tensorflow / keras LSTM模型拋出值錯誤

Question

我的劇本：

import sqlite3
from tensorflow import keras
from sklearn.model_selection import train_test_split as tts
import numpy as np
import pickle

def batchSequencer(tokenedCommentList,sequenceLength,incrementSize):
    filler = -1
    sequence = []
    index = 0
    if len(tokenedCommentList) > sequenceLength:
        while index <= len(tokenedCommentList)-sequenceLength:
            sequence.append(tokenedCommentList[index:index+sequenceLength])
            index +=incrementSize
    else:
        _slice = []
        for token in tokenedCommentList:
            _slice.append(token)
        for _ in range(len(_slice),sequenceLength):
            _slice.append(filler)
        sequence.append(_slice)
    return np.array(sequence)



class batch_generator():
    def __init__(self,tFeatures,tLabels,k_tk,Length,iSize):
        self.features = tFeatures
        self.labels = tLabels
        self.tk = k_tk
        self.length = Length
        self.iSize = iSize
        self.index = 0
        self.internalIter = 0
        self.storedSequences = []
        self.sequenceIndex = 0
        self.currentLabel = []

    def generate(self):
        result = batchSequencer(self.features[self.index],self.length,self.iSize)
        y_index = self.index
        self.index +=1
        if self.index > len(self.features):
            self.index = 0
        return result, self.labels[y_index]


    def batchGenerate(self):
        x = self.tk.texts_to_matrix(self.features[self.index])
        result = batchSequencer(x,self.length,self.iSize)
        y_index = self.index
        self.index +=1
        if self.index > len(self.features):
            self.index = 0

        self.storedSequences = result
        self.currentLabel = self.labels[y_index]
        #return np.array(result),np.array(self.labels[y_index])



    def miniSequencer(self):
        if self.sequenceIndex >= len(self.storedSequences):
            self.batchGenerate()
            self.internalIter = 0
            self.sequenceIndex = 0
        result = np.array(self.storedSequences[self.sequenceIndex])
        self.sequenceIndex +=1
        return result,np.array(self.currentLabel)



def Main():
    connection = sqlite3.connect('chatDataset.db')
    c = connection.cursor()
    trainFeatures = []
    trainLabels = []
    testFeatures = []
    testLabels = []
    vocabSize = 10000
    sequenceSize = 6
    hiddenSize = 500
    goodCount = 0
    badCount = 0
    num_epochs = 10
    for row in c.execute('SELECT * FROM posts'):
        if row[3] < 0:
            trainFeatures.append(row[1])
            trainLabels.append([1,0])
            badCount +=1
        else:
            if goodCount <= badCount:
                trainFeatures.append(row[1])
                trainLabels.append([0,1])
                goodCount +=1

    try:
        tk = pickle.load( open( "tokenizer2.pkl", "rb" ) )
        trainFeatures = tk.texts_to_sequences(trainFeatures)
        print("tokenizer loaded")

    except:
        print("no tokenizer found, creating new one")
        tk = keras.preprocessing.text.Tokenizer(num_words=vocabSize, filters='!"#$%&()*+,-./:;<=>?@[\]^_`{|}~ ', lower=True, split=' ', char_level=False, oov_token=None)
        tk.fit_on_texts(trainFeatures)
        trainFeatures = tk.texts_to_sequences(trainFeatures)
        pickle.dump(tk,open( "tokenizer2.pkl", "wb" ))

    try:
        model = keras.models.load_model("LSTM_Model.mdl")
        print("loaded model successfully!")

    except:
        print("No model information found, creating new model!")
        trainFeatures,testFeatures,trainLabels,testLabels = tts(trainFeatures,trainLabels,test_size = 0.01)

        checkpointer = keras.callbacks.ModelCheckpoint(filepath='/model-{epoch:02d}.hdf5', verbose=1)
        model = keras.Sequential()
        model.add(keras.layers.Embedding(vocabSize, hiddenSize, input_length = sequenceSize))
        model.add(keras.layers.LSTM(hiddenSize, return_sequences=True))
        model.add(keras.layers.LSTM(hiddenSize, return_sequences=True))
        model.add(keras.layers.Flatten())
        model.add(keras.layers.Dense(2, activation='relu')) 
        model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=["accuracy"])


        trainGenerator = batch_generator(trainFeatures,trainLabels, tk,sequenceSize,1)
        validGenerator = batch_generator(testFeatures,testLabels, tk,sequenceSize,1)
        model.fit_generator(trainGenerator.generate(), len(trainFeatures), num_epochs,
                        validation_data=validGenerator.generate(),
                        validation_steps=len(testFeatures),callbacks=[checkpointer])

        model.save("LSTM_Model.mdl")


    print("done training model")



if __name__ == "__main__":
    Main()

完整的回溯錯誤：

Traceback (most recent call last):

  File "<ipython-input-2-dfff27295b93>", line 1, in <module>
    runfile('D:/coding projects/py practice/machine learning/autoReporter/sequencer.py', wdir='D:/coding projects/py practice/machine learning/autoReporter')

  File "C:\Users\oxrock\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 701, in runfile
    execfile(filename, namespace)

  File "C:\Users\oxrock\Anaconda3\lib\site-packages\spyder\utils\site\sitecustomize.py", line 101, in execfile
    exec(compile(f.read(), filename, 'exec'), namespace)

  File "D:/coding projects/py practice/machine learning/autoReporter/sequencer.py", line 147, in <module>
    Main()

  File "D:/coding projects/py practice/machine learning/autoReporter/sequencer.py", line 137, in Main
    validation_steps=len(testFeatures),callbacks=[checkpointer])

  File "C:\Users\oxrock\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\training.py", line 1779, in fit_generator
    initial_epoch=initial_epoch)

  File "C:\Users\oxrock\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\training_generator.py", line 136, in fit_generator
    val_x, val_y, val_sample_weight)

  File "C:\Users\oxrock\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\training.py", line 917, in _standardize_user_data
    exception_prefix='target')

  File "C:\Users\oxrock\Anaconda3\lib\site-packages\tensorflow\python\keras\engine\training_utils.py", line 191, in standardize_input_data
    ' but got array with shape ' + str(data_shape))

ValueError: Error when checking target: expected dense_1 to have shape (2,) but got array with shape (1,)

該模型的目的是將Reddit評論分為負面/正面帖子。 從中提取訓練數據的sql數據庫包含我從reddit rip中提取並分類的帖子。 我必須通過批處理生成器類將數據分成小批，因為不可能將所有數據都保存在內存中。

我很難用“ ValueError：檢查目標時出錯：期望deny_1具有形狀（2，）但形狀為（1，）的數組”來訓練該模型，得到。

我已經堅持了一段時間，現在已經到了隨機改變事物的地步，希望能創造奇跡。 一臂之力將不勝感激。 如果需要任何其他信息，我將很樂意發布。

Answer 1

當前，您的代碼有些錯誤。 首先，數據生成器不僅應返回一個值，還應返回數據集的所有值。 通常，您可以通過循環執行此操作，並在循環內部使用yield來返回單個數據。

其次，不傳遞python列表而是將np.array傳遞給keras可能會有所幫助。 請嘗試一下，如果在訓練模型之前投射標簽會發生什么情況。 此外，請確保您的數據具有正確的形狀，即目標標簽應具有形狀(batch size, 2) ，輸入數據應具有類似的形狀(batch size, sequenceSize, vocabSize) 。 請打印訓練/驗證數據（輸入和目標數據）的形狀，並確保這是正確的。

最后，如果您正在執行一些分類任務，則網絡的最后激活功能應該是softmax。 因此，請更換

model.add(keras.layers.Dense(2, activation='relu'))

同

model.add(keras.layers.Dense(2, activation='softmax'))