def run_MFCC(self):
"""run classifier on MFCC feature (single modality)
"""
print("\nbuilding a classifier on MFCC features (both frame-level and session-level)")
X_train, y_train, train_inst, X_dev, y_dev, dev_inst = load_proc_baseline_feature('MFCC', verbose=True)
if self.model_name == 'RF_cv':
y_train, y_dev = np.ravel(y_train), np.ravel(y_dev)
train_inst, dev_inst = np.ravel(train_inst), np.ravel(dev_inst)
X = np.vstack((X_train, X_dev))
y = np.hstack((y_train, y_dev))
inst = np.hstack((train_inst, dev_inst))
assert len(X) == len(y) == len(inst)
cv_ids = k_fold_cv(len(X))
cv_res = []
for (ids_train, ids_dev) in cv_ids:
X_train = X[ids_train]
y_train = y[ids_train]
X_dev = X[ids_dev]
y_dev = y[ids_dev]
dev_inst = inst[ids_dev]
RF_MFCC = RandomForest(self.feature_name, X_train, y_train, X_dev, y_dev, baseline=True, test=self.test)
RF_MFCC.run()
y_pred_train, y_pred_dev = RF_MFCC.evaluate()
_, session_res = get_UAR(y_pred_dev, y_dev, dev_inst, self.model_name, self.feature_name, 'baseline', baseline=True, test=True)
cv_res.append(session_res)
save_cv_results(cv_res, self.model_name, self.feature_name, 'baseline')
print("\nupsampling training data to address class imbalance")
X_train, y_train, train_inst = upsample(X_train, y_train, train_inst)
print("\nobtaining sparse matrix for better classification")
# X_train = sp.csr_matrix(np.vstack((X_train, X_dev)))
# X_dev = sp.csr_matrix(X_dev)
# y_train = np.hstack((y_train, y_dev))
X_train, X_dev = sp.csr_matrix(X_train), sp.csr_matrix(X_dev)
if self.model_name == 'SVM':
SVM_MFCC = LinearSVM(self.feature_name, X_train, y_train, X_dev, y_dev, baseline=True, test=self.test)
SVM_MFCC.run()
y_pred_train, y_pred_dev = SVM_MFCC.evaluate()
elif self.model_name == 'RF':
RF_MFCC = RandomForest(self.feature_name, X_train, y_train, X_dev, y_dev, baseline=True, test=self.test)
RF_MFCC.run()
y_pred_train, y_pred_dev = RF_MFCC.evaluate()
get_UAR(y_pred_train, y_train, train_inst, self.model_name, self.feature_name, 'baseline', baseline=True, train_set=True, test=self.test)
get_UAR(y_pred_dev, y_dev, dev_inst, self.model_name, self.feature_name, 'baseline', baseline=True, test=self.test)
if not self.test:
get_post_probability(y_pred_dev, y_dev, dev_inst, np.array([]), self.model_name, self.feature_name)
def run_AU(self):
"""run classifier on AU feature (single modality)
"""
print("\nbuilding a classifier on AU features (already session-level)")
X_train, y_train, _, X_dev, y_dev, _ = load_proc_baseline_feature('AU', verbose=True)
if self.model_name == 'RF_cv':
X = np.vstack((X_train, X_dev))
y = np.hstack((y_train, y_dev))
assert len(X) == len(y)
cv_ids = k_fold_cv(len(X))
cv_res = []
for (ids_train, ids_dev) in cv_ids:
X_train = X[ids_train]
y_train = y[ids_train]
X_dev = X[ids_dev]
y_dev = y[ids_dev]
RF_MFCC = RandomForest(self.feature_name, X_train, y_train, X_dev, y_dev, baseline=True, test=self.test)
RF_MFCC.run()
y_pred_train, y_pred_dev = RF_MFCC.evaluate()
_, session_res = get_UAR(y_pred_dev, y_dev, np.array([]), self.model_name, self.feature_name, 'baseline', baseline=True, test=True)
cv_res.append(session_res)
save_cv_results(cv_res, self.model_name, self.feature_name, 'baseline')
print("\nupsampling training data to address class imbalance")
X_train, y_train, _ = upsample(X_train, y_train, np.array([]))
print("\nobtaining sparse matrix for better classification")
# X_train = sp.csr_matrix(np.vstack((X_train, X_dev)))
# X_dev = sp.csr_matrix(X_dev)
# y_train = np.hstack((y_train, y_dev))
X_train, X_dev = sp.csr_matrix(X_train), sp.csr_matrix(X_dev)
if self.model_name == 'SVM':
SVM_AU = LinearSVM(self.feature_name, X_train, y_train, X_dev, y_dev, baseline=True, test=self.test)
SVM_AU.run()
y_pred_train, y_pred_dev = SVM_AU.evaluate()
session_prob = SVM_AU.get_session_probability()
elif self.model_name == 'RF':
RF_AU = RandomForest(self.feature_name, X_train, y_train, X_dev, y_dev, baseline=True, test=self.test)
RF_AU.run()
y_pred_train, y_pred_dev = RF_AU.evaluate()
session_prob = RF_AU.get_session_probability()
get_UAR(y_pred_train, y_train, np.array([]), self.model_name, self.feature_name, 'baseline', baseline=True, train_set=True, test=self.test)
get_UAR(y_pred_dev, y_dev, np.array([]), self.model_name, self.feature_name, 'baseline', baseline=True, test=self.test)
def test_k_fold_cv(self):
ids = k_fold_cv(20)
for (ids_train, ids_dev) in ids:
print(ids_train, ids_dev)
def RF_CV(self):
print(
"\nrunning RF on features selected with RF with doc2vec embeddings"
)
feature_path = smart_open('./pre-trained/fusion/feature_list.txt',
'rb',
encoding='utf-8')
feature_list = []
for _, line in enumerate(feature_path):
feature_list.append(str(line).replace('\n', ''))
from sklearn.metrics import precision_recall_fscore_support
cv_results_UAR = dict()
cv_results_UAP = dict()
for feature in feature_list:
cv_results_UAR[feature] = []
cv_results_UAP[feature] = []
_, _, y_dev, y_train = load_label()
y_train = y_train.astype('int')
y_dev = y_dev.astype('int')
X_train = np.load(
os.path.join('pre-trained', 'fusion', feature, 'X_train.npy'))
X_dev = np.load(
os.path.join('pre-trained', 'fusion', feature, 'X_dev.npy'))
X = np.vstack((X_train, X_dev))
y = np.hstack((y_train, y_dev))
cv_ids = k_fold_cv(len(X))
for cv_id in cv_ids:
X_train = X[cv_id[0]]
y_train = y[cv_id[0]]
X_dev = X[cv_id[1]]
y_dev = y[cv_id[1]]
print('train on %d test on %d' % (len(y_train), len(y_dev)))
random_forest = RandomForest(feature,
X_train,
y_train,
X_dev,
y_dev,
test=False)
random_forest.run()
_, y_pred = random_forest.evaluate()
precision, recall, _, _ = precision_recall_fscore_support(
y_dev, y_pred, average='macro')
cv_results_UAR[feature].append(recall)
cv_results_UAP[feature].append(precision)
assert len(cv_results_UAR[feature]) == len(
cv_results_UAP[feature]) == 10
with open(os.path.join('results', 'cross-validation.json'),
'a+',
encoding='utf-8') as outfile:
json.dump(cv_results_UAR, outfile)
json.dump(cv_results_UAP, outfile)