AI实战：从入门到精通系列——用全连接神经网络实现情感分类（二）

本文链接：https://blog.csdn.net/zengNLP/article/details/101217887

AI实战：从入门到精通系列——用全连接神经网络实现情感分类（一）
AI实战：从入门到精通系列——用全连接神经网络实现情感分类（二）

用全连接神经网络实现情感分类

环境

Ubuntu16.04 LTS
python 3.x
numpy
gensim
jieba
pandas
数据集

weibo_senti_100k 微博情感数据集
数据处理

详见：AI实战：从入门到精通系列——用感知器实现情感分类（二）

数据加载
data_helper.py

import numpy as np
import pandas as pd
import jieba
import multiprocessing
from gensim.models import Word2Vec
from gensim.models.word2vec import LineSentence

max_text_len = 50#50个词
word2vec_dimension = 100

def load_train_test_data(train_path, test_path, w2v_model_path):
    
    #加载数据
    pd_all = pd.read_csv(train_path)
    train_data_x, train_data_y = pd_all.text.tolist(), pd_all.label.tolist()
    
    pd_all = pd.read_csv(test_path)
    test_data_x, test_data_y = pd_all.text.tolist(), pd_all.label.tolist()
    
    train_data_x = train_data_x[:5000]#数据太大了，只取5000个作为训练
    train_data_y = train_data_y[:5000]
    
    #test_data_x = test_data_x[:500]
    #test_data_y = test_data_y[:500]
    
    #分词
    train_data_x = [segment(k) for k in train_data_x]
    test_data_x = [segment(k) for k in test_data_x]
    
    
    #文本转向量表示
    w2v_model = Word2Vec.load(w2v_model_path)
    train_data_x = text_to_word2vec(w2v_model, train_data_x)
    test_data_x = text_to_word2vec(w2v_model, test_data_x)
    
    return train_data_x, test_data_x, train_data_y, test_data_y

模型训练
train.py

import os, sys
import data_helper
import fc_net
import numpy as np
from datetime import datetime

def get_result(vec):

    max_value_index = 0
    max_value = 0
    for i in range(len(vec)):
        if vec[i] > max_value:
            max_value = vec[i]
            max_value_index = i
    return max_value_index
    
    
def evaluate(network, test_data_set, test_labels):

    error = 0
    total = len(test_data_set)
    for i in range(total):
        label = get_result(test_labels[i])
        predict = get_result(network.predict(test_data_set[i]))
        if label != predict:
            error += 1
    return float(error) / float(total)
    
                
def now():
    return datetime.now().strftime('%c')

def to_sentence_vector(data):
    #求和，再求平均
    
    data_vec = np.zeros((len(data), len(data[0][0])))
    
    i = 0
    while i < len(data):
        n = data[i].shape[0]
        data_vec[i] = np.sum(data[i], axis=0)/n
        i += 1
    
    return data_vec
    
def to_one_hot(label_list):
    
    label_vec = []
    i = 0
    while i < len(label_list):
        if label_list[i] == 1:
            label_vec.append([0.1, 0.9])
        else:
            label_vec.append([0.9, 0.1])
        i += 1
    
    return label_vec
    
def transpose(args):
    return list(map(
        lambda arg: list(map(
            lambda line: np.array(line).reshape(len(line), 1)
            , arg))
        , args
    ))

def train_and_evaluate(train_path, test_path, w2v_model_path):

    x_train, x_test, y_train, y_test = data_helper.load_train_test_data(train_path, test_path, w2v_model_path)
    
    x_train = to_sentence_vector(x_train)
    x_test = to_sentence_vector(x_test)
    y_train = to_one_hot(y_train)
    y_test = to_one_hot(y_test)
    
    x_train, x_test, y_train, y_test = transpose((x_train, x_test, y_train, y_test))
    
    
    #1、句子的向量长度为100，隐藏层数量为1，节点数量为50，分类数量为2
    network = fc_net.Network([100, 50, 2])
    
    #2、句子的向量长度为100，隐藏层数量为5，节点数量为50，分类数量为2
    #network = fc_net.Network([100, 50, 50, 50, 50, 50, 2])
    
    last_error_ratio = 1.0
    epoch = 0
    while True:
        epoch += 1
        network.train(y_train, x_train, 0.1, 1)
        #print('%s epoch %d finished' % (now(), epoch))
        if epoch % 10 == 0:
            error_ratio = evaluate(network, x_test, y_test)
            print('%s after epoch %d, error ratio is %f' % (now(), epoch, error_ratio))
            if error_ratio > last_error_ratio:#准确率不再上升则提前停止训练
                break
            else:
                last_error_ratio = error_ratio
                
                
if __name__ == "__main__":  
    train_path = './data/train.csv'
    test_path = './data/test.csv'
    w2v_model_path = './data/word2vec.model'
    train_and_evaluate(train_path, test_path, w2v_model_path)

执行

python train.py
输出

1、句子的向量长度为100，隐藏层数量为1，节点数量为50，分类数量为2
network = fc_net.Network([100, 50, 2])
```
Mon Sep 23 20:27:46 2019 after epoch 10, error ratio is 0.124594
Mon Sep 23 20:27:50 2019 after epoch 20, error ratio is 0.121677
Mon Sep 23 20:27:54 2019 after epoch 30, error ratio is 0.115593
Mon Sep 23 20:27:58 2019 after epoch 40, error ratio is 0.139012	
```
结论：验证集上的准确率为88.45%，比AI实战：从入门到精通系列——用感知器实现情感分类（二）的感知器模型的81.89%要高一些。

2、句子的向量长度为100，隐藏层数量为5，节点数量为50，分类数量为2
network = fc_net.Network([100, 50, 50, 50, 50, 50, 2])
```
Mon Sep 23 20:30:57 2019 after epoch 10, error ratio is 0.144929
Mon Sep 23 20:31:06 2019 after epoch 20, error ratio is 0.103842
Mon Sep 23 20:31:16 2019 after epoch 30, error ratio is 0.113509
```
结论：验证集上的准确率为89.62%，5层隐藏层比1层隐藏层的准确率要高一些，说明神经网络的深度起作用了，这也是 “深度学习” “深度” 的魅力所在！