def record_res(cls, filename, taskname, f, n):
pos_feature, neg_feature = Ngram.build_features(flag=f, number=n)
shuffle(pos_feature)
shuffle(neg_feature)
index = int(((len(pos_feature) + len(neg_feature)) / 2) * 0.2)
x_train = pos_feature[index:] + neg_feature[index:]
x_test = pos_feature[:index] + neg_feature[:index]
BNB_A, BNB_PP, BNB_PR, BNB_PF1, BNB_NP, BNB_NR, BNB_NF1 = Ngram.score(
BernoulliNB(), x_train, x_test)
MNB_A, MNB_PP, MNB_PR, MNB_PF1, MNB_NP, MNB_NR, MNB_NF1 = Ngram.score(
MultinomialNB(), x_train, x_test)
LR_A, LR_PP, LR_PR, LR_PF1, LR_NP, LR_NR, LR_NF1 = Ngram.score(
LogisticRegression(), x_train, x_test)
LSVC_A, LSVC_PP, LSVC_PR, LSVC_PF1, LSVC_NP, LSVC_NR, LSVC_NF1 = Ngram.score(
LinearSVC(), x_train, x_test)
NSVC_A, NSVC_PP, NSVC_PR, NSVC_PF1, NSVC_NP, NSVC_NR, NSVC_NF1 = Ngram.score(
NuSVC(), x_train, x_test)
SVC_A, SVC_PP, SVC_PR, SVC_PF1, SVC_NP, SVC_NR, SVC_NF1 = Ngram.score(
SVC(), x_train, x_test)
# 记录 准确率（Accuracy）、正面精确率（pos_precision），正面召回率（pos_recall）、正面F1值、负面面精确率（neg_precision）、负面召回率（neg_recall）,负面F1值
log.console_out(
filename, taskname, n, ('BNB', BNB_A, 'pos:', BNB_PP, BNB_PR,
BNB_PF1, 'neg:', BNB_NP, BNB_NR, BNB_NF1),
('MNB', MNB_A, 'pos:', MNB_PP, MNB_PR, MNB_PF1, 'neg:', MNB_NP,
MNB_NR, MNB_NF1), ('LR', LR_A, 'pos:', LR_PP, LR_PR, LR_PF1,
'neg:', LR_NP, LR_NR, LR_NF1),
('SVC-rbf', SVC_A, 'pos:', SVC_PP, SVC_PR, SVC_PF1, 'neg:', SVC_NP,
SVC_NR, SVC_NF1), ('LSVC', LSVC_A, 'pos:', LSVC_PP, LSVC_PR,
LSVC_PF1, 'neg:', LSVC_NP, LSVC_NR, LSVC_NF1),
('NuSVC', NSVC_A, 'pos:', NSVC_PP, NSVC_PR, NSVC_PF1, 'neg:',
NSVC_NP, NSVC_NR, NSVC_NF1))
def record_res(cls, filename, taskname, f, n):
pos_feature, neg_feature = Ngram.build_features(flag=f, number=n)
shuffle(pos_feature)
shuffle(neg_feature)
index = int(((len(pos_feature) + len(neg_feature)) / 2) * 0.2)
x_train = pos_feature[index:] + neg_feature[index:]
x_test = pos_feature[:index] + neg_feature[:index]
SVC_A, SVC_PP, SVC_PR, SVC_PF1, SVC_NP, SVC_NR, SVC_NF1 = Ngram.score(SVC(gamma=0.015), x_train, x_test)
# 记录 准确率（Accuracy）、正面精确率（pos_precision），正面召回率（pos_recall）、正面F1值、负面面精确率（neg_precision）、负面召回率（neg_recall）,负面F1值
log.console_out(filename, taskname, n,
('SVC-rbf', SVC_A, 'pos:', SVC_PP, SVC_PR, SVC_PF1, 'neg:', SVC_NP, SVC_NR, SVC_NF1))
def train_lstm(cls, n_symbols, embedding_weights, x_train, y_train, x_test,
y_test, use_word_dim):
model = Sequential()
model.add(
Embedding(output_dim=cls.vocab_dim,
input_dim=n_symbols,
mask_zero=True,
weights=[embedding_weights],
input_length=cls.input_length)) # Adding Input Length
model.add(
LSTM(units=50,
activation='sigmoid',
recurrent_activation='hard_sigmoid'))
model.add(Dropout(0.55)) # 防止过拟合的
model.add(Dense(1))
model.add(Activation('sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy']) # 编译模型
loss_history = LossHistory()
# early_stopping = EarlyStopping(monitor='val_loss', patience=3) # 当经过3轮验证集的loss不再下降时，中断训练
# print(x_test[:1])
# print('\n\n')
# x_test, y_test, x_val, y_val = train_test_split(x_test, y_test, test_size=0.5, shuffle=True) # 训练模型10%测试集，10%验证集合
# print(x_test[:1])
hist = model.fit(x_train,
y_train,
batch_size=cls.batch_size,
epochs=cls.n_epoch,
verbose=1,
validation_split=0.11111,
callbacks=[loss_history])
print('=============>history ',
hist.history) # 在每个epoch后记录训练/测试的loss和正确率
score = model.evaluate(x_test, y_test,
batch_size=cls.batch_size) # 评估模型
yaml_string = model.to_yaml()
with open(LSTM_YML_PATH.format(use_word_dim), 'w') as outfile:
outfile.write(yaml.dump(yaml_string, default_flow_style=True))
model.save_weights(LSTM_MODEL_PATH.format(use_word_dim))
loss_history.loss_plot('epoch')
print('Test score:', score)
log.console_out('lstm_w2v.txt', (250, hist.history, score))
def load_file(cls):
neg = pd.read_excel('corpus/negT.xlsx', header=None, index=None)
pos = pd.read_excel('corpus/posT.xlsx', header=None, index=None)
# pos[1] 即excle表格中第二列
cw = lambda x: cls.text_parse(x) # 定义分词函数
pos['words'] = pos[1].apply(cw)
neg['words'] = neg[1].apply(cw)
# print(pos)
log.console_out("log_svm.txt", "pos length =", len(pos), "neg length =", len(neg))
# use 1 for positive sentiment, 0 for negative
y = np.concatenate((np.ones(len(pos)), np.zeros(len(neg)))) # 合并语料
x_train, x_test, y_train, y_test = train_test_split(np.concatenate((pos['words'], neg['words'])), y,
test_size=0.2) # 划分训练和测试集合8/2
# print(x_train, '\n', y_train)
np.save('svm_data/y_train.npy', y_train)
np.save('svm_data/y_test.npy', y_test)
log.console_out("log_svm.txt", "load_file done!")
return x_train, x_test
def train(cls):
x_train, x_test = cls.load_file()
cls.save_train_vecs(x_train, x_test) # w2v计算词向量
train_vecs, y_train, test_vecs, y_test = cls.get_data()
# clf = SVC(kernel='rbf', verbose=True, probability=True)
clf = BernoulliNB()
# clf = LinearSVC()
# clf = LogisticRegression()
clf.fit(train_vecs, y_train)
joblib.dump(clf, 'svm_data/svm_model/model.pkl')
# print(test_vecs)
predict_y = clf.predict(test_vecs) # 基于SVM对验证集做出预测，prodict_y 为预测的结果
test_accuracy = metrics.accuracy_score(y_test, predict_y) # 验证集上的准确率
# y_pred = clf.predict(y_test)
# test_precision = metrics.precision_score(y_test, y_pred, average='weighted')
# test_recall = metrics.recall_score(y_test, y_pred, average='weighted')
log.console_out("log_svm.txt", "SVM score = ", clf.score(test_vecs, y_test)) # 记录得分
log.console_out("log_svm.txt", "test_accuracy = ", test_accuracy) # 记录准确率
def save_train_vecs(cls, x_train, x_test):
n_dim = 128
# Initialize model and build vocab
comment_w2v = Word2Vec(size=n_dim, min_count=5)
comment_w2v.build_vocab(x_train)
# Train the model over train_reviews (this may take several minutes)
comment_w2v.train(x_train, total_examples=comment_w2v.corpus_count, epochs=comment_w2v.iter)
train_vecs = np.concatenate([cls.build_wordvector(z, n_dim, comment_w2v) for z in x_train])
# train_vecs = scale(train_vecs)
np.save('svm_data/train_vecs.npy', train_vecs)
log.console_out("log_svm.txt", "save train vecs done!")
# Train word2vec on test tweets
comment_w2v.train(x_test, total_examples=comment_w2v.corpus_count, epochs=comment_w2v.iter)
comment_w2v.save('svm_data/w2v_model/w2v_model.pkl')
# Build test tweet vectors then scale
test_vecs = np.concatenate([cls.build_wordvector(z, n_dim, comment_w2v) for z in x_test])
# test_vecs = scale(test_vecs)
np.save('svm_data/test_vecs.npy', test_vecs)
log.console_out("log_svm.txt", "save test vecs done!")
log.console_out("log_svm.txt", "save_train_vecs function done!")
'uni_bi_tri', 'uni_info_bi_tri'
]
ns = [1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000]
# nss = [11000, 12000, 13000, 14000, 15000, 16000, 17000, 18000, 19000, 20000]
# for i in range(6, 7):
# filename = 'record4.txt'
# if tasks[i] == tasks[6]: # uni_info_bi_tri
# for n in nss:
# Ngram.record_res(filename, tasks[i], i, n)
for i in range(0, 7):
filename = 'test.txt'
if tasks[i] == tasks[0]: # uni
Ngram.record_res(filename, tasks[i], i, 0)
elif tasks[i] == tasks[1]: # bi
Ngram.record_res(filename, tasks[i], i, 0)
elif tasks[i] == tasks[2]: # uni_bi
Ngram.record_res(filename, tasks[i], i, 0)
elif tasks[i] == tasks[3]: # uni_info
for n in ns:
Ngram.record_res(filename, tasks[i], i, n)
elif tasks[i] == tasks[4]: # uni_info_bi
for n in ns:
Ngram.record_res(filename, tasks[i], i, n)
elif tasks[i] == tasks[5]: # uni_bi_tri
Ngram.record_res(filename, tasks[i], i, 0)
elif tasks[i] == tasks[6]: # uni_info_bi_tri
for n in ns:
Ngram.record_res(filename, tasks[i], i, n)
log.console_out(filename, "All Tasks Done")
print(time.strftime('%Y-%m-%d %A %H:%M:%S', time.localtime(time.time())))