# -*- coding: utf-8 -*- # @Date : 2020/7/15 # @Author : mingming.xu # @Email : [email protected] # @File : extract_feature.py from toolkit4nlp.models import build_transformer_model from toolkit4nlp.tokenizers import Tokenizer import numpy as np config = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_config.json' ckpt = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_model.ckpt' vocab = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/vocab.txt' tokenizer = Tokenizer(vocab, do_lower_case=True) model = build_transformer_model(config, checkpoint_path=ckpt) token_ids, segment_ids = tokenizer.encode(u'我爱你中国') print('\n ===== predicting =====\n') print(model.predict([np.array([token_ids]), np.array([segment_ids])])) '''[[[-0.00827767 0.52711666 -0.2616654 ... 0.7717162 0.6682844 -0.3481327 ] [ 0.3665638 0.35970846 0.0772187 ... -0.5211092 -0.46724823 0.07845997] [ 0.6985213 -0.04391993 -1.3160559 ... 1.061864 0.8293197 0.07258661] ... [ 0.25169933 0.3048255 -1.2513847 ... 0.5438095 0.46753633 -0.61883307] [ 0.07904327 -0.08373377 -0.3963912 ... 0.29524678 0.74877214
for i, l in enumerate(f):
l = json.loads(l)
text, label, label_des = l['sentence'], l['label'], l['label_desc']
label = int(label) - 100 if int(label) < 105 else int(label) - 101
D.append((text, int(label), label_des))
return D
# 加载数据集
train_data = load_data(
'/home/mingming.xu/datasets/NLP/CLUE/tnews_public/train.json')
valid_data = load_data(
'/home/mingming.xu/datasets/NLP/CLUE/tnews_public/dev.json')
# tokenizer
tokenizer = Tokenizer(dict_path, do_lower_case=True)
all_data = train_data + valid_data
pretrain_data = [d[0] for d in all_data]
# whole word mask
pretrain_data = [jieba.lcut(d) for d in pretrain_data]
def random_masking(lines):
"""对输入进行随机mask, 支持多行
"""
if type(lines[0]) != list:
lines = [lines]
sources, targets = [tokenizer._token_start_id], [0]
"""
GPT闲聊demo，参考：https://github.com/thu-coai/CDial-GPT
"""
import numpy as np
from toolkit4nlp.models import build_transformer_model
from toolkit4nlp.tokenizers import Tokenizer
from toolkit4nlp.utils import AutoRegressiveDecoder
config_path = 'D:/pretrain/GPT_LCCC-base-tf/gpt_config.json'
checkpoint_path = 'D:/pretrain/GPT_LCCC-base-tf/gpt_model.ckpt'
dict_path = 'D:/pretrain/GPT_LCCC-base-tf/vocab.txt'
tokenizer = Tokenizer(dict_path, do_lower_case=True)
speakers = [
tokenizer.token_to_id('[speaker1]'),
tokenizer.token_to_id('[speaker2]')
]
model = build_transformer_model(
config_path=config_path,
checkpoint_path=checkpoint_path,
model='gpt',
)
model.summary()
class ChatBot(AutoRegressiveDecoder):
"""
随机采样生成对话
"""
D.append((text, int(label), label_des))
return D
# 加载数据集
train_data = load_data(
'/home/mingming.xu/datasets/NLP/CLUE/iflytek/train.json')
valid_data = load_data('/home/mingming.xu/datasets/NLP/CLUE/iflytek/dev.json')
# 加载并精简词表，建立分词器
token_dict, keep_tokens = load_vocab(
vocab_path=dict_path,
simplified=True,
startswith=['[PAD]', '[UNK]', '[CLS]', '[SEP]'],
)
tokenizer = Tokenizer(token_dict, do_lower_case=True)
class data_generator(DataGenerator):
"""迁移时由于没有额外的label data，所以通过data augmentation 来模拟。
方法是切分句子后重复后shuffle再重新组成新的句子
"""
def __init__(self, data_augmentation=False, transfer=False, **kwargs):
super(data_generator, self).__init__(**kwargs)
self.data_augmentation = data_augmentation
self.transfer = transfer
def __iter__(self, random=False):
batch_token_ids, batch_segment_ids, batch_labels = [], [], []
for is_end, (text, label, label_des) in self.get_sample():
if self.data_augmentation:
test_result_path = 'test_result.txt'
test_score_path = 'test_score.txt'
bert_config = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_config.json'
bert_dict = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/vocab.txt'
bert_checkpoint = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_model.ckpt'
maxlen = 256
bert_layers = 12
lr_multi = 2 # crf loss 放大
num_labes = 4
lr = 1e-5
batch_size = 16 # 实际梯度在此基础上累加2次，即 batch_size = 32
epochs = 5
tokenizer = Tokenizer(bert_dict, do_lower_case=True)
def load_data(data_path):
items = []
with open(data_path) as f:
for line in f:
chunk = re.split(' +', line.strip())
items.append(chunk)
return items
data = load_data(train_path)
row_nums = list(range(len(data)))
np.random.shuffle(row_nums)
from toolkit4nlp.tokenizers import Tokenizer
from toolkit4nlp.utils import ViterbiDecoder, pad_sequences, DataGenerator
from toolkit4nlp.optimizers import Adam
from toolkit4nlp.layers import *
from toolkit4nlp.backend import K, sequence_masking
vocab_dict = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/vocab.txt'
config_path = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_config.json'
checkpoint_path = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_model.ckpt'
data_dir = '/home/mingming.xu/datasets/NLP/ner/china-people-daily-ner-corpus/'
train_path = os.path.join(data_dir, 'example.train')
test_path = os.path.join(data_dir, 'example.test')
val_path = os.path.join(data_dir, 'example.dev')
tokenizer = Tokenizer(vocab_dict, do_lower_case=True)
maxlen = 256
lr = 1e-5
epochs = 5
batch_size = 16
def load_data(filename):
D = []
with open(filename, encoding='utf-8') as f:
f = f.read()
for l in f.split('\n\n'):
if not l:
continue
d, last_flag = [], ''
return x, y
return TrainingDataset.load_tfrecord(record_names, batch_size, parse_func)
if __name__ == "__main__":
from toolkit4nlp.tokenizers import Tokenizer
import re
import json
import glob
import jieba_fast as jieba
from tqdm import tqdm
jieba.initialize()
vocab = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/vocab.txt'
tokenizer = Tokenizer(vocab, do_lower_case=True)
seq_length = 512
def word_seg(text):
return jieba.lcut(text)
def generate_corp():
file_names = glob.glob('/home/mingming.xu/datasets/NLP/qa/dureader_robust-data/pretraining/*')
count, sentences = 0, []
for fname in file_names:
with open(fname) as fin:
for p in json.load(fin)['data'][0]['paragraphs']:
para = [qa['question'] for qa in p['qas']]
# -*- coding: utf-8 -*- # @Date : 2020/7/16 # @Author : mingming.xu # @Email : [email protected] # @File : mask_language_model.py import numpy as np from toolkit4nlp.tokenizers import Tokenizer from toolkit4nlp.models import build_transformer_model config = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_config.json' ckpt = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/bert_model.ckpt' vocab = '/home/mingming.xu/pretrain/NLP/chinese_L-12_H-768_A-12/vocab.txt' tokenizer = Tokenizer(vocab, do_lower_case=True) model = build_transformer_model(config, checkpoint_path=ckpt, with_mlm=True) # tokens, segs = tokenizer.encode('北京网聘技术有限公司') tokens, segs = tokenizer.encode('科学技术是第一生产力') tokens[3] = tokens[4] = tokenizer._token_dict['[MASK]'] prob = model.predict([np.array([tokens]), np.array([segs])])[0] print(tokenizer.decode(np.argmax(prob[3:5], axis=1))) ''' 正确结果应该是： 技术 '''
train_data = load_data('train')
test_data = load_data('test')
num_classes = 32
maxlen = 128
batch_size = 8
# BERT base
config_path = '/home/mingming.xu/pretrain/NLP/nezha_base_wwm/bert_config.json'
checkpoint_path = '/home/mingming.xu/pretrain/NLP/nezha_base_wwm/model.ckpt'
dict_path = '/home/mingming.xu/pretrain/NLP/nezha_base_wwm/vocab.txt'
# tokenizer
tokenizer = Tokenizer(dict_path, do_lower_case=True)
# pattern
pattern = '直接回答问题:'
mask_idx = [1]
id2label = {0: '间', 1: '直'}
label2id = {v: k for k, v in id2label.items()}
labels = list(id2label.values())
def random_masking(token_ids):
"""对输入进行随机mask
"""
rands = np.random.random(len(token_ids))