Press "Enter" to skip to content

原力计划【NLP】学不会打我 半小时学会基本操作 15 一百行实现 Bert 二分类 (附代码)

【NLP】:warning:学不会打我! 半小时学会基本操作 15:warning: 一百行实现 bert 二分类

 

从今天开始我们将开启一段自然语言处理 (NLP) 的旅程. 自然语言处理可以让来处理, 理解, 以及运用人类的语言, 实现机器语言和人类语言之间的沟通桥梁.

 

 

Bert (Bidirecrional Encoder Representation from Transformers) 是一个预训练的语言表征模型. Bert 主要利用了 Transformer 的 Encoder 结构, 这里就不多赘述.

 

 

简单说一说

 

在大家的鼓励下这届的 CCF 荣获 4 个冠军, 1 个亚军, 在天池中荣获第 4 名. 如图:

 

 

 

 

 

100 行实现 Bert

 

以下代码是全网最简单的 Bert 实现, 部分为比赛源码.

 

网络架构

 

Model: "model"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 780)]        0                                            
__________________________________________________________________________________________________
input_2 (InputLayer)            [(None, 780)]        0                                            
__________________________________________________________________________________________________
tf_bert_model (TFBertModel)     TFBaseModelOutputWit 102267648   input_1[0][0]                    
                                                                 input_2[0][0]                    
__________________________________________________________________________________________________
dense (Dense)                   (None, 2)           7690        tf_bert_model[0][1]              
==================================================================================================
Total params: 102,275,338
Trainable params: 102,275,338
Non-trainable params: 0
__________________________________________________________________________________________________

 

# 超参数
EPOCHS = 50  # 迭代次数
BATCH_SIZE = 8  # 单词训练样本数目
learning_rate = 0.00003  # 学习率
INPUT_DIM = 36782 + 1
MAX_LENGTH = 780
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)  # 优化器
loss = tf.keras.losses.CategoricalCrossentropy()  # 损失
bert_tokenizer = BertTokenizer.from_pretrained('Langboat/mengzi-bert-base')  # Bert的分词器

 

get_data

 

def get_data():
    """
    读取数据
    :return: 返回分批完的训练集和测试集
    """
    # 读取数据
    data_train = pd.read_csv("../data/train.csv")
    print(data_train.head(), data_train.shape)
    data_val = pd.read_csv("../data/val.csv")
    print(data_val.head(), data_val.shape)
    # 预处理
    data_train["label"] = data_train["label"].apply(lambda x: x.split(","))
    print(data_train.head())
    data_val["label"] = data_val["label"].apply(lambda x: x.split(","))
    print(data_val.head())
    # 获取X,y
    X_train = data_train["text"].values.tolist()
    y_train = data_train["label"].values.tolist()
    y_train = np.asarray(y_train, dtype=np.float32)
    X_val = data_val["text"].values.tolist()
    y_val = data_val["label"].values.tolist()
    y_val = np.asarray(y_val, dtype=np.float32)
    # Tokenizer
    X_train = bert_tokenizer(X_train, padding=True, truncation=True, max_length=MAX_LENGTH)
    X_val = bert_tokenizer(X_val, padding=True, truncation=True, max_length=MAX_LENGTH)
    print("=" * 20, "字个数:", bert_tokenizer.vocab_size, "=" * 20)
    # 获取input/mask
    train_input = X_train["input_ids"]
    train_mask = X_train["attention_mask"]
    train_input = np.asarray(train_input)
    train_mask = np.asarray(train_mask)
    val_input = X_val["input_ids"]
    val_mask = X_val["attention_mask"]
    val_input = np.asarray(val_input)
    val_mask = np.asarray(val_mask)
    return train_input, val_input, train_mask, val_mask, y_train, y_val

 

def main():
    # 获取数据
    X_train_input, X_test_input, X_train_mask, X_test_mask, y_train, y_test = get_data()
    
    # 调试输出
    print(X_train_input[:5], X_train_input.shape)
    print(X_test_input[:5], X_test_input.shape)
    print(X_train_mask[:5], X_train_mask.shape)
    print(X_test_mask[:5], X_test_mask.shape)
    print(y_train[:5], y_train.shape)
    print(y_test[:5], y_test.shape)
    # Bert模型
    bert = TFBertModel.from_pretrained("Langboat/mengzi-bert-base", from_pt=True)
    input_ids = tf.keras.Input(shape=(MAX_LENGTH,), dtype=tf.int32)
    masks = tf.keras.Input(shape=(MAX_LENGTH,), dtype=tf.int32)
    bert = bert([input_ids, masks])  
    bert = bert[1]
    classifier = Dense(2, activation="softmax")(bert)
    # 模型
    model = Model(inputs=[input_ids, masks], outputs=classifier)
    print(model.summary())
    # 组合
    model.compile(optimizer=optimizer, loss=loss, metrics=["accuracy"])
    # 保存
    checkpoint = tf.keras.callbacks.ModelCheckpoint(
        "../model/bert_mengzi/bert_mengzi.ckpt", monitor='val_loss',
        verbose=1, save_best_only=True, mode='min',
        save_weights_only=True
    )
    # 训练
    model.fit([X_train_input, X_train_mask], y_train, validation_data=([X_test_input, X_test_mask], y_test),
              epochs=EPOCHS, batch_size=BATCH_SIZE,
              callbacks=[checkpoint])

 

完整代码

 

import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow.python.keras import Model
from tensorflow.keras.layers import Dense
from transformers import BertTokenizer, TFBertModel
# 超参数
EPOCHS = 50  # 迭代次数
BATCH_SIZE = 8  # 单词训练样本数目
learning_rate = 0.00003  # 学习率
INPUT_DIM = 36782 + 1
MAX_LENGTH = 780
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)  # 优化器
loss = tf.keras.losses.CategoricalCrossentropy()  # 损失
bert_tokenizer = BertTokenizer.from_pretrained('Langboat/mengzi-bert-base')  # Bert的分词器
def get_data():
    """
    读取数据
    :return: 返回分批完的训练集和测试集
    """
    # 读取数据
    data_train = pd.read_csv("../data/train.csv")
    print(data_train.head(), data_train.shape)
    data_val = pd.read_csv("../data/val.csv")
    print(data_val.head(), data_val.shape)
    # 预处理
    data_train["label"] = data_train["label"].apply(lambda x: x.split(","))
    print(data_train.head())
    data_val["label"] = data_val["label"].apply(lambda x: x.split(","))
    print(data_val.head())
    # 获取X,y
    X_train = data_train["text"].values.tolist()
    y_train = data_train["label"].values.tolist()
    y_train = np.asarray(y_train, dtype=np.float32)
    X_val = data_val["text"].values.tolist()
    y_val = data_val["label"].values.tolist()
    y_val = np.asarray(y_val, dtype=np.float32)
    # Tokenizer
    X_train = bert_tokenizer(X_train, padding=True, truncation=True, max_length=MAX_LENGTH)
    X_val = bert_tokenizer(X_val, padding=True, truncation=True, max_length=MAX_LENGTH)
    print("=" * 20, "字个数:", bert_tokenizer.vocab_size, "=" * 20)
    # 获取input/mask
    train_input = X_train["input_ids"]
    train_mask = X_train["attention_mask"]
    train_input = np.asarray(train_input)
    train_mask = np.asarray(train_mask)
    val_input = X_val["input_ids"]
    val_mask = X_val["attention_mask"]
    val_input = np.asarray(val_input)
    val_mask = np.asarray(val_mask)
    return train_input, val_input, train_mask, val_mask, y_train, y_val
def main():
    # 获取数据
    X_train_input, X_test_input, X_train_mask, X_test_mask, y_train, y_test = get_data()
    
    # 调试输出
    print(X_train_input[:5], X_train_input.shape)
    print(X_test_input[:5], X_test_input.shape)
    print(X_train_mask[:5], X_train_mask.shape)
    print(X_test_mask[:5], X_test_mask.shape)
    print(y_train[:5], y_train.shape)
    print(y_test[:5], y_test.shape)
    # Bert模型
    bert = TFBertModel.from_pretrained("Langboat/mengzi-bert-base", from_pt=True)
    input_ids = tf.keras.Input(shape=(MAX_LENGTH,), dtype=tf.int32)
    masks = tf.keras.Input(shape=(MAX_LENGTH,), dtype=tf.int32)
    bert = bert([input_ids, masks])  
    bert = bert[1]
    classifier = Dense(2, activation="softmax")(bert)
    # 模型
    model = Model(inputs=[input_ids, masks], outputs=classifier)
    print(model.summary())
    # 组合
    model.compile(optimizer=optimizer, loss=loss, metrics=["accuracy"])
    # 保存
    checkpoint = tf.keras.callbacks.ModelCheckpoint(
        "../model/bert_mengzi/bert_mengzi.ckpt", monitor='val_loss',
        verbose=1, save_best_only=True, mode='min',
        save_weights_only=True
    )
    # 训练
    model.fit([X_train_input, X_train_mask], y_train, validation_data=([X_test_input, X_test_mask], y_test),
              epochs=EPOCHS, batch_size=BATCH_SIZE,
              callbacks=[checkpoint])
if __name__ == '__main__':
    main()

Be First to Comment

发表回复

您的电子邮箱地址不会被公开。 必填项已用*标注