def train():
train_data, train_label, word2id, word_embedding, max_sentence_len = load_all(settings.TRAIN_PATH,settings.VOCAB_PATH,
settings.VOCAB_EMBEDDING_PATH)
# test no embedding
# word_embedding=np.random.uniform(-0.25,0.25,word_embedding.shape)
ner_model = NerModel(word2id, word_embedding, settings.TAGS, max_sentence_len, settings.EMBEDDING_SIZE)
ner_model.train(train_data, train_label, save_path=settings.MODEL_PATH)
#Preprocessing Data
data_train = load_data()
getter = modifying(data_train)
getter.get_next()
tag2id,n_tags,word2id,n_words = getter.indexing()
text_sequences,label_sequences = getter.padding(args.max_len,word2id,tag2id) # making length of all sentences to be equal
train_dataset = tf.data.Dataset.from_tensor_slices((text_sequences, label_sequences)) # converting to tensorflow dataset
train_dataset = train_dataset.shuffle(len(text_sequences)).batch(args.batch_size, drop_remainder=True)
print("hidden_num:{}, vocab_size:{}, label_size:{}".format(args.hidden_num, len(word2id), len(tag2id)))
#######################################################################################################
model = NerModel(hidden_num = args.hidden_num, vocab_size = len(word2id)+1, label_size= len(tag2id), embedding_size = args.embedding_size)
optimizer = tf.keras.optimizers.Adam(args.lr)
ckpt = tf.train.Checkpoint(optimizer=optimizer, model=model)
ckpt.restore(tf.train.latest_checkpoint(args.output_dir))
ckpt_manager = tf.train.CheckpointManager(ckpt,args.output_dir,checkpoint_name='model.ckpt',max_to_keep=3)
#########################################################################################################
# @tf.function
def train_one_step(text_batch, labels_batch):
with tf.GradientTape() as tape:
logits, text_lens, log_likelihood = model(text_batch, labels_batch,training=True)
loss = - tf.reduce_mean(log_likelihood)
gradients = tape.gradient(loss, model.trainable_variables)
import tensorflow as tf
from model import NerModel
from utils import tokenize,read_vocab,format_result
import tensorflow_addons as tf_ad
from args_help import args
import json
vocab2id, id2vocab = read_vocab(args.vocab_file)
tag2id, id2tag = read_vocab(args.tag_file)
text_sequences ,label_sequences= tokenize(args.test_path,vocab2id,tag2id)
optimizer = tf.keras.optimizers.Adam(args.lr)
model = NerModel(hidden_num = args.hidden_num, vocab_size =len(vocab2id), label_size = len(tag2id), embedding_size = args.embedding_size)
# restore model
ckpt = tf.train.Checkpoint(optimizer=optimizer,model=model)
ckpt.restore(tf.train.latest_checkpoint(args.output_dir))
while True:
text = input("input:")
dataset = tf.keras.preprocessing.sequence.pad_sequences([[vocab2id.get(char,0) for char in text]], padding='post')
print(dataset)
logits, text_lens = model.predict(dataset)
paths = []
for logit, text_len in zip(logits, text_lens):
viterbi_path, _ = tf_ad.text.viterbi_decode(logit[:text_len], model.transition_params)
paths.append(viterbi_path)
print(paths[0])
parser.add_argument("--embedding_size", type=int, default=32,help="embedding dim")
parser.add_argument("--output_dir", type=str, default='./checkpoint',help="output dir")
parser.add_argument("--lr", type=float, default=1e-3,help="lr")
parser.add_argument("--batch_size", type=int, default=64,help="lr")
args = parser.parse_args()
gpus=tf.config.experimental.list_physical_devices(device_type='GPU')
tf.config.experimental.set_visible_devices(devices=gpus[2], device_type='GPU')
vocab2id, id2vocab = read_vocab(args.vocab_file)
tag2id, id2tag = read_vocab(args.tag_file)
text_sequences, text_lens ,label_sequences= tokenize_pred(args.test_file,vocab2id,tag2id)
train_dataset = tf.data.Dataset.from_tensor_slices((text_sequences, text_lens, label_sequences))
train_dataset = train_dataset.shuffle(len(text_sequences)).batch(args.batch_size, drop_remainder=True)
optimizer = tf.keras.optimizers.Adam(args.lr)
model = NerModel(hidden_num = args.hidden_num, vocab_size =len(vocab2id), label_size = len(tag2id), embedding_size = args.embedding_size)
# restore model
ckpt = tf.train.Checkpoint(optimizer=optimizer,model=model)
ckpt.restore(tf.train.latest_checkpoint(args.output_dir))
for text_batch, text_lens,labels_batch in train_dataset:
logits, _ = model.predict(text_batch)
paths = []
for logit, text_len, labels in zip(logits, text_lens, labels_batch):
viterbi_path, _ = tf_ad.text.viterbi_decode(logit[:text_len], model.transition_params)
paths.append(viterbi_path)
for i in range(len(text_batch)):
res = {'text':[],'pred':[],'label':[]}
for j,t in enumerate(paths[i]):
res['text'].append(id2vocab.get(text_batch[i][j].numpy(),'<UKN>'))
tag2id, id2tag = read_vocab(args.tag_file)
print(id2tag)
text_sequences, label_sequences, text_origin, label_origin = tokenize(
args.test_path, vocab2id, tag2id)
# text_sequences 的维度是（159，110）
embedded_matrix = build_embedding_matrix(args.pretrain_embedding_vec, vocab2id)
# print('查看 text_sequences 的值和维度:')
# print(text_sequences.shape)
# print(type(text_sequences))
# 载入模型
optimizer = tf.keras.optimizers.Adam(args.lr)
model = NerModel(hidden_num=args.hidden_num,
vocab_size=len(vocab2id),
label_size=len(tag2id),
embedding_size=args.embedding_size,
embedding_matrix=embedded_matrix)
# restore model
ckpt = tf.train.Checkpoint(optimizer=optimizer, model=model)
ckpt.restore(tf.train.latest_checkpoint(args.output_dir))
def evaluationMetrics(id2tag, logits_batch, labels_batch):
"""
（待加入模型）添加 presicion和 recall 作为测试集的评估方式
logits_batch 表示预测值（单位为batch）
labels_batch 表示真实值（单位为batch）
"""
entity = [] # 真实的实体
pre_entity = [] # 预测的实体
# 获取训练数据
train_dataset = tf.data.Dataset.from_tensor_slices((text_sequences, label_sequences))
train_dataset = train_dataset.shuffle(len(text_sequences)).batch(args.batch_size, drop_remainder=True) # Tensor序列中最后少于一个batch数量的不要了
# # 测试代码，用后即删
# for _, (text_batch, labels_batch) in enumerate(train_dataset):
# print(type(text_batch))
# print(text_batch.shape)
# print(text_batch)
# break
# 构建模型
logger.info("hidden_num:{}, vocab_size:{}, label_size:{}".format(args.hidden_num, len(vocab2id), len(tag2id)))
model = NerModel(hidden_num = args.hidden_num, vocab_size = len(vocab2id), label_size= len(tag2id),
embedding_size = args.embedding_size, embedding_matrix=embedded_matrix)
optimizer = tf.keras.optimizers.Adam(args.lr)
ckpt = tf.train.Checkpoint(optimizer=optimizer, model=model)
ckpt.restore(tf.train.latest_checkpoint(args.output_dir))
ckpt_manager = tf.train.CheckpointManager(ckpt,
args.output_dir,
checkpoint_name='model.ckpt',
max_to_keep=3)
# @tf.function
def train_one_step(text_batch, labels_batch):
with tf.GradientTape() as tape:
logits, text_lens, log_likelihood = model(text_batch, labels_batch, training=True) # 此时调用model的call方法
# log_likelihood 即表示loss结果
def get_model():
global __model
if not __model:
__model = NerModel()
return __model
#!/usr/bin/env python # encoding: utf-8 ''' @author: Ben @license: (C) Copyright 2013-2017, Node Supply Chain Manager Corporation Limited. @contact: [email protected] @file: keras_run.py @time: 2019/8/15 09:42 @desc: ''' from model import NerModel from utils import * if __name__ == '__main__': log.i('Start main function.') model = NerModel() model.train() if is_train() else model.predict() log.i('Process finish')
logger.info("vocab file exits!!")
vocab2id, id2vocab = read_vocab(args.vocab_file)
tag2id, id2tag = read_vocab(args.tag_file)
text_sequences, label_sequences = tokenize(args.train_path, vocab2id, tag2id)
train_dataset = tf.data.Dataset.from_tensor_slices(
(text_sequences, label_sequences))
train_dataset = train_dataset.shuffle(len(text_sequences)).batch(
args.batch_size, drop_remainder=True)
logger.info("hidden_num:{}, vocab_size:{}, label_size:{}".format(
args.hidden_num, len(vocab2id), len(tag2id)))
model = NerModel(hidden_num=args.hidden_num,
vocab_size=len(vocab2id),
label_size=len(tag2id),
embedding_size=args.embedding_size)
model.compile(optimizer=tf.keras.optimizers.Adam(args.lr),
# loss=
)
ckpt = tf.train.Checkpoint(optimizer=model.optimizer, model=model)
ckpt.restore(tf.train.latest_checkpoint(args.output_dir))
ckpt_manager = tf.train.CheckpointManager(ckpt,
args.output_dir,
checkpoint_name='model.ckpt',
max_to_keep=3)
model.fit(train_dataset, batch_size=args.batch_size, epoch=args.epoch)
# @tf.function
