为 tf.nn.embedding_lookup 预处理不同文本大小时 Pre-Padding 和 Post-Padding 文本的区别

Question

在馈入嵌入层时，我看到了两种类型的填充。

例如：

考虑两句话：

word1 = "我是个爱狗的人。"

word2 = "Krishni 和 Pradeepa 都喜欢猫。"

word1_int = [1,2,3,4,5,6]

word2_int = [7,8,9,10,11,12,13]

将两个单词填充到长度 = 8

填充方法1（在开头放0）

word1_int = [0,0,1,2,3,4,5,6]

word2_int = [0,7,8,9,10,11,12,13]

填充方法2（最后加0）

word1_int = [1,2,3,4,5,6,0,0]

word2_int = [7,8,9,10,11,12,13,0]

我正在尝试使用 20 个新闻组数据集进行在线分类。我目前正在使用第一种方法来填充我的文本。

问题：在我的实现中使用第一种方法比另一种方法有什么优势吗？

提前谢谢你！

我的代码如下所示：

from collections import Counter
import tensorflow as tf
from sklearn.datasets import fetch_20newsgroups
import matplotlib as mplt
mplt.use('agg') # Must be before importing matplotlib.pyplot or pylab!
import matplotlib.pyplot as plt
from string import punctuation
from sklearn.preprocessing import LabelBinarizer
import numpy as np
from nltk.corpus import stopwords
import nltk
nltk.download('stopwords')



def pre_process():
    newsgroups_data = fetch_20newsgroups(subset='all', remove=('headers', 'footers', 'quotes'))

    words = []
    temp_post_text = []
    print(len(newsgroups_data.data))

    for post in newsgroups_data.data:

        all_text = ''.join([text for text in post if text not in punctuation])
        all_text = all_text.split('\n')
        all_text = ''.join(all_text)
        temp_text = all_text.split(" ")

        for word in temp_text:
            if word.isalpha():
                temp_text[temp_text.index(word)] = word.lower()

        # temp_text = [word for word in temp_text if word not in stopwords.words('english')]
        temp_text = list(filter(None, temp_text))
        temp_text = ' '.join([i for i in temp_text if not i.isdigit()])
        words += temp_text.split(" ")
        temp_post_text.append(temp_text)

    # temp_post_text = list(filter(None, temp_post_text))

    dictionary = Counter(words)
    # deleting spaces
    # del dictionary[""]
    sorted_split_words = sorted(dictionary, key=dictionary.get, reverse=True)
    vocab_to_int = {c: i for i, c in enumerate(sorted_split_words,1)}

    message_ints = []
    for message in temp_post_text:
        temp_message = message.split(" ")
        message_ints.append([vocab_to_int[i] for i in temp_message])


    # maximum message length = 6577

    # message_lens = Counter([len(x) for x in message_ints])AAA

    seq_length = 6577
    num_messages = len(temp_post_text)
    features = np.zeros([num_messages, seq_length], dtype=int)
    for i, row in enumerate(message_ints):
        print(features[i, -len(row):])
        features[i, -len(row):] = np.array(row)[:seq_length]
        print(features[i, -len(row):])

    lb = LabelBinarizer()
    lbl = newsgroups_data.target
    labels = np.reshape(lbl, [-1])
    labels = lb.fit_transform(labels)

    return features, labels, len(sorted_split_words)+1


def get_batches(x, y, batch_size=1):
    for ii in range(0, len(y), batch_size):
        yield x[ii:ii + batch_size], y[ii:ii + batch_size]


def plot(noOfWrongPred, dataPoints):
    font_size = 14
    fig = plt.figure(dpi=100,figsize=(10, 6))
    mplt.rcParams.update({'font.size': font_size})
    plt.title("Distribution of wrong predictions", fontsize=font_size)
    plt.ylabel('Error rate', fontsize=font_size)
    plt.xlabel('Number of data points', fontsize=font_size)

    plt.plot(dataPoints, noOfWrongPred, label='Prediction', color='blue', linewidth=1.8)
    # plt.legend(loc='upper right', fontsize=14)

    plt.savefig('distribution of wrong predictions.png')
    # plt.show()



def train_test():
    features, labels, n_words = pre_process()

    print(features.shape)
    print(labels.shape)

    # Defining Hyperparameters

    lstm_layers = 1
    batch_size = 1
    lstm_size = 200
    learning_rate = 0.01

    # --------------placeholders-------------------------------------

    # Create the graph object
    graph = tf.Graph()
    # Add nodes to the graph
    with graph.as_default():

        tf.set_random_seed(1)

        inputs_ = tf.placeholder(tf.int32, [None, None], name="inputs")
        # labels_ = tf.placeholder(dtype= tf.int32)
        labels_ = tf.placeholder(tf.float32, [None, None], name="labels")

        # output_keep_prob is the dropout added to the RNN's outputs, the dropout will have no effect on the calculation of the subsequent states.
        keep_prob = tf.placeholder(tf.float32, name="keep_prob")

        # Size of the embedding vectors (number of units in the embedding layer)
        embed_size = 300

        # generating random values from a uniform distribution (minval included and maxval excluded)
        embedding = tf.Variable(tf.random_uniform((n_words, embed_size), -1, 1),trainable=True)
        embed = tf.nn.embedding_lookup(embedding, inputs_)

        print(embedding.shape)
        print(embed.shape)
        print(embed[0])

        # Your basic LSTM cell
        lstm = tf.contrib.rnn.BasicLSTMCell(lstm_size)


        # Add dropout to the cell
        drop = tf.contrib.rnn.DropoutWrapper(lstm, output_keep_prob=keep_prob)

        # Stack up multiple LSTM layers, for deep learning
        cell = tf.contrib.rnn.MultiRNNCell([drop] * lstm_layers)

        # Getting an initial state of all zeros
        initial_state = cell.zero_state(batch_size, tf.float32)

        outputs, final_state = tf.nn.dynamic_rnn(cell, embed, initial_state=initial_state)

        # hidden layer
        hidden = tf.layers.dense(outputs[:, -1], units=25, activation=tf.nn.relu)

        print(hidden.shape)

        logit = tf.contrib.layers.fully_connected(hidden, num_outputs=20, activation_fn=None)

        cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logit, labels=labels_))

        optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)

        saver = tf.train.Saver()

    # ----------------------------online training-----------------------------------------

    with tf.Session(graph=graph) as sess:
        tf.set_random_seed(1)
        sess.run(tf.global_variables_initializer())
        iteration = 1
        state = sess.run(initial_state)
        wrongPred = 0
        noOfWrongPreds = []
        dataPoints = []

        for ii, (x, y) in enumerate(get_batches(features, labels, batch_size), 1):

            feed = {inputs_: x,
                    labels_: y,
                    keep_prob: 0.5,
                    initial_state: state}

            embedzz = sess.run(embedding, feed_dict=feed)

            print(embedzz)


            predictions = tf.nn.softmax(logit).eval(feed_dict=feed)

            print("----------------------------------------------------------")
            print("Iteration: {}".format(iteration))

            isequal = np.equal(np.argmax(predictions[0], 0), np.argmax(y[0], 0))

            print(np.argmax(predictions[0], 0))
            print(np.argmax(y[0], 0))

            if not (isequal):
                wrongPred += 1

            print("nummber of wrong preds: ",wrongPred)

            if iteration%50 == 0:
                noOfWrongPreds.append(wrongPred/iteration)
                dataPoints.append(iteration)

            loss, states, _ = sess.run([cost, final_state, optimizer], feed_dict=feed)

            print("Train loss: {:.3f}".format(loss))
            iteration += 1

        saver.save(sess, "checkpoints/sentiment.ckpt")
        errorRate = wrongPred / len(labels)
        print("ERRORS: ", wrongPred)
        print("ERROR RATE: ", errorRate)
        plot(noOfWrongPreds, dataPoints)


if __name__ == '__main__':
    train_test()

这是我用来填充所有句子的代码示例。

  seq_length = 6577
  num_messages = len(temp_post_text)
    features = np.zeros([num_messages, seq_length], dtype=int)
    for i, row in enumerate(message_ints):
        print(features[i, -len(row):])
        features[i, -len(row):] = np.array(row)[:seq_length]
        print(features[i, -len(row):])

Answer 1

通常，当我们使用 LSTM 或 RNN 时，我们使用最终输出或隐藏状态并将其传递来进行预测。你也在做与这一行相同的事情：

logit = tf.contrib.layers.fully_connected(hidden, num_outputs=20, activation_fn=None)

这里区分了两种填充方法。如果您使用第二种填充方法，即后填充，那么最终的隐藏状态将被刷新，因为大多数情况下它将是0，而通过使用第一种方法，我们确保隐藏状态输出是正确的。

Answer 2

请记住我们在使用 RNN 时遇到的问题，即消失梯度。因此，如果您使用的是 Pre-padding，那么 RNN 会在最后记住实际值，而对于 post-padding，实际值在开头。因此，如果填充很长，模型很可能会忘记需要记住的实际值，并且模型可能表现不佳。 所以，我个人建议你使用 pre-padding 而不是 post-padding。