11-04 02:18 阅读 107

Tensorflow深度学习使用CNN分类英文文本

这篇文章主要为大家介绍了Tensorflow深度学习CNN实现英文文本分类示例解析，有需要的朋友可以借鉴参考下，希望能够有所帮助，祝大家多多进步

前言

Github源码地址

本文同时也是学习唐宇迪老师深度学习课程的一些理解与记录。

文中代码是实现在TensorFlow下使用卷积神经网络（CNN）做英文文本的分类任务（本次是垃圾邮件的二分类任务），当然垃圾邮件分类是一种应用环境，模型方法也可以推广到其它应用场景，如电商商品好评差评分类、正负面新闻等。

这里写图片描述

源码与数据

源码

- data_helpers.py
- train.py
- text_cnn.py
- eval.py（Save the evaluations to a csv, in case the user wants to inspect,analyze, or otherwise use the classifications generated by the neural net）

数据

- rt-polarity.neg

Python客栈送红包、纸质书

- rt-polarity.pos

这里写图片描述

train.py 源码及分析

import tensorflow as tf

import numpy as np

import os

import time

import datetime

import data_helpers

from text_cnn import TextCNN

from tensorflow.contrib import learn

# Parameters

# ==================================================

# Data loading params

# 语料文件路径定义

tf.flags.DEFINE_float("dev_sample_percentage", .1, "Percentage of the training data to use for validation")

tf.flags.DEFINE_string("positive_data_file", "./data/rt-polaritydata/rt-polarity.pos", "Data source for the positive data.")

tf.flags.DEFINE_string("negative_data_file", "./data/rt-polaritydata/rt-polarity.neg", "Data source for the negative data.")

# Model Hyperparameters

# 定义网络超参数

tf.flags.DEFINE_integer("embedding_dim", 128, "Dimensionality of character embedding (default: 128)")

tf.flags.DEFINE_string("filter_sizes", "3,4,5", "Comma-separated filter sizes (default: '3,4,5')")

tf.flags.DEFINE_integer("num_filters", 128, "Number of filters per filter size (default: 128)")

tf.flags.DEFINE_float("dropout_keep_prob", 0.5, "Dropout keep probability (default: 0.5)")

tf.flags.DEFINE_float("l2_reg_lambda", 0.0, "L2 regularization lambda (default: 0.0)")

# Training parameters

# 训练参数

tf.flags.DEFINE_integer("batch_size", 32, "Batch Size (default: 32)")

tf.flags.DEFINE_integer("num_epochs", 200, "Number of training epochs (default: 200)") # 总训练次数

tf.flags.DEFINE_integer("evaluate_every", 100, "Evaluate model on dev set after this many steps (default: 100)") # 每训练100次测试一下

tf.flags.DEFINE_integer("checkpoint_every", 100, "Save model after this many steps (default: 100)") # 保存一次模型

tf.flags.DEFINE_integer("num_checkpoints", 5, "Number of checkpoints to store (default: 5)")

# Misc Parameters

tf.flags.DEFINE_boolean("allow_soft_placement", True, "Allow device soft device placement") # 加上一个布尔类型的参数，要不要自动分配

tf.flags.DEFINE_boolean("log_device_placement", False, "Log placement of ops on devices") # 加上一个布尔类型的参数，要不要打印日志

# 打印一下相关初始参数

FLAGS = tf.flags.FLAGS

FLAGS._parse_flags()

print("\nParameters:")

for attr, value in sorted(FLAGS.__flags.items()):

    print("{}={}".format(attr.upper(), value))

print("")

# Data Preparation

# ==================================================

# Load data

print("Loading data...")

x_text, y = data_helpers.load_data_and_labels(FLAGS.positive_data_file, FLAGS.negative_data_file)

# Build vocabulary

max_document_length = max([len(x.split(" ")) for x in x_text]) # 计算最长邮件

vocab_processor = learn.preprocessing.VocabularyProcessor(max_document_length) # tensorflow提供的工具，将数据填充为最大长度，默认0填充

x = np.array(list(vocab_processor.fit_transform(x_text)))

# Randomly shuffle data

# 数据洗牌

np.random.seed(10)

# np.arange生成随机序列

shuffle_indices = np.random.permutation(np.arange(len(y)))

x_shuffled = x[shuffle_indices]

y_shuffled = y[shuffle_indices]

# 将数据按训练train和测试dev分块

# Split train/test set

# TODO: This is very crude, should use cross-validation

dev_sample_index = -1 * int(FLAGS.dev_sample_percentage * float(len(y)))

x_train, x_dev = x_shuffled[:dev_sample_index], x_shuffled[dev_sample_index:]

y_train, y_dev = y_shuffled[:dev_sample_index], y_shuffled[dev_sample_index:]

print("Vocabulary Size: {:d}".format(len(vocab_processor.vocabulary_)))

print("Train/Dev split: {:d}/{:d}".format(len(y_train), len(y_dev))) # 打印切分的比例

# Training

# ==================================================

with tf.Graph().as_default():

    session_conf = tf.ConfigProto(

        allow_soft_placement=FLAGS.allow_soft_placement,

        log_device_placement=FLAGS.log_device_placement)

    sess = tf.Session(config=session_conf)

    with sess.as_default():

        # 卷积池化网络导入

        cnn = TextCNN(

            sequence_length=x_train.shape[1],

            num_classes=y_train.shape[1], # 分几类

            vocab_size=len(vocab_processor.vocabulary_),

            embedding_size=FLAGS.embedding_dim,

            filter_sizes=list(map(int, FLAGS.filter_sizes.split(","))), # 上面定义的filter_sizes拿过来，"3,4,5"按","分割

            num_filters=FLAGS.num_filters, # 一共有几个filter

            l2_reg_lambda=FLAGS.l2_reg_lambda) # l2正则化项

        # Define Training procedure

        global_step = tf.Variable(0, name="global_step", trainable=False)

        optimizer = tf.train.AdamOptimizer(1e-3) # 定义优化器

        grads_and_vars = optimizer.compute_gradients(cnn.loss)

        train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)

        # Keep track of gradient values and sparsity (optional)

        grad_summaries = []

        for g, v in grads_and_vars:

            if g is not None:

                grad_hist_summary = tf.summary.histogram("{}/grad/hist".format(v.name), g)

                sparsity_summary = tf.summary.scalar("{}/grad/sparsity".format(v.name), tf.nn.zero_fraction(g))

                grad_summaries.append(grad_hist_summary)

                grad_summaries.append(sparsity_summary)

        grad_summaries_merged = tf.summary.merge(grad_summaries)

        # Output directory for models and summaries

        timestamp = str(int(time.time()))

        out_dir = os.path.abspath(os.path.join(os.path.curdir, "runs", timestamp))

        print("Writing to {}\n".format(out_dir))

        # Summaries for loss and accuracy

        # 损失函数和准确率的参数保存

        loss_summary = tf.summary.scalar("loss", cnn.loss)

        acc_summary = tf.summary.scalar("accuracy", cnn.accuracy)

        # Train Summaries

        # 训练数据保存

        train_summary_op = tf.summary.merge([loss_summary, acc_summary, grad_summaries_merged])

        train_summary_dir = os.path.join(out_dir, "summaries", "train")

        train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph)

        # Dev summaries

        # 测试数据保存

        dev_summary_op = tf.summary.merge([loss_summary, acc_summary])

        dev_summary_dir = os.path.join(out_dir, "summaries", "dev")

        dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph)

        # Checkpoint directory. Tensorflow assumes this directory already exists so we need to create it

        checkpoint_dir = os.path.abspath(os.path.join(out_dir, "checkpoints"))

        checkpoint_prefix = os.path.join(checkpoint_dir, "model")

        if not os.path.exists(checkpoint_dir):

            os.makedirs(checkpoint_dir)

        saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints) # 前面定义好参数num_checkpoints

        # Write vocabulary

        vocab_processor.save(os.path.join(out_dir, "vocab"))

        # Initialize all variables

        sess.run(tf.global_variables_initializer()) # 初始化所有变量

        # 定义训练函数

        def train_step(x_batch, y_batch):

"""

            A single training step

"""

            feed_dict = {

              cnn.input_x: x_batch,

              cnn.input_y: y_batch,

              cnn.dropout_keep_prob: FLAGS.dropout_keep_prob # 参数在前面有定义

}

            _, step, summaries, loss, accuracy = sess.run(

                [train_op, global_step, train_summary_op, cnn.loss, cnn.accuracy], feed_dict)

            time_str = datetime.datetime.now().isoformat() # 取当前时间，python的函数

            print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))

            train_summary_writer.add_summary(summaries, step)

        # 定义测试函数

        def dev_step(x_batch, y_batch, writer=None):

"""

            Evaluates model on a dev set

"""

            feed_dict = {

              cnn.input_x: x_batch,

              cnn.input_y: y_batch,

              cnn.dropout_keep_prob: 1.0 # 神经元全部保留

}

            step, summaries, loss, accuracy = sess.run(

                [global_step, dev_summary_op, cnn.loss, cnn.accuracy], feed_dict)

            time_str = datetime.datetime.now().isoformat()

            print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))

            if writer:

                writer.add_summary(summaries, step)

        # Generate batches

        batches = data_helpers.batch_iter(list(zip(x_train, y_train)), FLAGS.batch_size, FLAGS.num_epochs)

        # Training loop. For each batch...

        # 训练部分

        for batch in batches:

            x_batch, y_batch = zip(*batch) # 按batch把数据拿进来

            train_step(x_batch, y_batch)

            current_step = tf.train.global_step(sess, global_step) # 将Session和global_step值传进来

            if current_step % FLAGS.evaluate_every == 0: # 每FLAGS.evaluate_every次每100执行一次测试

                print("\nEvaluation:")

                dev_step(x_dev, y_dev, writer=dev_summary_writer)

                print("")

            if current_step % FLAGS.checkpoint_every == 0: # 每checkpoint_every次执行一次保存模型

                path = saver.save(sess, './', global_step=current_step) # 定义模型保存路径

                print("Saved model checkpoint to {}\n".format(path))

data_helpers.py 源码及分析

import numpy as np

import re

import itertools

from collections import Counter

def clean_str(string):

"""

    Tokenization/string cleaning for all datasets except for SST.

    Original taken from https://github.com/yoonkim/CNN_sentence/blob/master/process_data.py

"""

    # 清理数据替换掉无词义的符号

    string = re.sub(r"[^A-Za-z0-9(),!?\'\`]", " ", string)

    string = re.sub(r"\'s", " \'s", string)

    string = re.sub(r"\'ve", " \'ve", string)

    string = re.sub(r"n\'t", " n\'t", string)

    string = re.sub(r"\'re", " \'re", string)

    string = re.sub(r"\'d", " \'d", string)

    string = re.sub(r"\'ll", " \'ll", string)

    string = re.sub(r",", " , ", string)

    string = re.sub(r"!", " ! ", string)

    string = re.sub(r"\(", " \( ", string)

    string = re.sub(r"\)", " \) ", string)

    string = re.sub(r"\?", " \? ", string)

    string = re.sub(r"\s{2,}", " ", string)

    return string.strip().lower()

def load_data_and_labels(positive_data_file, negative_data_file):

"""

    Loads MR polarity data from files, splits the data into words and generates labels.

    Returns split sentences and labels.

"""

    # Load data from files

    positive = open(positive_data_file, "rb").read().decode('utf-8')

    negative = open(negative_data_file, "rb").read().decode('utf-8')

    # 按回车分割样本

    positive_examples = positive.split('\n')[:-1]

    negative_examples = negative.split('\n')[:-1]

    # 去空格

    positive_examples = [s.strip() for s in positive_examples]

    negative_examples = [s.strip() for s in negative_examples]

    #positive_examples = list(open(positive_data_file, "rb").read().decode('utf-8'))

    #positive_examples = [s.strip() for s in positive_examples]

    #negative_examples = list(open(negative_data_file, "rb").read().decode('utf-8'))

    #negative_examples = [s.strip() for s in negative_examples]

    # Split by words

    x_text = positive_examples + negative_examples

    x_text = [clean_str(sent) for sent in x_text] # 字符过滤，实现函数见clean_str()

    # Generate labels

    positive_labels = [[0, 1] for _ in positive_examples]

    negative_labels = [[1, 0] for _ in negative_examples]

    y = np.concatenate([positive_labels, negative_labels], 0) # 将两种label连在一起

    return [x_text, y]

# 创建batch迭代模块

def batch_iter(data, batch_size, num_epochs, shuffle=True): # shuffle=True洗牌

"""

    Generates a batch iterator for a dataset.

"""

    # 每次只输出shuffled_data[start_index:end_index]这么多

    data = np.array(data)

    data_size = len(data)

    num_batches_per_epoch = int((len(data)-1)/batch_size) + 1 # 每一个epoch有多少个batch_size

    for epoch in range(num_epochs):

        # Shuffle the data at each epoch

        if shuffle:

            shuffle_indices = np.random.permutation(np.arange(data_size)) # 洗牌

            shuffled_data = data[shuffle_indices]

        else:

            shuffled_data = data

        for batch_num in range(num_batches_per_epoch):

            start_index = batch_num * batch_size # 当前batch的索引开始

            end_index = min((batch_num + 1) * batch_size, data_size) # 判断下一个batch是不是超过最后一个数据了

            yield shuffled_data[start_index:end_index]

text_cnn.py 源码及分析

import tensorflow as tf

import numpy as np

# 定义CNN网络实现的类

class TextCNN(object):

"""

    A CNN for text classification.

    Uses an embedding layer, followed by a convolutional, max-pooling and softmax layer.

"""

    def __init__(self, sequence_length, num_classes, vocab_size,

                 embedding_size, filter_sizes, num_filters, l2_reg_lambda=0.0): # 把train.py中TextCNN里定义的参数传进来

        # Placeholders for input, output and dropout

        self.input_x = tf.placeholder(tf.int32, [None, sequence_length], name="input_x") # input_x输入语料,待训练的内容,维度是sequence_length,"N个词构成的N维向量"

        self.input_y = tf.placeholder(tf.float32, [None, num_classes], name="input_y") # input_y输入语料,待训练的内容标签,维度是num_classes,"正面 || 负面"

        self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob") # dropout_keep_prob dropout参数,防止过拟合,训练时用

        # Keeping track of l2 regularization loss (optional)

        l2_loss = tf.constant(0.0) # 先不用，写0

        # Embedding layer

        # 指定运算结构的运行位置在cpu非gpu,因为"embedding"无法运行在gpu

        # 通过tf.name_scope指定"embedding"

        with tf.device('/cpu:0'), tf.name_scope("embedding"): # 指定cpu

            self.W = tf.Variable(tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0), name="W") # 定义W并初始化

            self.embedded_chars = tf.nn.embedding_lookup(self.W, self.input_x)

            self.embedded_chars_expanded = tf.expand_dims(self.embedded_chars, -1) # 加一个维度，转换为4维的格式

        # Create a convolution + maxpool layer for each filter size

        pooled_outputs = []

        # filter_sizes卷积核尺寸,枚举后遍历

        for i, filter_size in enumerate(filter_sizes):

            with tf.name_scope("conv-maxpool-%s" % filter_size):

                # Convolution Layer

                filter_shape = [filter_size, embedding_size, 1, num_filters] # 4个参数分别为filter_size高h，embedding_size宽w，channel为1，filter个数

                W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1), name="W") # W进行高斯初始化

                b = tf.Variable(tf.constant(0.1, shape=[num_filters]), name="b") # b给初始化为一个常量

                conv = tf.nn.conv2d(

                    self.embedded_chars_expanded,

W,

                    strides=[1, 1, 1, 1],

                    padding="VALID", # 这里不需要padding

                    name="conv")

                # Apply nonlinearity 激活函数

                # 可以理解为,正面或者负面评价有一些标志词汇,这些词汇概率被增强，即一旦出现这些词汇,倾向性分类进正或负面评价,

                # 该激励函数可加快学习进度，增加稀疏性,因为让确定的事情更确定,噪声的影响就降到了最低。

                h = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu")

                # Maxpooling over the outputs

                # 池化

                pooled = tf.nn.max_pool(

h,

                    ksize=[1, sequence_length - filter_size + 1, 1, 1], # (h-filter+2padding)/strides+1=h-f+1

                    strides=[1, 1, 1, 1],

                    padding='VALID', # 这里不需要padding

                    name="pool")

                pooled_outputs.append(pooled)

        # Combine all the pooled features

        num_filters_total = num_filters * len(filter_sizes)

        self.h_pool = tf.concat(3, pooled_outputs)

        self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total]) # 扁平化数据，跟全连接层相连

        # Add dropout

        # drop层,防止过拟合,参数为dropout_keep_prob

        # 过拟合的本质是采样失真,噪声权重影响了判断，如果采样足够多,足够充分,噪声的影响可以被量化到趋近事实,也就无从过拟合。

        # 即数据越大,drop和正则化就越不需要。

        with tf.name_scope("dropout"):

            self.h_drop = tf.nn.dropout(self.h_pool_flat, self.dropout_keep_prob)

        # Final (unnormalized) scores and predictions

        # 输出层

        with tf.name_scope("output"):

            W = tf.get_variable(

                "W",

                shape=[num_filters_total, num_classes], #前面连扁平化后的池化操作

                initializer=tf.contrib.layers.xavier_initializer()) # 定义初始化方式

            b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name="b")

            # 损失函数导入

            l2_loss += tf.nn.l2_loss(W)

            l2_loss += tf.nn.l2_loss(b)

            # xw+b

            self.scores = tf.nn.xw_plus_b(self.h_drop, W, b, name="scores") # 得分函数

            self.predictions = tf.argmax(self.scores, 1, name="predictions") # 预测结果

        # CalculateMean cross-entropy loss

        with tf.name_scope("loss"):

            # loss，交叉熵损失函数

            losses = tf.nn.softmax_cross_entropy_with_logits(logits=self.scores, labels=self.input_y)

            self.loss = tf.reduce_mean(losses) + l2_reg_lambda * l2_loss

        # Accuracy

        with tf.name_scope("accuracy"):

            # 准确率，求和计算算数平均值

            correct_predictions = tf.equal(self.predictions, tf.argmax(self.input_y, 1))

            self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy")

这里写图片描述

以上就是Tensorflow深度学习CNN实现英文文本分类的详细内容

原文链接：https://blog.csdn.net/lwplwf/article/details/72723140

Tensorflow深度学习使用CNN分类英文文本

目录

前言

源码与数据

源码

数据

train.py 源码及分析

data_helpers.py 源码及分析

text_cnn.py 源码及分析