python - Keras 文本分类 : Validation Accuracy doesn't change

Question

所以，我正在构建一个基本的 Keras 文本分类器，但无论我做什么，我都无法使验证准确率高于 49-50%(或更低)。我的训练准确率攀升非常正常，从 50% 左右开始，在 4-5 个 epoch 后攀升至 80% 左右。

这是一个输出示例:

 - 54s - loss: 0.6982 - acc: 0.5064 - val_loss: 0.6932 - val_acc: 0.4950
Epoch 2/3
 - 57s - loss: 0.6560 - acc: 0.6580 - val_loss: 0.7324 - val_acc: 0.4950
Epoch 3/3
 - 60s - loss: 0.5359 - acc: 0.7047 - val_loss: 0.7339 - val_acc: 0.4955

这是我的代码:

import logging
logging.getLogger("tensorflow").setLevel(logging.ERROR)
import numpy as np
import os
from keras.preprocessing.text import Tokenizer
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Flatten
from keras.layers.embeddings import Embedding
from keras.layers import Dropout
from keras.preprocessing import sequence
from keras.layers.convolutional import Conv1D
from keras.layers.convolutional import MaxPooling1D
from keras import optimizers
np.random.seed(7)
class TextClassifier:

    def __init__(self):
        self.tokenizer = Tokenizer(num_words = 5000)
        self.top_words = 5000
        self.max_words = 500
        self.model = model = Sequential()
        model.add(Embedding(self.top_words,64,input_length = self.max_words))
        model.add(Conv1D(filters = 64,kernel_size=3,padding='same',activation='relu'))
        model.add(MaxPooling1D(pool_size=2))
        model.add(Dropout(0.5))
        model.add(Flatten())
        model.add(Dense(250,activation='relu'))
        model.add(Dense(1,activation='sigmoid'))
        model.compile(loss='binary_crossentropy',optimizer = 'adam',metrics = ['accuracy'])

    def train(self, X_train, y_train, X_test, y_test):
        self.model.fit(X_train,y_train, shuffle = True,
        validation_data = (X_test,y_test),epochs = 3,batch_size=512,verbose = 2)

    def predict(self,X):
        return self.model.predict(X)

    def init_tokenizer(self,X):
        self.tokenizer.fit_on_texts(X)

    def eval(self,X,y):
        return self.model.evaluate(X, y, verbose=0)

    def proccess_text(self,X):
        vocab_text = self.tokenizer.texts_to_sequences(X)
        vocab_text = sequence.pad_sequences(vocab_text,maxlen=500)
        return vocab_text

    def organize_text(self,pos_path,neg_path):
        data = {'label':[],'text':[]}
        pos_texts = self.text_to_array(pos_path)
        neg_texts = self.text_to_array(neg_path)
        for i in pos_texts:
            data['label'].append(0)
            data['text'].append(i)
        for i in neg_texts:
            data['label'].append(1)
            data['text'].append(i)
        return data

    def text_to_array(self,path):
            ''' Takes a path argument and retrieves all the text lines from
                within a folder'''
            name_list = []
            texts = []
            for file_ in os.listdir(path):
                name_list.append(file_)
            for i in name_list:
                file_ = open(path+i)
                texts.append(file_.read().splitlines())
            return texts





from TextClassifier import *

path = './data/train/'
test_path = './data/test/'
model = TextClassifier()

data = model.organize_text(path+'pos/',path+'neg/')
tests = model.organize_text(test_path+'pos/',test_path+'neg/')
model.init_tokenizer(data['text'])
model.init_tokenizer(tests['text'])
X_train  = np.array(model.proccess_text(data['text']))
X_test = np.array(model.proccess_text(tests['text']))
y_train = data['label']
y_test = tests['label']
model.train(X_train,y_train,X_test,y_test)
scores = model.eval(X_test,y_test)
input_ = model.proccess_text(['It was very good! Awesome! Enjoyable!'])
print("Predict: ")
predict = model.predict(input_)
print(predict)
print("Accuracy: %.2f%%" % (scores[1]*100))

我的语料库来自这里:http://ai.stanford.edu/~amaas/data/sentiment/并且我只是将每个类的前 1000 个拆分为验证数据。 (他们是无序的 afaik)

Answer 1

我不知道 Dropout(0.5) 对您的模型有多大帮助。看起来 BatchNormalization() 可能是一个很好的替代品。您可能需要在嵌入层之后考虑一个 LSTM 层。您可以直接将 dropout 添加到 LSTM。例如:

model.add(LSTM(lstm_out, dropout_U=0.2, dropout_W=0.2))

对于一个很好的完整示例，我邀请您查看此模型: https://www.kaggle.com/ngyptr/lstm-sentiment-analysis-keras

希望这对您有所帮助，祝您好运!

编辑:根据我下面的评论

我认为您的模型可能正在训练非字符标记，因为如果您使用此 text_to_array(self,path) 函数，您的验证准确度为 .79、.85 和 .87:

    import re
def text_to_array(self,path):
            ''' Takes a path argument and retrieves all the text lines from
                within a folder'''
            name_list = []
            texts = []
            for file_ in os.listdir(path):
                name_list.append(file_)
            for i in name_list:
                file_ = open(path+i)
                for line in file_.read().splitlines():
                    # I just preprocessed your text here
                    texts.append(re.sub('[^a-zA-z0-9\s]','',line.lower()))
            return texts

如果这有帮助，请告诉我。