def _search_best(self, **Xy):
# Fit/predict CV grid search
self.cv.store = StoreMem() # local store erased at each fit
from epac.workflow.pipeline import Pipe
self.cv.top_down(**Xy)
# Pump-up results
cv_result_set = self.cv.reduce(store_results=False)
key_val = [(result.key(), result[self.score]) \
for result in cv_result_set]
scores = np.asarray(zip(*key_val)[1])
scores_opt = np.max(scores) if self.arg_max else np.min(scores)
idx_best = np.where(scores == scores_opt)[0][0]
best_key = key_val[idx_best][0]
# Find nodes that match the best
nodes_dict = {n.get_signature(): n for n in self.cv.walk_true_nodes() \
if n.get_signature() in key_split(best_key)}
to_refit = Pipe(*[nodes_dict[k].estimator for k in key_split(best_key)])
best_params = [dict(sig) for sig in key_split(best_key, eval=True)]
return to_refit, best_params
def reduce(self, result):
from epac.workflow.pipeline import Pipe
# Pump-up results
cv_result_set = result
key_val = [(result.key(), result[self.NodeBestSearchRefit.score])
for result in cv_result_set]
scores = np.asarray(zip(*key_val)[1])
scores_opt = np.max(scores)\
if self.NodeBestSearchRefit.arg_max else np.min(scores)
idx_best = np.where(scores == scores_opt)[0][0]
best_key = key_val[idx_best][0]
# Find nodes that match the best
nodes_dict = \
{n.get_signature():
n for n in self.NodeBestSearchRefit.children[0].walk_true_nodes()
if n.get_signature() in key_split(best_key)}
to_refit = Pipe(*[nodes_dict[k].wrapped_node
for k in key_split(best_key)])
best_params = [dict(sig) for sig in key_split(best_key, eval=True)]
return to_refit, best_params
def _search_best(self, **Xy):
# Fit/predict CV grid search
self.cv.store = StoreMem() # local store erased at each fit
from epac.workflow.pipeline import Pipe
self.cv.top_down(**Xy)
# Pump-up results
cv_result_set = self.cv.reduce(store_results=False)
key_val = [(result.key(), result[self.score])
for result in cv_result_set]
scores = np.asarray(zip(*key_val)[1])
scores_opt = np.max(scores) if self.arg_max else np.min(scores)
idx_best = np.where(scores == scores_opt)[0][0]
best_key = key_val[idx_best][0]
# Find nodes that match the best
nodes_dict = {n.get_signature(): n for n in self.cv.walk_true_nodes()
if n.get_signature() in key_split(best_key)}
to_refit = Pipe(*[nodes_dict[k].wrapped_node
for k in key_split(best_key)])
best_params = [dict(sig) for sig in key_split(best_key, eval=True)]
return to_refit, best_params