前言
仅记录学习过程,有问题欢迎讨论
* 1. 输入问题
* 2. 匹配问题库(基础资源,FAQ)
* 3. 返回答案
文本匹配算法:
编辑距离算法(缺点)
- 字符之间没有语义相似度;
受无关词/停用词影响大;
受语序影响大
- 字符之间没有语义相似度;
Jaccard相似度(元素的交集/元素的并集)、
词向量(基于窗口;解决了语义相似的问题;文本转为数字,计算cos值来判断相似度)
深度学习-表示型(问题匹配问题比较合适,因为二者都是问题,所以转向量也方便)
两个话使用同一个Encoder向量 语义相似的score = 1,类似二分类- (Triplet Loss):
使得相同标签的样本再Embedding空间尽量接近(anchor和positive接近 away negative) - loss = max((D(p,a)-D(p,n)+margin,0)
- 优点:训练好的模型可以对知识库内的问题计算向量,在实际查找过程中,只对输入文本做一次向量化
- 缺点:在向量化的过程中不知道文本重点
- (Triplet Loss):
深度学习-交互型
- 输入一句话,但是两个样本拼接,利用attention机制来判断是否匹配(Q&A拼接去学习)
- 优点:通过对比把握句子重点
- 缺点:每次计算需要都需要两个输入
对比学习
- 输入一个样本,通过函数把样本改动,但还是相似,得到两个相似样本,进行bertEconder,pooling操作
海量向量查找:
- 可以用开源写好的库^^ Faiss Pinecore
- 避免遍历,避免和所有向量做距离计算(空间切割KD树,Kmeans方式切割)
代码
实现一个智能问答demo
"""
配置参数信息
"""
Config = {
"model_path": "./output/",
"model_name": "model.pt",
"schema_path": r"D:\NLP\video\第八周\week8 文本匹配问题\data\schema.json",
"train_data_path": r"D:\NLP\video\第八周\week8 文本匹配问题\data\data.json",
"valid_data_path": r"D:\NLP\video\第八周\week8 文本匹配问题\data\valid.json",
"vocab_path": r"D:\NLP\video\第七周\data\vocab.txt",
"model_type": "rnn",
# 正样本比例
"positive_sample_rate": 0.5,
"use_bert": False,
# 文本向量大小
"char_dim": 32,
# 文本长度
"max_len": 20,
# 词向量大小
"hidden_size": 128,
# 训练 轮数
"epoch_size": 15,
# 批量大小
"batch_size": 32,
# 训练集大小
"simple_size": 300,
# 学习率
"lr": 1e-3,
# dropout
"dropout": 0.5,
# 优化器
"optimizer": "adam",
# 卷积核
"kernel_size": 3,
# 最大池 or 平均池
"pooling_style": "max",
# 模型层数
"num_layers": 2,
"bert_model_path": r"D:\NLP\video\第六周\bert-base-chinese",
# 输出层大小
"output_size": 2,
# 随机数种子
"seed": 987
}
load.py j加载数据文件
"""
数据加载
"""
import json
from collections import defaultdict
import random
import torch
import torch.utils.data as Data
from torch.utils.data import DataLoader
from transformers import BertTokenizer
# 获取字表集
def load_vocab(path):
vocab = {}
with open(path, 'r', encoding='utf-8') as f:
for index, line in enumerate(f):
word = line.strip()
# 0留给padding位置,所以从1开始
vocab[word] = index + 1
vocab['unk'] = len(vocab) + 1
return vocab
# 数据预处理 裁剪or填充
def padding(input_ids, length):
if len(input_ids) >= length:
return input_ids[:length]
else:
padded_input_ids = input_ids + [0] * (length - len(input_ids))
return padded_input_ids
# 文本预处理
# 转化为向量
def sentence_to_index(text, length, vocab):
input_ids = []
for char in text:
input_ids.append(vocab.get(char, vocab['unk']))
# 填充or裁剪
input_ids = padding(input_ids, length)
return input_ids
class DataGenerator:
def __init__(self, data_path, config):
# 加载json数据
self.load_know_base(config["train_data_path"])
# 加载schema 相当于答案集
self.schema = self.load_schema(config["schema_path"])
self.data_path = data_path
self.config = config
if self.config["model_type"] == "bert":
self.tokenizer = BertTokenizer.from_pretrained(config["bert_model_path"])
self.vocab = load_vocab(config["vocab_path"])
self.config["vocab_size"] = len(self.vocab)
self.train_flag = None
self.load_data()
def __len__(self):
if self.train_flag:
return self.config["simple_size"]
else:
return len(self.data)
# 这里需要返回随机的样本
def __getitem__(self, idx):
if self.train_flag:
# return self.random_train_sample() # 随机生成一个训练样本
# triplet loss:
return self.random_train_sample_for_triplet_loss()
else:
return self.data[idx]
# 针对获取的文本 load_know_base = {target : [questions]} 做处理
# 传入两个样本 正样本为相同target数据 负样本为不同target数据
# 训练集和验证集不一致
def load_data(self):
self.train_flag = self.config["train_flag"]
dataset_x = []
dataset_y = []
self.knwb = defaultdict(list)
if self.train_flag:
for target, questions in self.target_to_questions.items():
for question in questions:
input_id = sentence_to_index(question, self.config["max_len"], self.vocab)
input_id = torch.LongTensor(input_id)
# self.schema[target] 下标 把每个question转化为向量append放入一个target下
self.knwb[self.schema[target]].append(input_id)
else:
with open(self.data_path, encoding="utf8") as f:
for line in f:
line = json.loads(line)
assert isinstance(line, list)
question, target = line
input_id = sentence_to_index(question, self.config["max_len"], self.vocab)
# input_id = torch.LongTensor(input_id)
label_index = torch.LongTensor([self.schema[target]])
# self.data.append([input_id, label_index])
dataset_x.append(input_id)
dataset_y.append(label_index)
self.data = Data.TensorDataset(torch.tensor(dataset_x), torch.tensor(dataset_y))
return
# 加载知识库
def load_know_base(self, know_base_path):
self.target_to_questions = {}
with open(know_base_path, encoding="utf8") as f:
for index, line in enumerate(f):
content = json.loads(line)
questions = content["questions"]
target = content["target"]
self.target_to_questions[target] = questions
return
# 加载schema 相当于答案集
def load_schema(self, param):
with open(param, encoding="utf8") as f:
return json.loads(f.read())
# 训练集随机生成一个样本
# 依照一定概率生成负样本或正样本
# 负样本从随机两个不同的标准问题中各随机选取一个
# 正样本从随机一个标准问题中随机选取两个
def random_train_sample(self):
target = random.choice(list(self.knwb.keys()))
# 随机正样本:
# 随机正样本
if random.random() <= self.config["positive_sample_rate"]:
if len(self.knwb[target]) <= 1:
return self.random_train_sample()
else:
question1 = random.choice(self.knwb[target])
question2 = random.choice(self.knwb[target])
# 一组
# dataset_x.append([question1, question2])
# # 二分类任务 同一组的question target = 1
# dataset_y.append([1])
return [question1, question2, torch.LongTensor([1])]
else:
# 随机负样本:
p, n = random.sample(list(self.knwb.keys()), 2)
question1 = random.choice(self.knwb[p])
question2 = random.choice(self.knwb[n])
# dataset_x.append([question1, question2])
# dataset_y.append([-1])
return [question1, question2, torch.LongTensor([-1])]
# triplet_loss随机生成3个样本 锚样本A, 正样本P, 负样本N
def random_train_sample_for_triplet_loss(self):
target = random.choice(list(self.knwb.keys()))
# question1锚样本 question2为同一个target下的正样本 question3 为其他target下样本
question1 = random.choice(self.knwb[target])
question2 = random.choice(self.knwb[target])
question3 = random.choice(self.knwb[random.choice(list(self.knwb.keys()))])
return [question1, question2, question3]
# 用torch自带的DataLoader类封装数据
def load_data_batch(data_path, config, shuffle=True):
dg = DataGenerator(data_path, config)
if config["train_flag"]:
dl = DataLoader(dg, batch_size=config["batch_size"], shuffle=shuffle)
else:
dl = DataLoader(dg.data, batch_size=config["batch_size"], shuffle=shuffle)
return dl
if __name__ == '__main__':
from config import Config
Config["train_flag"] = True
# dg = DataGenerator(Config["train_data_path"], Config)
dataset = load_data_batch(Config["train_data_path"], Config)
# print(len(dg))
# print(dg[0])
for index, dataset in enumerate(dataset):
input_id1, input_id2, input_id3 = dataset
print(input_id1)
print(input_id2)
print(input_id3)
main.py 主方法
import torch
import os
import random
import os
import numpy as np
import logging
from config import Config
from model import TorchModel, choose_optimizer, SiameseNetwork
from loader import load_data_batch
from evaluate import Evaluator
# [DEBUG, INFO, WARNING, ERROR, CRITICAL]
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
"""
模型训练主程序
"""
# 通过设置随机种子来复现上一次的结果(避免随机性)
seed = Config["seed"]
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def main(config):
# 保存模型的目录
if not os.path.isdir(config["model_path"]):
os.mkdir(config["model_path"])
# 加载数据
dataset = load_data_batch(config["train_data_path"], config)
# 加载模型
model = SiameseNetwork(config)
# 是否使用gpu
if torch.cuda.is_available():
logger.info("gpu可以使用,迁移模型至gpu")
model.cuda()
# 选择优化器
optim = choose_optimizer(config, model)
# 加载效果测试类
evaluator = Evaluator(config, model, logger)
for epoch in range(config["epoch_size"]):
epoch += 1
logger.info("epoch %d begin" % epoch)
epoch_loss = []
# 训练模型
model.train()
for batch_data in dataset:
if torch.cuda.is_available():
batch_data = [d.cuda() for d in batch_data]
# x, y = dataiter
# 反向传播
optim.zero_grad()
s1, s2, s3 = batch_data # 输入变化时这里需要修改,比如多输入,多输出的情况
# 计算梯度
loss = model(s1, s2, s3)
# 梯度更新
loss.backward()
# 优化器更新模型
optim.step()
# 记录损失
epoch_loss.append(loss.item())
logger.info("epoch average loss: %f" % np.mean(epoch_loss))
# 测试模型效果
acc = evaluator.eval(epoch)
# 可以用model_type model_path epoch 三个参数来保存模型
model_path = os.path.join(config["model_path"], "epoch_%d_%s.pth" % (epoch, config["model_type"]))
torch.save(model.state_dict(), model_path) # 保存模型权重
return
if __name__ == "__main__":
from config import Config
Config["train_flag"] = True
main(Config)
# for model in ["cnn"]:
# Config["model_type"] = model
# print("最后一轮准确率:", main(Config), "当前配置:", Config["model_type"])
# 对比所有模型
# 中间日志可以关掉,避免输出过多信息
# 超参数的网格搜索
# for model in ["gated_cnn"]:
# Config["model_type"] = model
# for lr in [1e-3, 1e-4]:
# Config["learning_rate"] = lr
# for hidden_size in [128]:
# Config["hidden_size"] = hidden_size
# for batch_size in [64, 128]:
# Config["batch_size"] = batch_size
# for pooling_style in ["avg"]:
# Config["pooling_style"] = pooling_style
# 可以把输出放入文件中 便于查看
# print("最后一轮准确率:", main(Config), "当前配置:", Config)
evaluate.py 评估模型文件
"""
模型效果测试
"""
import torch
from loader import load_data_batch
class Evaluator:
def __init__(self, config, model, logger):
self.config = config
self.model = model
self.logger = logger
# 选择验证集合
config['train_flag'] = False
self.valid_data = load_data_batch(config["valid_data_path"], config, shuffle=False)
config['train_flag'] = True
self.train_data = load_data_batch(config["train_data_path"], config)
self.stats_dict = {"correct": 0, "wrong": 0} # 用于存储测试结果
def eval(self, epoch):
self.logger.info("开始测试第%d轮模型效果:" % epoch)
self.stats_dict = {"correct": 0, "wrong": 0} # 清空前一轮的测试结果
self.model.eval()
self.knwb_to_vector()
for index, batch_data in enumerate(self.valid_data):
if torch.cuda.is_available():
batch_data = [d.cuda() for d in batch_data]
input_id, labels = batch_data # 输入变化时这里需要修改,比如多输入,多输出的情况
with torch.no_grad():
test_question_vectors = self.model(input_id) # 不输入labels,使用模型当前参数进行预测
self.write_stats(test_question_vectors, labels)
self.show_stats()
return
def write_stats(self, test_question_vectors, labels):
assert len(labels) == len(test_question_vectors)
for test_question_vector, label in zip(test_question_vectors, labels):
# 通过一次矩阵乘法,计算输入问题和知识库中所有问题的相似度
# test_question_vector shape [vec_size] knwb_vectors shape = [n, vec_size]
res = torch.mm(test_question_vector.unsqueeze(0), self.knwb_vectors.T)
hit_index = int(torch.argmax(res.squeeze())) # 命中问题标号
hit_index = self.question_index_to_standard_question_index[hit_index] # 转化成标准问编号
if int(hit_index) == int(label):
self.stats_dict["correct"] += 1
else:
self.stats_dict["wrong"] += 1
return
# 将知识库中的问题向量化,为匹配做准备
# 每轮训练的模型参数不一样,生成的向量也不一样,所以需要每轮测试都重新进行向量化
def knwb_to_vector(self):
self.question_index_to_standard_question_index = {}
self.question_ids = []
for standard_question_index, question_ids in self.train_data.dataset.knwb.items():
for question_id in question_ids:
# 记录问题编号到标准问题标号的映射,用来确认答案是否正确
self.question_index_to_standard_question_index[len(self.question_ids)] = standard_question_index
self.question_ids.append(question_id)
with torch.no_grad():
question_matrixs = torch.stack(self.question_ids, dim=0)
if torch.cuda.is_available():
question_matrixs = question_matrixs.cuda()
self.knwb_vectors = self.model(question_matrixs)
# 将所有向量都作归一化 v / |v|
self.knwb_vectors = torch.nn.functional.normalize(self.knwb_vectors, dim=-1)
return
def show_stats(self):
correct = self.stats_dict["correct"]
wrong = self.stats_dict["wrong"]
self.logger.info("预测集合条目总量:%d" % (correct + wrong))
self.logger.info("预测正确条目:%d,预测错误条目:%d" % (correct, wrong))
self.logger.info("预测准确率:%f" % (correct / (correct + wrong)))
self.logger.info("--------------------")
return correct / (correct + wrong)
model.py
import torch
import torch.nn as nn
from torch.optim import Adam, SGD
from transformers import BertModel
"""
建立网络模型结构
"""
class TorchModel(nn.Module):
def __init__(self, config):
super(TorchModel, self).__init__()
hidden_size = config["hidden_size"]
vocab_size = config["vocab_size"] + 1
output_size = config["output_size"]
model_type = config["model_type"]
num_layers = config["num_layers"]
self.use_bert = config["use_bert"]
self.emb = nn.Embedding(vocab_size + 1, hidden_size, padding_idx=0)
if model_type == 'rnn':
self.encoder = nn.RNN(input_size=hidden_size, hidden_size=hidden_size, num_layers=num_layers,
batch_first=True)
elif model_type == 'lstm':
# 双向lstm,输出的是 hidden_size * 2(num_layers 要写2)
self.encoder = nn.LSTM(hidden_size, hidden_size, num_layers=num_layers)
elif self.use_bert:
self.encoder = BertModel.from_pretrained(config["bert_model_path"])
# 需要使用预训练模型的hidden_size
hidden_size = self.encoder.config.hidden_size
elif model_type == 'cnn':
self.encoder = CNN(config)
elif model_type == "gated_cnn":
self.encoder = GatedCNN(config)
elif model_type == "bert_lstm":
self.encoder = BertLSTM(config)
# 需要使用预训练模型的hidden_size
hidden_size = self.encoder.config.hidden_size
self.classify = nn.Linear(hidden_size, output_size)
self.pooling_style = config["pooling_style"]
self.loss = nn.functional.cross_entropy # loss采用交叉熵损失
def forward(self, x, y=None):
if self.use_bert:
# 输入x为[batch_size, seq_len]
# bert返回的结果是 (sequence_output, pooler_output)
# sequence_output:batch_size, max_len, hidden_size
# pooler_output:batch_size, hidden_size
x = self.encoder(x)[0]
else:
x = self.emb(x)
x = self.encoder(x)
# 判断x是否是tuple
if isinstance(x, tuple):
x = x[0]
# 池化层
if self.pooling_style == "max":
# shape[1]代表列数,shape是行和列数构成的元组
self.pooling_style = nn.MaxPool1d(x.shape[1])
elif self.pooling_style == "avg":
self.pooling_style = nn.AvgPool1d(x.shape[1])
x = self.pooling_style(x.transpose(1, 2)).squeeze()
y_pred = self.classify(x)
if y is not None:
return self.loss(y_pred, y.squeeze())
else:
return y_pred
# 定义孪生网络 (计算两个句子之间的相似度)
class SiameseNetwork(nn.Module):
def __init__(self, config):
super(SiameseNetwork, self).__init__()
self.sentence_encoder = TorchModel(config)
# 使用的是cos计算
# self.loss = nn.CosineEmbeddingLoss()
# 使用triplet_loss
self.triplet_loss = self.cosine_triplet_loss
# 计算余弦距离 1-cos(a,b)
# cos=1时两个向量相同,余弦距离为0;cos=0时,两个向量正交,余弦距离为1
def cosine_distance(self, tensor1, tensor2):
tensor1 = torch.nn.functional.normalize(tensor1, dim=-1)
tensor2 = torch.nn.functional.normalize(tensor2, dim=-1)
cosine = torch.sum(torch.mul(tensor1, tensor2), axis=-1)
return 1 - cosine
# 3个样本 2个为一类 另一个一类 计算triplet loss
def cosine_triplet_loss(self, a, p, n, margin=None):
ap = self.cosine_distance(a, p)
an = self.cosine_distance(a, n)
if margin is None:
diff = ap - an + 0.1
else:
diff = ap - an + margin.squeeze()
return torch.mean(diff[diff.gt(0)]) # greater than
# 使用triplet_loss
def forward(self, sentence1, sentence2=None, sentence3=None, margin=None):
vector1 = self.sentence_encoder(sentence1)
# 同时传入3 个样本
if sentence2 is None:
if sentence3 is None:
return vector1
# 计算余弦距离
else:
vector3 = self.sentence_encoder(sentence3)
return self.cosine_distance(vector1, vector3)
else:
vector2 = self.sentence_encoder(sentence2)
if sentence3 is None:
return self.cosine_distance(vector1, vector2)
else:
vector3 = self.sentence_encoder(sentence3)
return self.triplet_loss(vector1, vector2, vector3, margin)
# CosineEmbeddingLoss
# def forward(self,sentence1, sentence2=None, target=None):
# # 同时传入两个句子
# if sentence2 is not None:
# vector1 = self.sentence_encoder(sentence1) # vec:(batch_size, hidden_size)
# vector2 = self.sentence_encoder(sentence2)
# # 如果有标签,则计算loss
# if target is not None:
# return self.loss(vector1, vector2, target.squeeze())
# # 如果无标签,计算余弦距离
# else:
# return self.cosine_distance(vector1, vector2)
# # 单独传入一个句子时,认为正在使用向量化能力
# else:
# return self.sentence_encoder(sentence1)
# 优化器的选择
def choose_optimizer(config, model):
optimizer = config["optimizer"]
learning_rate = config["lr"]
if optimizer == "adam":
return Adam(model.parameters(), lr=learning_rate)
elif optimizer == "sgd":
return SGD(model.parameters(), lr=learning_rate)
# 定义CNN模型
class CNN(nn.Module):
def __init__(self, config):
super(CNN, self).__init__()
hidden_size = config["hidden_size"]
kernel_size = config["kernel_size"]
pad = int((kernel_size - 1) / 2)
self.cnn = nn.Conv1d(hidden_size, hidden_size, kernel_size, bias=False, padding=pad)
def forward(self, x): # x : (batch_size, max_len, embeding_size)
return self.cnn(x.transpose(1, 2)).transpose(1, 2)
# 定义GatedCNN模型
class GatedCNN(nn.Module):
def __init__(self, config):
super(GatedCNN, self).__init__()
self.cnn = CNN(config)
self.gate = CNN(config)
# 定义前向传播函数 比普通cnn多了一次sigmoid 然后互相卷积
def forward(self, x):
a = self.cnn(x)
b = self.gate(x)
b = torch.sigmoid(b)
return torch.mul(a, b)
# 定义BERT-LSTM模型
class BertLSTM(nn.Module):
def __init__(self, config):
super(BertLSTM, self).__init__()
self.bert = BertModel.from_pretrained(config["bert_model_path"], return_dict=False)
self.rnn = nn.LSTM(self.bert.config.hidden_size, self.bert.config.hidden_size, batch_first=True)
def forward(self, x):
x = self.bert(x)[0]
x, _ = self.rnn(x)
return x
if __name__ == "__main__":
from config import Config
Config["vocab_size"] = 10
Config["max_length"] = 4
model = SiameseNetwork(Config)
s1 = torch.LongTensor([[1, 2, 3, 0], [2, 2, 0, 0]])
s2 = torch.LongTensor([[1, 2, 3, 4], [3, 2, 3, 4]])
l = torch.LongTensor([[1], [0]])
y = model(s1, s2, l)
print(y)