def weka_arff_ttest(model='nc'):
# importing config
print("\nimporting config file...")
config = utility.load_config()
print("\npreparing the components...\n")
db = DBConnector(**config['db'])
if model is 'nc':
n_class = 9
dataset = shuffle(pd.DataFrame(db.find_trainingset()), random_state=42)
# dataset = dataset.head(100)
print("\ntokenizing...\n")
ds = preprocessing.tokenize_list(dataset)
labels = dataset['tag'].to_numpy()
txt_labels = classification.news_categories
else:
n_class = 2
dataset = shuffle(pd.DataFrame(db.find_likabilityset()),
random_state=42)
print("\ntokenizing...\n")
ds = pd.DataFrame()
ds['content'] = preprocessing.tokenize_list(dataset)
ds['tag'] = dataset['tag']
labels = np.asarray([
classification.labelize_likability(a)[0]
for _, a in dataset.iterrows()
])
txt_labels = ['LIKE', 'DISLIKE']
# preparing classifiers
print("\n\nt-testing %s ...\n\n" % (model))
classifiers = classification.init_simple_classifiers(model)
classifiers = classifiers + classification.init_ensmeta_classifiers(
classifiers, model) # putting together simple and ens/meta
classification.weka_ttest(classifiers,
ds,
labels,
n_class=n_class,
txt_labels=txt_labels,
model=model)
def build_feed(self, parsed_feed):
self.feed = pd.DataFrame(parsed_feed)
features = pd.DataFrame()
features['content'] = pp.tokenize_list(parsed_feed)
features['tag'] = self.nws_clf.predict(features['content'])
self.feed['likability'] = self.lik_prd.predict(features)
self.feed['predicted_tag'] = features['tag']
self.feed = self.feed[self.feed['likability'] == 'LIKE'].drop(
'likability', axis=1)
def deploy_news_classifier(dataset, dir_path='home/miner/model'):
# this function should provide a trained pipeline (so apparently bagging is the best now)
clf = BaggingClassifier(base_estimator=classifier['svc'](C=0.51, random_state=42),
n_estimators=100,
random_state=42)
model = build_nc_model(('clf', clf))
dataset = shuffle(dataset, random_state=42)
ds = pp.tokenize_list(dataset)
labels = dataset['tag'].to_numpy()
cross_validate_fullscores(model, ds, labels, n_class=9, txt_labels=news_categories)
joblib.dump(model, dir_path + '/nws_clf.pkl')
return model
def meta_classify_lc(dataset, show_mat=False, tuning=False, plot=False, load_pretokenized=False):
# preparing the inputs
ds = pd.DataFrame()
dataset = shuffle(dataset, random_state=42)
# avoid wasting time during test...
# if load_pretokenized:
# with open('home/pretokenized_dataset/lcds.pkl', 'rb') as f:
# ds['content'] = pickle.load(f)
# else:
ds['content'] = pp.tokenize_list(dataset)
ds['tag'] = dataset['tag']
# preparing the targets
labels = np.asarray([labelize_likability(a)[0] for _,a in dataset.iterrows()])
# classifiers initialization
classifiers = init_simple_classifiers('lc')
print("\n\nSimple Classifiers:\n")
for c in classifiers:
print('\n---------------------------\n')
# building the model
pl = build_lc_model(c)
print('\nCrossValidation with 10 folds for ' + c[0] + '\n')
cross_validate(pl, ds, labels, 2, show_mat=show_mat, txt_labels=['LIKE', 'DISLIKE'], random_state=42)
if plot:
plot_learning_curve(pl, c[0], ds, labels, random_state=42)
print('\n---------------------------\n')
print("\n\nEnsembles and Meta-Classifiers:\n")
np.random.seed(42)
ens_meta_classifiers = init_ensmeta_classifiers(classifiers, 'lc')
for c in ens_meta_classifiers:
print("CV 10 folds - " + c[0])
# building the pipeline vectorizer-classifier
pl = build_lc_model(c)
# Cross_validating the model (dunno y its not working with the )
cross_validate(pl, ds, labels, 2, show_mat=show_mat, txt_labels=['LIKE', 'DISLIKE'], random_state=42)
if plot:
plot_learning_curve(pl, c[0], ds, labels, random_state=42)
def deploy_likability_predictor(dataset, dir_path='home/miner/model'):
# preparing the inputs
ds = pd.DataFrame()
dataset = shuffle(dataset, random_state=42)
ds['content'] = pp.tokenize_list(dataset)
ds['tag'] = dataset['tag']
# preparing the targets
labels = np.asarray([labelize_likability(a)[0] for _,a in dataset.iterrows()])
# building the model...
clf = classifier['log_reg'](solver='lbfgs', multi_class='auto', random_state=42)
model = build_lc_model(('clf', clf))
cross_validate_fullscores(model, ds, labels, n_class=2, txt_labels=['LIKE', 'DISIKE'])
joblib.dump(model, dir_path + '/lik_prd.pkl')
return model
def meta_classify_nc(dataset, show_mat=False, tuning=False, plot=False, load_pretokenized=False):
# preparing the trainingset
dataset = shuffle(dataset, random_state=42)
# avoid wasting time during test...
# if load_pretokenized:
# with open('home/pretokenized_dataset/ncds.pkl', 'rb') as f:
# ds = pickle.load(f)
# else:
ds = pp.tokenize_list(dataset) # pp.vectorize_list(dataset) (doc_to_vector stuff, not really working)
# preparing the targets
labels = dataset['tag'].to_numpy()
n_class = len(news_categories)
# classifiers initialization
classifiers = init_simple_classifiers('nc')
# params tuning
params = {
'rf' : {
'clf__n_estimators': [2000],
'clf__bootstrap': [True],
'clf__max_depth': [None],
'clf__max_features': ['sqrt'],
'clf__min_samples_leaf': [2, 4],
'clf__min_samples_split': [5, 10]
},
'mnb' : {},
'svc' : {
# 'vect__ngram_range': [(1, 1), (1, 2), (1, 3), (2, 2)],
# 'vect__min_df': (1, 0.01, 0.025),
# 'vect__norm': [None, 'l1', 'l2']
# 'vect__max_df': (0.5, 0.55, 0.65, 0.70, 0.75, 1.0),
# 'vect__max_features' : [12000, 14500, 15000], # best MF: 14500
'clf__C': np.arange(0.001, 1, 0.001) # best C: 0.51
},
'ada' : {
'clf__n_estimators': [2000],
'clf__algorithm': ['SAMME']
},
'lr' : {},
'dt' : {}
}
print("\n\nSimple Classifiers:\n")
for c in classifiers:
print('\n---------------------------\n')
# building the model
model = build_nc_model(c)
if tuning:
print("\nTuning {} Hyper-Parameters with GridSearchCV: \n".format(c[0]))
tuned_model = GridSearchCV(model, params[c[0]], iid=True,
scoring='accuracy', cv=StratifiedKFold(random_state=42, shuffle=True, n_splits=10),
verbose=1,
n_jobs=-1
)
tuned_model.fit(ds, labels)
print(tuned_model.best_score_, tuned_model.best_params_)
else:
print('\n CrossValidation with 10 folds for ' + c[0] + '\n')
cross_validate(model, ds, labels, n_class, show_mat=show_mat, txt_labels=news_categories, random_state=42)
if plot:
plot_learning_curve(model, c[0], ds, labels, random_state=42)
print('\n---------------------------\n')
print("\n\nSimple Classifiers WITH Independent Features Selection (chi2-40Percentile):\n")
for c in classifiers:
print('\n---------------------------\n')
# building the model
model = build_nc_model(c, feature_selection=SelectPercentile(chi2, percentile=40))
if tuning:
print("\nTuning {} Hyper-Parameters with GridSearchCV: \n".format(c[0]))
tuned_model = GridSearchCV(model, params[c[0]], iid=True,
scoring='accuracy', cv=StratifiedKFold(random_state=42, shuffle=True, n_splits=10),
verbose=1,
n_jobs=-1
)
tuned_model.fit(ds, labels)
print(tuned_model.best_score_, tuned_model.best_params_)
else:
print('\n CrossValidation with 10 folds for ' + c[0] + '\n')
cross_validate(model, ds, labels, n_class, show_mat=show_mat, txt_labels=news_categories, random_state=42)
if plot:
plot_learning_curve(model, c[0], ds, labels, random_state=42)
print('\n---------------------------\n')
print("\n\nEnsembles and Meta-Classifiers:\n")
# set the seed for some classifiers...
np.random.seed(42)
# init the list...
ens_meta_classifiers = init_ensmeta_classifiers(classifiers, 'nc')
for c in ens_meta_classifiers:
print("\nCV 10 folds - " + c[0])
# building the pipeline vectorizer-classifier
model = build_nc_model(c)
# # Cross_validating the model
cross_validate(model, ds, labels, n_class, show_mat=show_mat, txt_labels=news_categories, random_state=42)
if plot:
plot_learning_curve(model, c[0], ds, labels, random_state=42)