def run():
if FLAGS.log_file_name:
sys.stdout = open(FLAGS.log_file_name, 'w')
tf.reset_default_graph()
localtime = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("***********localtime: ", localtime)
x_data, y_data, sen_len_data, doc_len_data, word_distance, word_embedding, pos_embedding = func.load_data(
)
word_embedding = tf.constant(word_embedding,
dtype=tf.float32,
name='word_embedding')
pos_embedding = tf.constant(pos_embedding,
dtype=tf.float32,
name='pos_embedding')
print('build model...')
start_time = time.time()
x = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len, FLAGS.max_sen_len])
y = tf.placeholder(tf.float32, [None, FLAGS.max_doc_len, FLAGS.n_class])
sen_len = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len])
doc_len = tf.placeholder(tf.int32, [None])
word_dis = tf.placeholder(tf.int32,
[None, FLAGS.max_doc_len, FLAGS.max_sen_len])
keep_prob1 = tf.placeholder(tf.float32)
keep_prob2 = tf.placeholder(tf.float32)
placeholders = [x, y, sen_len, doc_len, word_dis, keep_prob1, keep_prob2]
pred, reg, pred_assist_list, reg_assist_list = build_model(
x, sen_len, doc_len, word_dis, word_embedding, pos_embedding,
keep_prob1, keep_prob2)
with tf.name_scope('loss'):
valid_num = tf.cast(tf.reduce_sum(doc_len), dtype=tf.float32)
loss_op = -tf.reduce_sum(
y * tf.log(pred)) / valid_num + reg * FLAGS.l2_reg
loss_assist_list = []
for i in range(FLAGS.n_layers - 1):
loss_assist = -tf.reduce_sum(y * tf.log(pred_assist_list[
i])) / valid_num + reg_assist_list[i] * FLAGS.l2_reg
loss_assist_list.append(loss_assist)
with tf.name_scope('train'):
optimizer = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_main).minimize(loss_op)
optimizer_assist_list = []
for i in range(FLAGS.n_layers - 1):
if i == 0:
optimizer_assist = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_assist).minimize(
loss_assist_list[i])
else:
optimizer_assist = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_main).minimize(loss_assist_list[i])
optimizer_assist_list.append(optimizer_assist)
true_y_op = tf.argmax(y, 2)
pred_y_op = tf.argmax(pred, 2)
pred_y_assist_op_list = []
for i in range(FLAGS.n_layers - 1):
pred_y_assist_op = tf.argmax(pred_assist_list[i], 2)
pred_y_assist_op_list.append(pred_y_assist_op)
print('build model done!\n')
prob_list_pr, y_label = [], []
# Training Code Block
print_training_info()
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
kf, fold, SID = KFold(n_splits=10), 1, 0
Id = []
p_list, r_list, f1_list = [], [], []
for train, test in kf.split(x_data):
tr_x, tr_y, tr_sen_len, tr_doc_len, tr_word_dis = map(
lambda x: x[train],
[x_data, y_data, sen_len_data, doc_len_data, word_distance])
te_x, te_y, te_sen_len, te_doc_len, te_word_dis = map(
lambda x: x[test],
[x_data, y_data, sen_len_data, doc_len_data, word_distance])
precision_list, recall_list, FF1_list = [], [], []
pre_list, true_list, pre_list_prob = [], [], []
sess.run(tf.global_variables_initializer())
print('############# fold {} ###############'.format(fold))
fold += 1
max_f1 = 0.0
print('train docs: {} test docs: {}'.format(
len(tr_y), len(te_y)))
'''*********GP*********'''
for layer in range(FLAGS.n_layers - 1):
if layer == 0:
training_iter = FLAGS.training_iter
else:
training_iter = FLAGS.training_iter - 5
for i in range(training_iter):
step = 1
for train, _ in get_batch_data(tr_x, tr_y, tr_sen_len,
tr_doc_len, tr_word_dis,
FLAGS.keep_prob1,
FLAGS.keep_prob2,
FLAGS.batch_size):
_, loss, pred_y, true_y, pred_prob, doc_len_batch = sess.run(
[
optimizer_assist_list[layer],
loss_assist_list[layer],
pred_y_assist_op_list[layer], true_y_op,
pred_assist_list[layer], doc_len
],
feed_dict=dict(zip(placeholders, train)))
acc_assist, p_assist, r_assist, f1_assist = func.acc_prf(
pred_y, true_y, doc_len_batch)
if step % 10 == 0:
print(
'GL{}: epoch {}: step {}: loss {:.4f} acc {:.4f}'
.format(layer + 1, i + 1, step, loss,
acc_assist))
step = step + 1
'''*********Train********'''
for epoch in range(FLAGS.training_iter):
step = 1
for train, _ in get_batch_data(tr_x, tr_y, tr_sen_len,
tr_doc_len, tr_word_dis,
FLAGS.keep_prob1,
FLAGS.keep_prob2,
FLAGS.batch_size):
_, loss, pred_y, true_y, pred_prob, doc_len_batch = sess.run(
[
optimizer, loss_op, pred_y_op, true_y_op, pred,
doc_len
],
feed_dict=dict(zip(placeholders, train)))
acc, p, r, f1 = func.acc_prf(pred_y, true_y, doc_len_batch)
if step % 5 == 0:
print(
'epoch {}: step {}: loss {:.4f} acc {:.4f}'.format(
epoch + 1, step, loss, acc))
step = step + 1
'''*********Test********'''
test = [
te_x, te_y, te_sen_len, te_doc_len, te_word_dis, 1., 1.
]
loss, pred_y, true_y, pred_prob = sess.run(
[loss_op, pred_y_op, true_y_op, pred],
feed_dict=dict(zip(placeholders, test)))
end_time = time.time()
true_list.append(true_y)
pre_list.append(pred_y)
pre_list_prob.append(pred_prob)
acc, p, r, f1 = func.acc_prf(pred_y, true_y, te_doc_len)
precision_list.append(p)
recall_list.append(r)
FF1_list.append(f1)
if f1 > max_f1:
max_acc, max_p, max_r, max_f1 = acc, p, r, f1
print(
'\ntest: epoch {}: loss {:.4f} acc {:.4f}\np: {:.4f} r: {:.4f} f1: {:.4f} max_f1 {:.4f}\n'
.format(epoch + 1, loss, acc, p, r, f1, max_f1))
Id.append(len(te_x))
SID = np.sum(Id) - len(te_x)
_, maxIndex = func.maxS(FF1_list)
print("maxIndex:", maxIndex)
print('Optimization Finished!\n')
pred_prob = pre_list_prob[maxIndex]
for i in range(pred_y.shape[0]):
for j in range(te_doc_len[i]):
prob_list_pr.append(pred_prob[i][j][1])
y_label.append(true_y[i][j])
p_list.append(max_p)
r_list.append(max_r)
f1_list.append(max_f1)
print("running time: ", str((end_time - start_time) / 60.))
print_training_info()
p, r, f1 = map(lambda x: np.array(x).mean(), [p_list, r_list, f1_list])
print("f1_score in 10 fold: {}\naverage : {} {} {}\n".format(
np.array(f1_list).reshape(-1, 1), round(p, 4), round(r, 4),
round(f1, 4)))
return p, r, f1
def run():
if FLAGS.log_file_name:
sys.stdout = open(FLAGS.log_file_name, 'w')
tf.reset_default_graph()
localtime = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("***********localtime: ", localtime)
x_data, y_data, sen_len_data, doc_len_data, word_distance, word_embedding, pos_embedding = func.load_data(
)
word_embedding = tf.constant(word_embedding,
dtype=tf.float32,
name='word_embedding')
pos_embedding = tf.constant(pos_embedding,
dtype=tf.float32,
name='pos_embedding')
print('build model...')
x = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len, FLAGS.max_sen_len])
word_dis = tf.placeholder(tf.int32,
[None, FLAGS.max_doc_len, FLAGS.max_sen_len])
sen_len = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len])
doc_len = tf.placeholder(tf.int32, [None])
keep_prob1 = tf.placeholder(tf.float32)
keep_prob2 = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32, [None, FLAGS.max_doc_len, FLAGS.n_class])
placeholders = [x, word_dis, sen_len, doc_len, keep_prob1, keep_prob2, y]
with tf.name_scope('loss'):
pred, reg = build_model(word_embedding, pos_embedding, x, word_dis,
sen_len, doc_len, keep_prob1, keep_prob2)
valid_num = tf.cast(tf.reduce_sum(doc_len), dtype=tf.float32)
loss_op = -tf.reduce_sum(
y * tf.log(pred)) / valid_num + reg * FLAGS.l2_reg
with tf.name_scope('train'):
optimizer = tf.train.AdamOptimizer(
learning_rate=FLAGS.learning_rate).minimize(loss_op)
true_y_op = tf.argmax(y, 2)
pred_y_op = tf.argmax(pred, 2)
print('build model done!\n')
# Training Code Block
print_training_info()
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
kf, fold, SID = KFold(n_splits=10), 1, 0
Id = []
p_list, r_list, f1_list = [], [], []
true_result_all, pre_result_all = [], []
start_time = time.time()
all0_sum, multi1_sum, pred_multi_cause_sum = [], [], []
for train, test in kf.split(x_data):
tr_x, tr_y, tr_sen_len, tr_doc_len, tr_word_dis = map(
lambda x: x[train],
[x_data, y_data, sen_len_data, doc_len_data, word_distance])
te_x, te_y, te_sen_len, te_doc_len, te_word_dis = map(
lambda x: x[test],
[x_data, y_data, sen_len_data, doc_len_data, word_distance])
precision_list, recall_list, FF1_list = [], [], []
pre_list, true_list, pre_list_prob = [], [], []
sess.run(tf.global_variables_initializer())
print('############# fold {} ###############'.format(fold))
fold += 1
max_f1, max_f1_rectify = 0.0, 0.0
print('train docs: {} test docs: {}'.format(
len(tr_y), len(te_y)))
for epoch in range(FLAGS.training_iter):
step = 1
# ************train************
for train, _ in get_batch_data(tr_x, tr_word_dis, tr_sen_len,
tr_doc_len, FLAGS.keep_prob1,
FLAGS.keep_prob2, tr_y,
FLAGS.batch_size):
_, loss, pred_y, true_y, pred_prob, doc_len_batch = sess.run(
[
optimizer, loss_op, pred_y_op, true_y_op, pred,
doc_len
],
feed_dict=dict(zip(placeholders, train)))
acc, p, r, f1 = func.acc_prf(pred_y, true_y, doc_len_batch)
if step % 5 == 0:
print(
'epoch {}: step {}: loss {:.4f} acc {:.4f}'.format(
epoch + 1, step, loss, acc))
step = step + 1
# ************test************
test = [
te_x, te_word_dis, te_sen_len, te_doc_len, 1., 1., te_y
]
loss, pred_y, true_y, pred_prob = sess.run(
[loss_op, pred_y_op, true_y_op, pred],
feed_dict=dict(zip(placeholders, test)))
true_list.append(true_y)
pre_list.append(pred_y)
pre_list_prob.append(pred_prob)
acc, p, r, f1 = func.acc_prf(pred_y, true_y, te_doc_len)
precision_list.append(p)
recall_list.append(r)
FF1_list.append(f1)
if f1 > max_f1:
max_acc, max_p, max_r, max_f1 = acc, p, r, f1
print(
'\nepoch {}: loss {:.4f} acc {:.4f}\n\nnorectify: p {:.4f} r {:.4f} f1 {:.4f} max_f1 {:.4f}'
.format(epoch + 1, loss, acc, p, r, f1, max_f1))
Id.append(len(te_x))
SID = np.sum(Id) - len(te_x)
_, maxIndex = func.maxS(FF1_list)
print('Optimization Finished!\n')
p_list.append(max_p)
r_list.append(max_r)
f1_list.append(max_f1)
end_time = time.time()
print("running time: ", str((end_time - start_time) / 60.))
print_training_info()
p, r, f1 = map(lambda x: np.array(x).mean(), [p_list, r_list, f1_list])
print("f1_score in 10 fold: {}\naverage : {} {} {}\n".format(
np.array(f1_list).reshape(-1, 1), round(p, 4), round(r, 4),
round(f1, 4)))
return p, r, f1
def New(self):
self.cause_edt.clear()
senID = random.randint(1, 2105)
the_line = linecache.getline('./data/datacsv_2105.csv', senID)
the_line = the_line.strip().split(',')
the_line = the_line[2]
self.clause_edt.setPlainText(the_line)
# 生成测试集
#func.load_data()：return x, y_position, y, sen_len, doc_len, relative_pos, relative_pos_a, embedding, embedding_pos
x_data, y_position_data, y_data, sen_len_data, doc_len_data, word_distance, word_distance_a, word_distance_e, word_embedding, pos_embedding, pos_embedding_a, pos_embedding_e = func.load_data(
)
print("senID:{}\n".format(senID))
te_x = x_data[senID]
te_x = te_x[np.newaxis, :]
te_sen_len = sen_len_data[senID]
te_sen_len = te_sen_len[np.newaxis, :]
te_doc_len = np.array([doc_len_data[senID]])
te_word_dis = word_distance_e[senID]
te_word_dis = te_word_dis[np.newaxis, :]
print("te_x.shape:{}\n".format(te_x.shape))
print("te_sen_len.shape:{}\n".format(te_sen_len.shape))
print("te_doc_len:{}\n".format(te_doc_len))
print("te_doc_len.shape:{}\n".format(te_doc_len.shape))
print("te_word_dis.shape:{}\n".format(te_word_dis.shape))
test = [
te_x, te_sen_len, te_doc_len, te_word_dis, 1., 1., word_embedding
]
tf.reset_default_graph()
with tf.Session() as sess:
# reload训练好的模型
saver = tf.train.import_meta_graph(
'run_final_ee/model.ckpt-13.meta')
model_file = tf.train.latest_checkpoint('run_final_ee/')
saver.restore(sess, model_file)
tenboard_dir = './tensorboard/RTHN_EE'
graph = tf.get_default_graph()
writer = tf.summary.FileWriter(tenboard_dir, graph)
writer.add_graph(sess.graph)
x = graph.get_tensor_by_name("x:0")
word_dis = graph.get_tensor_by_name("word_dis:0")
sen_len = graph.get_tensor_by_name("sen_len:0")
doc_len = graph.get_tensor_by_name("doc_len:0")
keep_prob1 = graph.get_tensor_by_name("keep_prob1:0")
keep_prob2 = graph.get_tensor_by_name("keep_prob2:0")
pred_y_op = graph.get_tensor_by_name("pred_y_op:0")
pred_pos_op = graph.get_tensor_by_name("pred_pos_op:0")
word_embedding = graph.get_tensor_by_name("word_embedding:0")
placeholders = [
x, sen_len, doc_len, word_dis, keep_prob1, keep_prob2,
word_embedding
]
# placeholders = [x, y, sen_len, doc_len, word_dis, keep_prob1, keep_prob2]
# 将测试集传入模型进行训练
pred_pos, pred_y = sess.run([pred_pos_op, pred_y_op],
feed_dict=dict(zip(placeholders,
test)))
# print("pred_pos.shape:{}",pred_pos.shape)
# print("pred_y.shape:{}",pred_y.shape)
pred_pos = pred_pos.reshape(75, )
pred_y = pred_y.reshape(75, )
print("pred_pos:{}", pred_pos)
print("pred_y:{}", pred_y)
emo_ind = np.argwhere(pred_pos == 1)[:, 0]
cla_ind = np.argwhere(pred_y == 1)[:, 0]
# cla_ind = int(cla_ind)
print("emo_ind:{}".format(emo_ind))
print("cla_ind:{}".format(cla_ind))
emo_num = len(emo_ind)
cau_num = len(cla_ind)
print("emo_num:{}".format(emo_num))
print("cau_num:{}".format(cau_num))
if cau_num == 0:
self.cause_edt.setPlainText("该句子中不存在原因子句")
else:
# 通过训练结果找出原因子句
self.cause_edt.clear()
inputFile1 = codecs.open('./data/clause_keywords.csv', 'r',
'utf-8')
i = 0
clause_all = []
for line in inputFile1.readlines():
line = line.strip().split(',')
sen_id, cla_id, clause = int(line[0]), int(
line[1]), line[-1]
if sen_id == senID:
clause_all.append(clause.replace(' ', ''))
clause_all = np.array(clause_all)
for i in range(0, cau_num):
self.cause_edt.append(clause_all[cla_ind[i]])
# self.cause_edt.setPlainText(clause.replace(' ', ''))
# print(clause_all)
inputFile1.close()
if emo_num == 0:
self.emotion_edt.setPlainText("该句子中不存在情感子句")
else:
# 通过训练结果找出情感子句
self.emotion_edt.clear()
inputFile2 = codecs.open('./data/clause_keywords.csv', 'r',
'utf-8')
i = 0
clause_e_all = []
for line in inputFile2.readlines():
line = line.strip().split(',')
sen_id, cla_id, clause = int(line[0]), int(
line[1]), line[-1]
if sen_id == senID:
for j in range(0, emo_num):
if cla_id == emo_ind[i] + 1:
self.emotion_edt.append(clause.replace(
' ', ''))
# self.emotion_edt.setPlainText(clause.replace(' ', ''))
# print(clause_e_all)
inputFile2.close()
def run():
if FLAGS.log_file_name:
sys.stdout = open(FLAGS.log_file_name, 'w')
tf.reset_default_graph()
localtime = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("***********localtime: ", localtime)
#func.load_data()：return x, y, sen_len, doc_len, relative_pos, embedding, embedding_pos
#需要将word_distance改为自己计算的结果
x_data, y_position_data, y_data, sen_len_data, doc_len_data, word_distance, word_distance_a, word_distance_e, word_embedding, pos_embedding, pos_embedding_a, pos_embedding_ap = func.load_data(
)
# print("x_data.shape:{}\n".format(x_data.shape))
# print("y_data.shape:{}\n".format(y_data.shape))
# print("sen_len_data.shape:{}\n".format(sen_len_data.shape))
# print("doc_len_data.shape:{}\n".format(doc_len_data.shape))
# print("word_distance.shape:{}\n".format(word_distance.shape))
# print("word_embedding.shape:{}\n".format(word_embedding.shape))
# print("pos_embedding.shape:{}\n".format(pos_embedding.shape))
# print("pos_embedding:{}\n".format(pos_embedding[1]))
word_embedding = tf.constant(word_embedding,
dtype=tf.float32,
name='word_embedding')
pos_embedding = tf.constant(pos_embedding,
dtype=tf.float32,
name='pos_embedding')
print('build model...')
start_time = time.time()
#定义placeholder
x = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len, FLAGS.max_sen_len],
name="x")
y = tf.placeholder(tf.float32, [None, FLAGS.max_doc_len, FLAGS.n_class],
name="y")
sen_len = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len],
name="sen_len")
doc_len = tf.placeholder(tf.int32, [None], name="doc_len")
word_dis = tf.placeholder(tf.int32,
[None, FLAGS.max_doc_len, FLAGS.max_sen_len],
name="word_dis")
keep_prob1 = tf.placeholder(tf.float32, name="keep_prob1")
keep_prob2 = tf.placeholder(tf.float32, name="keep_prob2")
placeholders = [x, y, sen_len, doc_len, word_dis, keep_prob1, keep_prob2]
pred, reg = build_model(x, sen_len, doc_len, word_dis, word_embedding,
pos_embedding, keep_prob1, keep_prob2)
# print(pred)
with tf.name_scope('loss'):
valid_num = tf.cast(tf.reduce_sum(doc_len), dtype=tf.float32)
loss_op = -tf.reduce_sum(
y * tf.log(pred)) / valid_num + reg * FLAGS.l2_reg
with tf.name_scope('train'):
optimizer = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_main).minimize(loss_op)
true_y_op = tf.argmax(y, 2, name="true_y_op")
pred_y_op = tf.argmax(pred, 2, name="pred_y_op")
print('build model done!\n')
########训练和验证过程#########
prob_list_pr, y_label = [], []
# Training Code Block
print_training_info()
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
# saver = tf.train.Saver(max_to_keep=4)
#
# tenboard_dir = './tensorboard/RTHN'
# graph = tf.get_default_graph()
# writer = tf.summary.FileWriter(tenboard_dir, graph)
with tf.Session(config=tf_config) as sess:
# writer.add_graph(sess.graph)
kf, fold, SID = KFold(n_splits=10), 1, 0 #十折交叉验证
Id = []
p_list, r_list, f1_list = [], [], []
for train, test in kf.split(x_data):
tr_x, tr_y, tr_sen_len, tr_doc_len, tr_word_dis = map(
lambda x: x[train],
[x_data, y_data, sen_len_data, doc_len_data, word_distance])
te_x, te_y, te_sen_len, te_doc_len, te_word_dis = map(
lambda x: x[test],
[x_data, y_data, sen_len_data, doc_len_data, word_distance])
precision_list, recall_list, FF1_list = [], [], []
pre_list, true_list, pre_list_prob = [], [], []
sess.run(tf.global_variables_initializer())
print('############# fold {} ###############'.format(fold))
fold += 1
max_f1 = 0.0
print('train docs: {} test docs: {}'.format(
len(tr_y), len(te_y)))
'''*********Train********'''
for epoch in range(FLAGS.training_iter):
step = 1
#train：feed_list = [x[index], y[index], sen_len[index], doc_len[index], word_dis[index], keep_prob1, keep_prob2]
for train, _ in get_batch_data(tr_x, tr_y, tr_sen_len,
tr_doc_len, tr_word_dis,
FLAGS.keep_prob1,
FLAGS.keep_prob2,
FLAGS.batch_size):
_, loss, pred_y, true_y, pred_prob, doc_len_batch = sess.run(
[
optimizer, loss_op, pred_y_op, true_y_op, pred,
doc_len
],
feed_dict=dict(zip(placeholders, train)))
acc, p, r, f1 = func.acc_prf(pred_y, true_y, doc_len_batch)
# if step % 10 == 0:
# print('epoch {}: step {}: loss {:.4f} acc {:.4f}'.format(epoch + 1, step, loss, acc))
step = step + 1
# print("begin save!")
# saver.save(sess, "./run_final/model.ckpt", global_step=step)
'''*********Test********'''
test = [
te_x, te_y, te_sen_len, te_doc_len, te_word_dis, 1., 1.
]
loss, pred_y, true_y, pred_prob = sess.run(
[loss_op, pred_y_op, true_y_op, pred],
feed_dict=dict(zip(placeholders, test)))
end_time = time.time()
true_list.append(true_y)
pre_list.append(pred_y)
pre_list_prob.append(pred_prob)
#计算精确率准确率召回率和F值
acc, p, r, f1 = func.acc_prf(pred_y, true_y, te_doc_len)
precision_list.append(p)
recall_list.append(r)
FF1_list.append(f1)
if f1 > max_f1:
max_acc, max_p, max_r, max_f1 = acc, p, r, f1
# print('\ntest: epoch {}: loss {:.4f} acc {:.4f}\np: {:.4f} r: {:.4f} f1: {:.4f} max_f1 {:.4f}\n'.format(
# epoch + 1, loss, acc, p, r, f1, max_f1))
Id.append(len(te_x))
SID = np.sum(Id) - len(te_x)
_, maxIndex = func.maxS(FF1_list)
# print("maxIndex:", maxIndex)
# print('Optimization Finished!\n')
pred_prob = pre_list_prob[maxIndex]
for i in range(pred_y.shape[0]):
for j in range(te_doc_len[i]):
prob_list_pr.append(pred_prob[i][j][1])
y_label.append(true_y[i][j])
p_list.append(max_p)
r_list.append(max_r)
f1_list.append(max_f1)
print("running time: ", str((end_time - start_time) / 60.))
print_training_info()
p, r, f1 = map(lambda x: np.array(x).mean(), [p_list, r_list, f1_list])
# print("f1_score in 10 fold: {}\naverage : {} {} {}\n".format(np.array(f1_list).reshape(-1, 1), round(p, 4), round(r, 4), round(f1, 4)))
# writer.close()
return p, r, f1
def run():
save_dir = 'result_data_{}/'.format(FLAGS.save_dir)
if FLAGS.log_file_name:
sys.stdout = open(FLAGS.log_file_name, 'w')
tf.reset_default_graph()
localtime = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("***********localtime: ", localtime)
x_data, y_data, y_emotion_data, y_cause_data, sen_len_data, doc_len_data, word_embedding, doc_id_data, idx_word_dict, clause_position_data, embedding_pos, _ = func.load_data(
)
word_embedding = tf.constant(word_embedding,
dtype=tf.float32,
name='word_embedding')
embedding_pos = tf.constant(embedding_pos,
dtype=tf.float32,
name='embedding_pos')
print('build model...')
start_time = time.time()
x = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len, FLAGS.max_sen_len])
y = tf.placeholder(tf.float32, [None, FLAGS.max_doc_len, FLAGS.n_class])
y_emotion = tf.placeholder(
tf.float32, [None, FLAGS.max_doc_len, FLAGS.emotion_n_class])
y_cause = tf.placeholder(tf.float32,
[None, FLAGS.max_doc_len, FLAGS.cause_n_class])
sen_len = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len])
doc_len = tf.placeholder(tf.int32, [None])
clause_position = tf.placeholder(tf.int32, [None, FLAGS.max_doc_len])
keep_prob1 = tf.placeholder(tf.float32)
keep_prob2 = tf.placeholder(tf.float32)
placeholders_emotion_assist = [
x, y_emotion, sen_len, doc_len, clause_position, keep_prob1, keep_prob2
]
placeholders_cause_assist = [
x, y_cause, sen_len, doc_len, clause_position, keep_prob1, keep_prob2
]
placeholders_main = [
x, y, sen_len, doc_len, clause_position, keep_prob1, keep_prob2
]
pred, reg, pred_emotion_assist_list, reg_emotion_assist_list, pred_cause_assist_list, reg_cause_assist_list =\
build_model(x, sen_len, doc_len, word_embedding, clause_position, embedding_pos, keep_prob1, keep_prob2)
with tf.name_scope('loss'):
valid_num = tf.cast(tf.reduce_sum(doc_len), dtype=tf.float32)
loss_op = -tf.reduce_sum(
y * tf.log(pred)) / valid_num + reg * FLAGS.l2_reg
loss_emotion_assist_list, loss_cause_assist_list = [], []
for i in range(FLAGS.assist_n_layers - 1):
loss_emotion_assist = -tf.reduce_sum(
y_emotion * tf.log(pred_emotion_assist_list[i])
) / valid_num + reg_emotion_assist_list[i] * FLAGS.l2_reg
loss_emotion_assist_list.append(loss_emotion_assist)
loss_cause_assist = -tf.reduce_sum(
y_cause * tf.log(pred_cause_assist_list[i])
) / valid_num + reg_cause_assist_list[i] * FLAGS.l2_reg
loss_cause_assist_list.append(loss_cause_assist)
with tf.name_scope('train'):
optimizer = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_main).minimize(loss_op)
optimizer_emotion_assist_list, optimizer_cause_assist_list = [], []
for i in range(FLAGS.assist_n_layers - 1):
if i == 0:
optimizer_emotion_assist = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_assist).minimize(
loss_emotion_assist_list[i])
optimizer_cause_assist = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_assist).minimize(
loss_cause_assist_list[i])
else:
optimizer_emotion_assist = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_main).minimize(
loss_emotion_assist_list[i])
optimizer_cause_assist = tf.train.AdamOptimizer(
learning_rate=FLAGS.lr_assist).minimize(
loss_cause_assist_list[i])
optimizer_emotion_assist_list.append(optimizer_emotion_assist)
optimizer_cause_assist_list.append(optimizer_cause_assist)
true_y_op = tf.argmax(y, 2)
pred_y_op = tf.argmax(pred, 2)
emotion_true_y_op = tf.argmax(y_emotion, 2)
cause_true_y_op = tf.argmax(y_cause, 2)
pred_y_emotion_assist_op_list, pred_y_cause_assist_op_list = [], []
for i in range(FLAGS.assist_n_layers - 1):
pred_y_emotion_assist_op = tf.argmax(pred_emotion_assist_list[i], 2)
pred_y_emotion_assist_op_list.append(pred_y_emotion_assist_op)
pred_y_cause_assist_op = tf.argmax(pred_cause_assist_list[i], 2)
pred_y_cause_assist_op_list.append(pred_y_cause_assist_op)
print('build model done!\n')
prob_list_pr, y_label = [], []
# Training Code Block
print_training_info()
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
kf, fold, SID = KFold(n_splits=10), 1, 0
Id = []
p_list, r_list, f1_list = [], [], []
for train, test in kf.split(x_data):
tr_x, tr_y, tr_y_emotion, tr_y_cause, tr_sen_len, tr_doc_len, tr_doc_id, tr_clause_pos = map(
lambda x: x[train], [
x_data, y_data, y_emotion_data, y_cause_data, sen_len_data,
doc_len_data, doc_id_data, clause_position_data
])
te_x, te_y, te_y_emotion, te_y_cause, te_sen_len, te_doc_len, te_doc_id, te_clause_pos = map(
lambda x: x[test], [
x_data, y_data, y_emotion_data, y_cause_data, sen_len_data,
doc_len_data, doc_id_data, clause_position_data
])
precision_list, recall_list, FF1_list = [], [], []
pre_list, true_list, pre_list_prob = [], [], []
# file_dir = './ECPredictor_demo_assistlayer4_mainlayer1/fold{}/'.format(fold)
# if not os.path.exists(file_dir):
# os.mkdir(file_dir)
sess.run(tf.global_variables_initializer())
print('############# fold {} ###############'.format(fold))
fold += 1
max_f1 = 0.0
emo_max_f1 = 0.0
cause_max_f1 = 0.0
print('train docs: {} test docs: {}'.format(
len(tr_y), len(te_y)))
for layer in range(FLAGS.assist_n_layers - 1):
if layer == 0:
training_iter = FLAGS.training_iter
else:
training_iter = FLAGS.training_iter - 5
for i in range(training_iter):
'''预训练情感分类'''
emotion_step = 1
for train, _ in get_emotion_batch_data(
tr_x, tr_y_emotion, tr_sen_len, tr_doc_len,
tr_clause_pos, FLAGS.keep_prob1, FLAGS.keep_prob2,
FLAGS.batch_size):
_, loss, pred_y, true_y, doc_len_batch = sess.run(
[
optimizer_emotion_assist_list[layer],
loss_emotion_assist_list[layer],
pred_y_emotion_assist_op_list[layer],
emotion_true_y_op, doc_len
],
feed_dict=dict(
zip(placeholders_emotion_assist, train)))
acc_assist, p_assist, r_assist, f1_assist = func.acc_prf_binary(
pred_y, true_y, doc_len_batch)
if emotion_step % 10 == 0:
print(
'Emotion {}: epoch {}: step {}: loss {:.4f} acc {:.4f} p {:.4f}'
.format(layer + 1, i + 1, emotion_step, loss,
acc_assist, p_assist))
emotion_step = emotion_step + 1
"""****test emotion extraction****"""
test = [
te_x, te_y_emotion, te_sen_len, te_doc_len,
te_clause_pos, 1., 1.
]
loss, pred_y, true_y, doc_len_batch = sess.run(
[
loss_emotion_assist_list[layer],
pred_y_emotion_assist_op_list[layer],
emotion_true_y_op, doc_len
],
feed_dict=dict(zip(placeholders_emotion_assist, test)))
acc_test, p_test, r_test, f1_test = func.acc_prf_binary(
pred_y, true_y, doc_len_batch)
if emo_max_f1 < f1_test:
emo_max_f1 = f1_test
print(
'\nemotion-test: epoch {}: loss {:.4f} acc {:.4f}\np: {:.4f} r: {:.4f} f1: {:.4f} max_f1 {:.4f}\n'
.format(i + 1, loss, acc_test, p_test, r_test, f1_test,
emo_max_f1))
# file_name = file_dir + 'emotion_test_layer{}_epoch{}.txt'.format(layer + 1, i+1)
# func.output_pred(file_name, te_doc_id, te_x, te_doc_len, te_sen_len, true_y, pred_y, idx_word_dict)
"""预训练原因分类"""
cause_step = 1
for train, _ in get_cause_batch_data(
tr_x, tr_y_cause, tr_sen_len, tr_doc_len,
tr_clause_pos, FLAGS.keep_prob1, FLAGS.keep_prob2,
FLAGS.batch_size):
_, loss, pred_y, true_y, doc_len_batch = sess.run(
[
optimizer_cause_assist_list[layer],
loss_cause_assist_list[layer],
pred_y_cause_assist_op_list[layer],
cause_true_y_op, doc_len
],
feed_dict=dict(
zip(placeholders_cause_assist, train)))
acc_assist, p_assist, r_assist, f1_assist = func.acc_prf_binary(
pred_y, true_y, doc_len_batch)
if cause_step % 10 == 0:
print(
'Cause {}: epoch {}: step {}: loss {:.4f} acc {:.4f} p {:.4f}'
.format(layer + 1, i + 1, cause_step, loss,
acc_assist, p_assist))
cause_step = cause_step + 1
"""****test cause extraction****"""
test = [
te_x, te_y_cause, te_sen_len, te_doc_len,
te_clause_pos, 1., 1.
]
loss, pred_y, true_y, doc_len_batch = sess.run(
[
loss_cause_assist_list[layer],
pred_y_cause_assist_op_list[layer],
cause_true_y_op, doc_len
],
feed_dict=dict(zip(placeholders_cause_assist, test)))
acc_test, p_test, r_test, f1_test = func.acc_prf_binary(
pred_y, true_y, doc_len_batch)
if cause_max_f1 < f1_test:
cause_max_f1 = f1_test
print(
'\ncause-test: epoch {}: loss {:.4f} acc {:.4f}\np: {:.4f} r: {:.4f} f1: {:.4f} max_f1 {:.4f}\n'
.format(i + 1, loss, acc_test, p_test, r_test, f1_test,
cause_max_f1))
# file_name = file_dir + 'cause_test_layer{}_epoch{}.txt'.format(layer+1, i+1)
# func.output_pred(file_name, te_doc_id, te_x, te_doc_len, te_sen_len, true_y, pred_y, idx_word_dict)
'''*********Train********'''
for epoch in range(FLAGS.training_iter):
step = 1
for train, _ in get_batch_data(tr_x, tr_y, tr_sen_len,
tr_doc_len, tr_clause_pos,
FLAGS.keep_prob1,
FLAGS.keep_prob2,
FLAGS.batch_size):
_, loss, pred_y, true_y, pred_prob, doc_len_batch = sess.run(
[
optimizer, loss_op, pred_y_op, true_y_op, pred,
doc_len
],
feed_dict=dict(zip(placeholders_main, train)))
acc, p, r, f1 = func.acc_prf_multiclass(
pred_y, true_y, doc_len_batch)
if step % 5 == 0:
print(
'epoch {}: step {}: loss {:.4f} acc {:.4f} p {:.4}'
.format(epoch + 1, step, loss, acc, p))
step = step + 1
'''*********Test********'''
test = [
te_x, te_y, te_sen_len, te_doc_len, te_clause_pos, 1., 1.
]
loss, pred_y, true_y, pred_prob = sess.run(
[loss_op, pred_y_op, true_y_op, pred],
feed_dict=dict(zip(placeholders_main, test)))
end_time = time.time()
true_list.append(true_y)
pre_list.append(pred_y)
pre_list_prob.append(pred_prob)
acc, p, r, f1 = func.acc_prf_multiclass(
pred_y, true_y, te_doc_len)
precision_list.append(p)
recall_list.append(r)
FF1_list.append(f1)
if f1 > max_f1:
max_acc, max_p, max_r, max_f1 = acc, p, r, f1
print(
'\ntest: epoch {}: loss {:.4f} acc {:.4f}\np: {:.4f} r: {:.4f} f1: {:.4f} max_f1 {:.4f}\n'
.format(epoch + 1, loss, acc, p, r, f1, max_f1))
# file_name = file_dir + 'emotion_cause_test_{}.txt'.format(epoch+1)
# func.output_pred(file_name, te_doc_id, te_x, te_doc_len, te_sen_len, true_y, pred_y, idx_word_dict)
Id.append(len(te_x))
SID = np.sum(Id) - len(te_x)
_, maxIndex = func.maxS(FF1_list)
print("maxIndex:", maxIndex)
print('Optimization Finished!\n')
pred_prob = pre_list_prob[maxIndex]
for i in range(pred_y.shape[0]):
for j in range(te_doc_len[i]):
prob_list_pr.append(pred_prob[i][j][1])
y_label.append(true_y[i][j])
print("*********prob_list_pr", len(prob_list_pr))
print("*********y_label", len(y_label))
p_list.append(max_p)
r_list.append(max_r)
f1_list.append(max_f1)
print("running time: ", str((end_time - start_time) / 60.))
print_training_info()
p, r, f1 = map(lambda x: np.array(x).mean(), [p_list, r_list, f1_list])
print("f1_score in 10 fold: {}\naverage : {} {} {}\n".format(
np.array(f1_list).reshape(-1, 1), round(p, 4), round(r, 4),
round(f1, 4)))
return p, r, f1