def optimize(self,
train,
rare_thresh=100,
size=5000,
tune_mode="paramwise",
as_text=False,
cached_params=False):
# Estimate useful features on a random sample of |size| instances
sys.stderr.write("o Tuning hyperparameters\n\n")
# Optimize hyperparameters via grid search
if cached_params:
clf, best_params, _ = get_best_params(self.corpus, self.name)
sys.stderr.write("\no Using cached best hyperparameters\n")
else:
clf, best_params = self.train(train,
rare_thresh=rare_thresh,
tune_mode=tune_mode,
size=size,
as_text=as_text)
sys.stderr.write("\no Found best hyperparameters\n")
for key, val in best_params.items():
sys.stderr.write(key + "\t" + str(val) + "\n")
sys.stderr.write("\n")
return clf, [], best_params
def optimize(self,
train,
rare_thresh=100,
size=5000,
tune_mode="paramwise",
cached_params=False,
as_text=False):
# Estimate useful features on a random sample of |size| instances
selected_cat, selected_num = self.train(train,
model_path=None,
rare_thresh=100,
as_text=False,
size=size,
tune_mode="importances")
selected_feats = selected_cat + selected_num
sys.stderr.write("o Chose " + str(len(selected_feats)) +
" features: " + ",".join(selected_feats) + "\n")
sys.stderr.write("o Tuning hyperparameters\n\n")
# Optimize hyperparameters via grid search or hyperopt
if cached_params:
best_clf, best_params, _ = get_best_params(self.corpus, self.name)
sys.stderr.write("\no Using cached best hyperparameters\n")
else:
best_clf, best_params = self.train(train,
rare_thresh=rare_thresh,
tune_mode=tune_mode,
size=size,
as_text=as_text)
sys.stderr.write("\no Found best hyperparameters\n")
for key, val in best_params.items():
sys.stderr.write(key + "\t" + str(val) + "\n")
sys.stderr.write("\n")
return best_clf, selected_feats, best_params
chosen_feats=vars,
rare_thresh=200,
clf_params=best_params,
as_text=False,
multitrain=True)
else:
est.train(train,
rare_thresh=200,
clf_params=best_params,
as_text=False,
chosen_clf=clf)
elif "train" in mode:
if opts.best_params and est.name in [
"SubtreeSegmenter", "EnsembleSegmenter"
]:
best_clf, params, feats = get_best_params(corpus, est.name)
else:
best_clf = None
params = None
feats = None
if est.name == "SubtreeSegmenter":
est.train(train,
rare_thresh=200,
as_text=False,
multitrain=True,
chosen_clf=best_clf,
clf_params=params,
chosen_feats=feats)
elif est.name == "RNNSegmenter":
est.train(train, as_text=False, multifolds=5)
elif est.name == "EnsembleSegmenter":
# Now train on whole training set with those variables
sys.stderr.write("\no Training best configuration\n")
e.train(train,
rare_thresh=200,
clf_params=best_params,
as_text=False,
chosen_clf=best_clf,
chosen_feats=vars,
size=220000)
elif "train" in opts.mode:
tune_mode = None if opts.tune_mode != "hyperopt" else "hyperopt"
feats = None
params = None
best_clf = None
if opts.best_params:
best_clf, params, feats = get_best_params(
corpus, "EnsembleSentencer")
if len(feats) == 0:
feats = None
e.train(train,
chosen_feats=feats,
as_text=False,
tune_mode=tune_mode,
clf_params=params,
chosen_clf=best_clf,
size=220000)
if "test" in opts.mode:
if opts.eval_test:
conf_mat, prec, rec, f1 = e.predict(test,
eval_gold=True,
as_text=False)
else:
size=5000,
tune_mode=opts.tune_mode,
cached_params=opts.best_params)
if "best_score" in best_params:
best_params.pop("best_score")
# Now train on whole training set with those variables
sys.stderr.write("\no Training best configuration\n")
e.train(train,
rare_thresh=100,
clf_params=best_params,
as_text=False,
chosen_clf=clf)
elif "train" in opts.mode:
if opts.best_params:
best_clf, best_params, _ = get_best_params(corpus,
e.name,
auto=auto)
else:
best_clf = None
sys.stderr.write("\no Training on corpus " + corpus + "\n")
tune_mode = None if opts.tune_mode != "hyperopt" else "hyperopt"
e.train(train,
as_text=False,
tune_mode=tune_mode,
chosen_clf=best_clf,
clf_params=best_params)
if "test" in opts.mode:
if opts.eval_test:
conf_mat, prec, rec, f1 = e.predict(test,
eval_gold=True,