Tensorflow 中的形状不匹配

Question

我对 TensorFlow 比较陌生，正在尝试在生产环境中实施我的第一个模型。该模型经过良好的训练和测试，但我发现使用此算法转移到生产中非常具有挑战性。谁能告诉我为什么我的评估行会出现以下错误？

ValueError: Cannot feed value of shape (1, 1095277) for Tensor 'input:0', which has shape '(?, 2912)'

我正在实现的代码是（我尝试了各种不同的方法来让它工作）：

哪个张量的长度为 1x1095277？

def use_neural_network(input_data, lexicon,stopWords):
    x= tf.placeholder('float', shape=[None, 2912], name='input')
    y= tf.placeholder('float', name='output')

    #x = tf.Variable('float', [None, 2912]', name='input')
    #y = tf.Variable('float', name='output')

    hidden_1_layer = {'weights':tf.Variable(tf.random_normal([2912, 1])),'biases':tf.Variable(tf.random_normal([1]))}
    output_layer = {'weights':tf.Variable(tf.random_normal([1, 2])),'biases':tf.Variable(tf.random_normal([2])),}
    def neural_network_model(data):
        l1 = tf.add(tf.matmul(data,hidden_1_layer['weights']), hidden_1_layer['biases'])
        l1 = tf.nn.relu(l1)
        output = tf.matmul(l1,output_layer['weights']) + output_layer['biases']
        return output

    prediction = neural_network_model(x) 
    saver=tf.train.Saver()
    with tf.Session() as sess:
        saver.restore(sess,"model.ckpt")   
        lemmatizer = WordNetLemmatizer()
        current_words = word_tokenize(input_data.lower())
        current_words = [re.sub("[^a-zA-Z]"," ", i) for i in current_words]
        current_words = [re.sub("\s{1,10}"," ", i) for i in current_words]
        current_words = [i for i in current_words if i not in stopWords]   
        current_words = [lemmatizer.lemmatize(i) for i in current_words]
        features = np.zeros(len(lexicon))
        for word in current_words:
            if word.lower() in lexicon:
                index_value = lexicon.index(word.lower())
                features[index_value] += 1
                print(pd.Series(features).sum())
            features = np.array(list(features))
            result = (sess.run(tf.argmax(prediction.eval(feed_dict={x:[features]}),1)))
            if result[0] == 0:
                print('No:',input_data)
            elif result[0] == 1:
                print('Yes:',input_data)

with open('lexicon_1.pickle','rb') as f:
    lexicon = pickle.load(f)
stopWords = set(stopwords.words('english'))
use_neural_network('I do not understand the problem', lexicon, stopWords)

Answer 1

您的网络似乎需要大小为 [2912, 1] 的输入，如 hidden_1_layer

所定义

hidden_1_layer = {'weights':tf.Variable(tf.random_normal([2912, 1])), ...

当您调用预测时，您不会使用大小为 [2912, 1] 的输入来调用它，而是使用等于您词典长度的输入来调用它，该词典（可能）包含 1095277 个数字。

features = np.zeros(len(lexicon))

我还怀疑您将 features 数组包装了两次，首先是 features = np.array(list(features))，然后是 x:[features]。对你的数据不是很自信，但感觉不对。

就个人而言，我发现从教程中复制并修改行比尝试从头开始编写更容易学习。

Tensorflow 中的形状不匹配

Shape Mismatch in Tensorflow

nltk

python-3.x

tensorflow

tensor