作者:十方
bert的大名如雷贯耳,无论在比赛,还是实际上的应用早已普及开来。想到十方第一次跑bert模型用的框架还是paddlepaddle,那时候用自己的训练集跑bert还是比较痛苦的,不仅要看很多配置文件,预处理代码,甚至报错了都不知道怎么回事,当时十方用的是bert双塔做文本向量的语义召回。如今tf都已经更新到了2.4了,tensorflow-hub的出现更是降低了使用预训练模型的门槛,接下来带大家看下,如何花十分钟时间快速构建bert双塔召回模型。
tensorflow hub
打开tensorflow官网,找到tensorflow-hub点进去,我们就能看到各种预训练好的模型了,找到一个预训练好的模型(如下图),下载下来,如介绍所说,这是个12层,768维,12头的模型。
在往下看,我们看到有配套的预处理工具:
同样下载下来,然后我们就可以构建bert双塔了。
Bert双塔
代码语言:javascript复制import os
import shutil
import pickle
import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_text as text
from official.nlp import optimization
from tensorflow.keras import *
from tqdm import tqdm
import numpy as np
import pandas as pd
import json
import re
import random
# 这里读你自己的文本数据集
with open('./data/train_data.pickle', 'rb') as f:
train_data = pickle.load(f)
# 读数据用的generater
def train_generator():
np.random.shuffle(train_data)
for i in range(len(train_data)):
yield train_data[i][0], train_data[i][1]
# 训练数据 dataset
ds_tr = tf.data.Dataset.from_generator(train_generator, output_types=(tf.string, tf.string))
# bert 双塔 dim_size是维度 model_name是下载模型的路径
def get_model(dim_size, model_name):
# 下载的预处理工具路径
preprocessor = hub.load('./bert_en_uncased_preprocess/3')
# 左边塔的文本
text_source = tf.keras.layers.Input(shape=(), dtype=tf.string)
# 右边塔的文本
text_target = tf.keras.layers.Input(shape=(), dtype=tf.string)
tokenize = hub.KerasLayer(preprocessor.tokenize)
tokenized_inputs_source = [tokenize(text_source)]
tokenized_inputs_target = [tokenize(text_target)]
seq_length = 512 # 这里指定你序列文本的最大长度
bert_pack_inputs = hub.KerasLayer(
preprocessor.bert_pack_inputs,
arguments=dict(seq_length=seq_length))
encoder_inputs_source = bert_pack_inputs(tokenized_inputs_source)
encoder_inputs_target = bert_pack_inputs(tokenized_inputs_target)
# 加载预训练参数
bert_model = hub.KerasLayer(model_name)
bert_encoder_source, bert_encoder_target = bert_model(encoder_inputs_source), bert_model(encoder_inputs_target)
# 这里想尝试in-batch loss
# 也可以直接对 bert_encoder_source['pooled_output'], bert_encoder_target['pooled_output'] 做点积操作
matrix_logit = tf.linalg.matmul(bert_encoder_source['pooled_output'], bert_encoder_target['pooled_output'], transpose_a=False, transpose_b=True)
matrix_logit = matrix_logit / tf.sqrt(dim_size)
model = models.Model(inputs = [text_source, text_target], outputs = [bert_encoder_source['pooled_output'], bert_encoder_target['pooled_output'], matrix_logit])
return model
bert_double_tower = get_model(128.0, './small_bert_bert_en_uncased_L-2_H-128_A-2_1/3')
bert_double_tower.summary()
我们看到bert双塔模型已经构建完成:
然后定义loss,就可以训练啦!
代码语言:javascript复制optimizer = optimizers.Adam(learning_rate=5e-5)
loss_func_softmax = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
train_loss = metrics.Mean(name='train_loss')
train_acc = metrics.CategoricalAccuracy(name='train_accuracy')
def train_step(model, features):
with tf.GradientTape() as tape:
p_source, p_target, pred = model(features)
label = tf.eye(tf.shape(pred)[0])
loss = loss_func_softmax(label, pred)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss.update_state(loss)
train_acc.update_state(label, pred)
def train_model(model, bz, epochs):
for epoch in tf.range(epochs):
steps = 0
for feature in ds_tr.prefetch(buffer_size = tf.data.experimental.AUTOTUNE).batch(bz):
logs_s = 'At Epoch={},STEP={}'
tf.print(tf.strings.format(logs_s,(epoch, steps)))
train_step(model, feature)
steps = 1
train_loss.reset_states()
train_acc.reset_states()