def _bench_ml(sensor,
x_train,
y_train,
x_test,
*,
x_other=None,
verbose=False,
seed=42,
bagging=True,
gridsearch=False,
scale=True,
methods=None,
**kwargs):
from sklearn.preprocessing import RobustScaler, MinMaxScaler
from sklearn.model_selection import GridSearchCV
from sklearn.multioutput import MultiOutputRegressor
from sklearn.ensemble import BaggingRegressor
from .ML import models
args = getattr(kwargs, 'args', None)
seed = getattr(args, 'seed', seed)
gridsearch_kwargs = {
'refit': False,
'scoring': 'neg_median_absolute_error'
}
bagging_kwargs = {
'n_estimators': getattr(args, 'n_rounds', 10),
'max_samples': 0.75,
'bootstrap': False,
'random_state': seed,
}
if len(y_train.shape) == 1: y_train = y_train[:, None]
valid = np.isfinite(x_train).all(-1) & np.isfinite(y_train).all(-1)
x_train = x_train[valid]
y_train = y_train[valid]
if scale:
# x_scaler = TransformerPipeline([AUCTransformer(list(get_sensor_bands(sensor))), RobustScaler()])
x_scaler = TransformerPipeline([RobustScaler()])
y_scaler = TransformerPipeline(
[LogTransformer(), MinMaxScaler((-1, 1))])
x_scaler.fit(x_train)
y_scaler.fit(y_train)
x_test = x_scaler.transform(x_test)
x_train = x_scaler.transform(x_train)
y_train = y_scaler.transform(y_train)
preprocess = lambda m: m.fit(x_train.copy(), y_train.copy())
postprocess = None if not scale else y_scaler.inverse_transform
if verbose and gridsearch:
print('\nPerforming gridsearch...')
if methods is None:
methods = list(models.keys())
other = {}
estim = {}
for method, params in models.items():
if method not in methods: continue
methods.remove(method)
params['grid']['random_state'] = params['default'][
'random_state'] = seed
model_kwargs = params['default']
model_class = params['class']
n_jobs = 1 if method == 'MDN' else 3
if y_train.shape[1] > 1:
model_class = lambda *args, **kwargs: MultiOutputRegressor(params[
'class'](*args, **kwargs))
with GlobalRandomManager(seed):
if gridsearch and method != 'SVM':
model = GridSearchCV(model_class(),
params['grid'],
n_jobs=n_jobs,
**gridsearch_kwargs)
model.fit(x_train.copy(), y_train.copy())
model_kwargs = model.best_params_
if verbose: print(f'Best {method} params: {model_kwargs}')
model = model_class(**model_kwargs)
if bagging: model = BaggingRegressor(model, **bagging_kwargs)
model.__name__ = method
estim[method] = _create_estimates(model,
x_test,
postprocess,
preprocess,
verbose=verbose,
**kwargs)
if x_other is not None:
other[method] = _create_estimates(model, x_other, postprocess)
if len(methods):
print(f'Unknown ML benchmark methods requested: {methods}')
if len(other):
return estim, other
return estim
def _bench_ml(sensor,
x_train,
y_train,
x_test,
*,
x_other=None,
verbose=False,
seed=42,
bagging=True,
gridsearch=False,
scale=True,
**kwargs):
from sklearn.preprocessing import RobustScaler, MinMaxScaler
from sklearn.model_selection import GridSearchCV
from sklearn.ensemble import BaggingRegressor
from .Benchmarks.ML import models
seed = getattr(getattr(kwargs, 'args', None), 'seed', seed)
gridsearch_kwargs = {
'refit': False,
'scoring': 'neg_median_absolute_error'
}
bagging_kwargs = {
'n_estimators': 10,
'max_samples': 0.75,
'bootstrap': False,
'random_state': seed,
}
if scale:
x_scaler = TransformerPipeline([RobustScaler()])
y_scaler = TransformerPipeline(
[LogTransformer(), MinMaxScaler((-1, 1))])
x_scaler.fit(x_train)
y_scaler.fit(y_train)
x_test = x_scaler.transform(x_test)
x_train = x_scaler.transform(x_train)
y_train = y_scaler.transform(y_train).flatten()
preprocess = lambda m: m.fit(x_train.copy(), y_train.copy())
postprocess = None if not scale else y_scaler.inverse_transform
if verbose and gridsearch:
print('\nPerforming gridsearch...')
other = {}
estim = {}
for method, params in models.items():
params['grid']['random_state'] = params['default'][
'random_state'] = seed
model_kwargs = params['default']
with GlobalRandomManager(seed):
if gridsearch and method != 'SVM':
model = GridSearchCV(params['class'](),
params['grid'],
n_jobs=3 if method != 'MDN' else 1,
**gridsearch_kwargs)
model.fit(x_train.copy(), y_train.copy())
model_kwargs = model.best_params_
if verbose: print(f'Best {method} params: {model_kwargs}')
model = params['class'](**model_kwargs)
if bagging: model = BaggingRegressor(model, **bagging_kwargs)
model.__name__ = method
estim[method] = _create_estimates(model,
x_test,
postprocess,
preprocess,
verbose=verbose,
**kwargs)
if x_other is not None:
other[method] = _create_estimates(model, x_other, postprocess)
if len(other):
return estim, other
return estim
