test, oof, y_train, mean_squared_error, higher_is_better=False)
result2 = averaging_opt(test, oof, y_train, mean_squared_error, higher_is_better=False, method='Nelder-Mead')
result3 = averaging_opt(test, oof, y_train, mean_squared_error, higher_is_better=False, method='SLSQP')
assert result1.score != result2.score
assert result1.score == result3.score
assert result1.score <= best_single_model
assert result2.score <= best_single_model
def test_rank_averaging_opt_maximize():
X, y = make_classification_df(n_samples=1024)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)