def train_requestable(data_tmp_path):
"""
用于训练模型，先训练完存好了才能用
训练用 early stopping
:param data_tmp_path: data tmp 文件夹位置
"""
print('载入数据管理器...')
data_manager = DataManager(data_tmp_path)
print('载入训练数据...')
informable_slot_datasets, requestable_slot_datasets = generate_dataset(
data_manager.DialogData)
print('载入 requestable slot detector...')
init_learning_rate = 0.005
graph = tf.Graph()
with graph.as_default():
requestable_slots_models = {}
for k, v in All_requestable_slots_order.items():
requestable_slots_models[k] = RequestableSlotDector(
str(v), learning_rate=init_learning_rate)
with tf.Session(graph=graph,
config=tf.ConfigProto(allow_soft_placement=True)) as sess:
sess.run(tf.group(tf.global_variables_initializer()))
saver = tf.train.Saver()
# saver.restore(sess, "./ckpt/requestable/model.ckpt")
# 训练 requestable slots
requestable_slots_F1s = {}
for slot, model in requestable_slots_models.items():
average_loss = 0
learning_rate = init_learning_rate
best_F1 = 0
tolerance = 30
tolerance_count = 0
display_step = 10
for step in range(5000):
step += 1
batch_data, batch_output = requestable_slot_datasets[
slot].next_batch()
char_emb_matrix, word_emb_matrix, seqlen = data_manager.sent2num(
batch_data)
_, training_loss = sess.run(
[model.train_op, model.final_loss],
feed_dict={
model.char_emb_matrix: char_emb_matrix,
model.word_emb_matrix: word_emb_matrix,
model.output: batch_output
})
average_loss += training_loss / display_step
if step % display_step == 0:
batch_data, batch_output = requestable_slot_datasets[
slot].get_testset()
char_emb_matrix, word_emb_matrix, seqlen = data_manager.sent2num(
batch_data)
pred, accu = sess.run(
[model.predict, model.accuracy],
feed_dict={
model.char_emb_matrix: char_emb_matrix,
model.word_emb_matrix: word_emb_matrix,
model.output: batch_output
})
F1 = get_F1score(batch_output, pred.tolist())
if best_F1 < F1:
best_F1 = F1
tolerance_count = 0
if not os.path.exists("./ckpt/requestable/"):
os.makedirs("./ckpt/requestable/")
saver.save(sess, "./ckpt/requestable/model.ckpt")
if tolerance_count == tolerance:
break
print("%s, step % 4d, loss %0.4f, F1 %0.4f, accu %0.4f" %
(slot, step, average_loss, F1, accu))
average_loss = 0
tolerance_count += 1
learning_rate = max(learning_rate * 0.98, 0.001)
sess.run(model.update_lr,
feed_dict={model.new_lr: learning_rate})
print("requestable slot: %s, best F1 %0.4f" % (slot, best_F1))
requestable_slots_F1s[slot] = best_F1
print(requestable_slots_F1s)
print(
sum(requestable_slots_F1s.values()) /
len(requestable_slots_F1s.values()))
def train_informable(data_tmp_path):
"""
用于训练模型，先训练完存好了才能用
训练用 early stopping
:param data_tmp_path: data tmp 文件夹位置
"""
print('载入数据管理器...')
data_manager = DataManager(data_tmp_path)
print('载入训练数据...')
informable_slot_datasets, requestable_slot_datasets = generate_dataset(
data_manager.DialogData)
print('载入 informable slot detector ...')
init_learning_rate = 0.005
informable_batch_ratios = { # 不同slot 的minibatch ratio
"通话时长": [2, 8, 8, 8],
"流量": [4, 8, 8, 8],
"功能费": [4, 8, 8, 8]
}
graph = tf.Graph()
with graph.as_default():
informable_slots_models = {
"功能费": InformableSlotDector('cost',
learning_rate=init_learning_rate),
"流量": InformableSlotDector('data',
learning_rate=init_learning_rate),
"通话时长": InformableSlotDector('time',
learning_rate=init_learning_rate),
}
with tf.Session(graph=graph,
config=tf.ConfigProto(allow_soft_placement=True)) as sess:
sess.run(tf.group(tf.global_variables_initializer()))
saver = tf.train.Saver()
# saver.restore(sess, "./ckpt/informable/model.ckpt")
# 训练 informable slots
informable_slots_accus = []
for slot, model in informable_slots_models.items():
learning_rate = init_learning_rate
average_loss = 0
best_accu = 0
tolerance = 20
tolerance_count = 0
display_step = 10
for step in range(5000):
step += 1
batch_data, batch_output = informable_slot_datasets[
slot].next_batch(informable_batch_ratios[slot])
char_emb_matrix, word_emb_matrix, seqlen = data_manager.sent2num(
batch_data)
_, training_loss = sess.run(
[model.train_op, model.final_loss],
feed_dict={
model.char_emb_matrix: char_emb_matrix,
model.word_emb_matrix: word_emb_matrix,
model.output: batch_output
})
average_loss += training_loss / display_step
if step % display_step == 0:
batch_data, batch_output = informable_slot_datasets[
slot].get_testset()
char_emb_matrix, word_emb_matrix, seqlen = data_manager.sent2num(
batch_data)
pred, accu = sess.run(
[model.predict, model.accuracy],
feed_dict={
model.char_emb_matrix: char_emb_matrix,
model.word_emb_matrix: word_emb_matrix,
model.output: batch_output
})
if best_accu < accu:
best_accu = accu
tolerance_count = 0
if not os.path.exists("./ckpt/informable/"):
os.makedirs("./ckpt/informable/")
saver.save(sess, "./ckpt/informable/model.ckpt")
if tolerance_count == tolerance:
break
print("%s, step % 4d, loss %0.4f, accu %0.4f" %
(slot, step, average_loss, accu))
average_loss = 0
tolerance_count += 1
learning_rate = max(learning_rate * 0.95, 0.0001)
sess.run(model.update_lr,
feed_dict={model.new_lr: learning_rate})
print("informable slot: %s, best accu %0.4f" % (slot, best_accu))
informable_slots_accus.append(best_accu)