赞
踩
说先看一下这个图,它大体介绍了CNN的自然语言处理流程:
1.首先每个单词对应一行,d=5表示分了5个维度,一般是分128维,300维之类的,这里为了方便,用d=5。
这样的话矩阵就是7*5
2.然后第一步进行卷积的操作,分别使用了四行的卷积核两个,三行的卷积核两个,两行的卷积核两个。然后分别对7*5的矩阵进行卷积,对于7*5的话,4*5放上去可以向下移动4次,所以产生了4*1矩阵(feature map),3*5的则可以移动5次,所以得到5*1的矩阵,同理,2*5的得到6*1的矩阵。
3.然后第二步进行池化操作,图中使用的是max pooling,是最大值池化,所以前面产生的两个4*1的feature map分别取最大的元素,组成一个2*1的矩阵。同理前面产生的5*1、6*1的矩阵也要执行该操作。最终6个feature map产生了3个池化之后的矩阵
4.然后再把3个池化后的矩阵拼接起来形成6*1的feature vector。
5.然后这是一个全连接层,假如我们需要分两个类,则对这6个神经元进行分类,产生两个神经元作为输出。
下面是代码实现,
第一步,各种超参数设定
- 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
- import pickle
-
- # Parameters
- # ==================================================
-
- # Data loading params
- #validation数据集占比 0.1 百分之十
- tf.flags.DEFINE_float("dev_sample_percentage", .1, "Percentage of the training data to use for validation")
- #正样本 电影正面评论
- tf.flags.DEFINE_string("positive_data_file", "./data/rt-polaritydata/rt-polarity.pos", "Data source for the positive data.")
- #负样本 电影负面评论
- tf.flags.DEFINE_string("negative_data_file", "./data/rt-polaritydata/rt-polarity.neg", "Data source for the negative data.")
-
- # Model Hyperparameters
- #词向量长度
- tf.flags.DEFINE_integer("embedding_dim", 128, "Dimensionality of character embedding (default: 128)")
- #卷积核大小 使用3 4 5行的卷积核,列肯定是128列
- tf.flags.DEFINE_string("filter_sizes", "3,4,5", "Comma-separated filter sizes (default: '3,4,5')")
- #每一种卷积核的个数
- tf.flags.DEFINE_integer("num_filters", 64, "Number of filters per filter size (default: 128)")
- #dropout参数
- tf.flags.DEFINE_float("dropout_keep_prob", 0.5, "Dropout keep probability (default: 0.5)")
- #l2正则化参数
- tf.flags.DEFINE_float("l2_reg_lambda", 0.0005, "L2 regularization lambda (default: 0.0)")
-
- # Training parameters
- #批次大小,就是每个批次有多少行的数据
- tf.flags.DEFINE_integer("batch_size", 64, "Batch Size (default: 64)")
- #迭代周期
- tf.flags.DEFINE_integer("num_epochs", 6, "Number of training epochs (default: 200)")
- #多少step测试一次 每运算多少次做一个测试
- tf.flags.DEFINE_integer("evaluate_every", 100, "Evaluate model on dev set after this many steps (default: 100)")
- #每多少次保存一次模型
- tf.flags.DEFINE_integer("checkpoint_every", 100, "Save model after this many steps (default: 100)")
- #最多保存5个模型,如果满了,每新增一个就删除一个,只保留最新的5个模型
- tf.flags.DEFINE_integer("num_checkpoints", 5, "Number of checkpoints to store (default: 5)")
- # Misc Parameters
- #tensorflow会自动选择一个存在并且支持的设备上来运行operation
- tf.flags.DEFINE_boolean("allow_soft_placement", True, "Allow device soft device placement")
- #获取你的operations核Tensor被指派到哪个设备上运行
- tf.flags.DEFINE_boolean("log_device_placement", False, "Log placement of ops on devices")
- #flags解析
- FLAGS = tf.flags.FLAGS
- FLAGS._parse_flags()
- #打印所有参数
- print("\nParameters:")
- for attr, value in sorted(FLAGS.__flags.items()):
- print("{}={}".format(attr.upper(), value))
- print("")
注释讲的很清楚了,这里不过多描述
第二步
- # Data Preparation
- # ==================================================
-
- # Load data
- print("Loading data...")
- #把正样本和负样本传入
- x_text, y = data_helpers.load_data_and_labels(FLAGS.positive_data_file, FLAGS.negative_data_file)
-
- # Build vocabulary
- #一行数据最多的词汇数 因为我们要保证句子的尺寸是一样的,即长度是一样的。
- max_document_length = max([len(x.split(" ")) for x in x_text])
- #词典
- vocab_processor = learn.preprocessing.VocabularyProcessor(max_document_length)
- # 其中VocabularyProcessor(max_document_length,min_frequency=0,vocabulary=None, tokenizer_fn=None)的构造函数中有4个参数
- # max_document_length是文档的最大长度。如果文本的长度大于最大长度,那么它会被剪切,反之则用0填充
- # min_frequency词频的最小值,出现次数>最小词频 的词才会被收录到词表中
- # vocabulary CategoricalVocabulary 对象,不太清楚使用方法
- # tokenizer_fn tokenizer function,讲句子或给定文本格式 token化得函数,可以理解为分词函数
- #把原来的句子进行特殊处理,单词赋予编号
- x = np.array(list(vocab_processor.fit_transform(x_text)))
- # 如果长度低于56自动补0
- # x:[
- # [4719 59 182 34 190 804 0 0 0 0
- # 0 0 0 0 0……]
- # [129 ……]
- # ]
-
- print("x_shape:", x.shape)
- print("y_shape:", y.shape)
- # Randomly shuffle data
- #对数据进行打乱的操作
- np.random.seed(10)
- shuffle_indices = np.random.permutation(np.arange(len(y)))
- x_shuffled = x[shuffle_indices]
- y_shuffled = y[shuffle_indices]
-
- # Split train/test set
- # TODO: This is very crude, should use cross-validation
- #切分 数据集分为两部分 90%训练 10验证
- 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)))
- print("x:",x_train[0:5])
- print("y:",y_train[0:5])
这里调用了data_helpers.py里的写好的一个函数来引入正样本和负样本,我们看一下这个函数是怎么写的
- 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_examples = list(open(positive_data_file, "r", encoding='utf-8').readlines())
- #去掉头尾的空格 形状['I like english','how are you',……]
- positive_examples = [s.strip() for s in positive_examples]
- negative_examples = list(open(negative_data_file, "r", encoding='utf-8').readlines())
- negative_examples = [s.strip() for s in negative_examples]
- # Split by words
- #合并
- x_text = positive_examples + negative_examples
- #句子预处理
- #['句子','句子','']
- x_text = [clean_str(sent) for sent in x_text]
- # Generate labels
- #生成labels,因为就分两个类,正面评论,负面评论
- positive_labels = [[0, 1] for _ in positive_examples]#形状[[0,1],[0,1],……]
- negative_labels = [[1, 0] for _ in negative_examples]#形状[[1,0],[1,0],……]
- y = np.concatenate([positive_labels, negative_labels], 0)#[[0,1],[0,1],……,[1,0],[1,0]]
- return [x_text, y]
额,都是简单的基本操作,这些东西我觉着自己看看就行,注释里也写的很清楚了,我们的目标就是得到句子以及对应的标签
这个函数返回的就是我们处理好的数据,格式:[['句子',……],[[0,1],[0,1],……,[1,0],[1,0],……]
然后计算所有句子中,单词数最多是几,方便padding操作。然后就是定义词典以及进行padding补零操作,这里用到了learn.preprocessing.VocabularyProcessor函数,很强。
然后将数据集进行切分,9比1。
最后就是训练喽~
- # 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():
- cnn = TextCNN(
- sequence_length=x_train.shape[1],#sequence_length:最长词汇数
- num_classes=y_train.shape[1],#num_classes:分类数
- vocab_size=len(vocab_processor.vocabulary_),#vocab_size:总词汇数
- embedding_size=FLAGS.embedding_dim,#embedding_size:词向量长度
- filter_sizes=list(map(int, FLAGS.filter_sizes.split(","))),#filter_sizes:卷积核的尺寸3 4 5
- num_filters=FLAGS.num_filters,#num_filters:卷积核的数量
- l2_reg_lambda=FLAGS.l2_reg_lambda)#l2_reg_lambda_l2正则化系数
-
- # Define Training procedure
- #其实就是统计当前训练的步数
- global_step = tf.Variable(0, name="global_step", trainable=False)
- #优化器
- optimizer = tf.train.AdamOptimizer(1e-3)
- #计算梯度 ——minimize()的第一部分
- grads_and_vars = optimizer.compute_gradients(cnn.loss)
- #将计算出来的梯度应用到变量上,是函数minimize()的第二部分
- #返回一个应用指定的梯度的操作operation,对global_step做自增操作
- ##综上所述,上面三行代码,就相当于global_step = tf.Variable(0, name="global_step", trainable=False).minimize(cnn.loss)
- train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)
-
- # Keep track of gradient values and sparsity (optional)
- #保存变量的梯度值
- grad_summaries = []
- for g, v in grads_and_vars:#g是梯度,v是变量
- 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
- #定义输出路径
- 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
- #记录一下loss值和准确率的值
- loss_summary = tf.summary.scalar("loss", cnn.loss)
- acc_summary = tf.summary.scalar("accuracy", cnn.accuracy)
-
- # Train Summaries
- #训练部分
- #merge一下
- train_summary_op = tf.summary.merge([loss_summary, acc_summary, grad_summaries_merged])
- #路径定义
- train_summary_dir = os.path.join(out_dir, "summaries", "train")
- #写入
- train_summary_writer = tf.summary.FileWriter(train_summary_dir, sess.graph)
-
- # Dev summaries
- #测试部分
- 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
- #保存模型的路径
- 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)
- #保存模型,最多保存5个
- saver = tf.train.Saver(tf.global_variables(), max_to_keep=FLAGS.num_checkpoints)
-
- # Write vocabulary
- vocab_processor.save(os.path.join(out_dir, "vocab"))
-
- # Initialize all variables
- sess.run(tf.global_variables_initializer())
-
- 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
- }
- _, 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()
- print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
- train_summary_writer.add_summary(summaries, step)
-
- 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
- }
- 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...
- for batch in batches:
- x_batch, y_batch = zip(*batch)
- train_step(x_batch, y_batch)
- current_step = tf.train.global_step(sess, global_step)
- #测试
- if current_step % FLAGS.evaluate_every == 0:
- print("\nEvaluation:")
- dev_step(x_dev, y_dev, writer=dev_summary_writer)
- print("")
- #保存模型
- if current_step % FLAGS.checkpoint_every == 0:
- path = saver.save(sess, checkpoint_prefix, global_step=current_step)
- print("Saved model checkpoint to {}\n".format(path))
创建一个会话,然后我们看一下TextCNN干了嘛,
text_cnn.py
- import tensorflow as tf
- import numpy as np
-
-
- class TextCNN(object):
- """
- A CNN for text classification.
- Uses an embedding layer, followed by a convolutional, max-pooling and softmax layer.
- """
- #sequence_length:最长词汇数
- #num_classes:分类数
- #vocab_size:总词汇数
- #embedding_size:词向量长度
- #filter_sizes:卷积核的尺寸3 4 5
- #num_filters:卷积核的数量
- #l2_reg_lambda_l2正则化系数
- def __init__(
- self, sequence_length, num_classes, vocab_size,
- embedding_size, filter_sizes, num_filters, l2_reg_lambda=0.0):
-
- # Placeholders for input, output and dropout
- self.input_x = tf.placeholder(tf.int32, [None, sequence_length], name="input_x")
- self.input_y = tf.placeholder(tf.float32, [None, num_classes], name="input_y")
- #dropout系数
- self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
-
- # Keeping track of l2 regularization loss (optional)
- l2_loss = tf.constant(0.0)
-
- # Embedding layer
- with tf.device('/cpu:0'), tf.name_scope("embedding"):
- #vocab_size词典大小
- self.W = tf.Variable(
- tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0),
- name="W")
- #[batch_size,sequeue_length,embedding_size]
- self.embedded_chars = tf.nn.embedding_lookup(self.W, self.input_x)
- #添加一个维度,[batch_size,sequence_length,embedding_size,1]有点类似于图片哦~
- self.embedded_chars_expanded = tf.expand_dims(self.embedded_chars, -1)
-
- # Create a convolution + maxpool layer for each filter size
- pooled_outputs = []
- #三种不同尺寸的卷积,所以用了一个循环
- 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]
- #下面是卷积的参数
- W = tf.Variable(tf.truncated_normal(filter_shape, stddev=0.1), name="W")
- b = tf.Variable(tf.constant(0.1, shape=[num_filters]), name="b")
- #卷积操作
- conv = tf.nn.conv2d(
- self.embedded_chars_expanded,#需要卷积的矩阵
- W,#权值
- strides=[1, 1, 1, 1],#步长
- padding="VALID",
- name="conv")
- # Apply nonlinearity
- #激活函数
- 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],
- strides=[1, 1, 1, 1],
- padding='VALID',
- name="pool")
- pooled_outputs.append(pooled)
-
- # Combine all the pooled features
- #卷积核的数量*卷积核的尺寸种类
- num_filters_total = num_filters * len(filter_sizes)
- #三个池化层的输出拼接起来
- self.h_pool = tf.concat(pooled_outputs, 3)
- #转化为一维的向量
- self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total])
-
- # Add dropout
- with tf.name_scope("dropout"):
- self.h_drop = tf.nn.dropout(self.h_pool_flat, self.dropout_keep_prob)
-
- # Final (unnormalized) scores and predictions
- #全连接层
- with tf.name_scope("output"):
- W = tf.get_variable(
- "W",
- shape=[num_filters_total, num_classes],#[64*3,2]
- initializer=tf.contrib.layers.xavier_initializer())
- b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name="b")
- #l2正则化
- l2_loss += tf.nn.l2_loss(W)
- l2_loss += tf.nn.l2_loss(b)
- self.scores = tf.nn.softmax(tf.nn.xw_plus_b(self.h_drop, W, b, name="scores"))
- #最终预测值
- self.predictions = tf.argmax(self.scores, 1, name="predictions")
-
- # Calculate mean cross-entropy loss
- with tf.name_scope("loss"):
- 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"):
- correct_predictions = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
- self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy")
在这个类的对象初始化的过程中,我们传入了几个超参数,分别是最长词汇数,分类数,总词汇数,词向量长度,卷积核尺寸,卷积核数量,l2正则化系数。
然后设置了几个占位符,然后重点嵌入层来了
- with tf.device('/cpu:0'), tf.name_scope("embedding"):
- #vocab_size词典大小
- self.W = tf.Variable(
- tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0),
- name="W")
- #[batch_size,sequeue_length,embedding_size]
- self.embedded_chars = tf.nn.embedding_lookup(self.W, self.input_x)
- #添加一个维度,[batch_size,sequence_length,embedding_size,1]有点类似于图片哦~
- self.embedded_chars_expanded = tf.expand_dims(self.embedded_chars, -1)
W就是所有词的词向量矩阵,然后embedded_chars是输入句子所对应的的词向量矩阵,返回的就是[batch_size,sequence_length,embedding_size],然后手动为其添加了一个维度。
然后又是一个重点:卷积池化层。
我们先看一下tf.nn.conv2d函数
- def conv2d(x,W):
- #2d是二维的意思
- #x是一个tensor,形状是[batch,in_height,in_width,in_channels]NHWC关系,分别是批次大小(本例batch_size=100),图片高度,图片宽度,通道数(黑白照片是1,彩色是3)
- #w是一个滤波器,tensor,形状是[filter_height,filter_width,in_channels,out_channels],滤波器长,宽,输入和输出通道数
- #步长参数,strides[0]=strides[3]=1,strides[1]代表x方向的步长,strides[2]代表y方向的步长
- #padding:一个字符串,要么是'SAME'要么是'VALID',对应两种卷积方法,前者补零,后者不会超出平面外部
- ————————————————
- 版权声明:本文为CSDN博主「CtrlZ1」的原创文章,遵循CC 4.0 by-sa版权协议,转载请附上原文出处链接及本声明。
- 原文链接:https://blog.csdn.net/qq_41076797/article/details/99212303
我们看tf.nn.conv2d执行卷积,输入需要卷积的矩阵self.embedded_chars_expanded,要求它的格式是[batch_size,in_height,in_width,in_channels],这就是为什么要加一个维度的原因,这里和照片有点相似,其实这个in_height和in_width很耐人寻味,前一个是传入的句子的最大长度,后面是对应的句子中每个词的词向量维度,从矩阵来看正好是高和宽。w是一个[卷积核尺寸,词向量长度,1,卷积核个数]的矩阵,步长是1,1,padding方案是valid,即不进行补零,那样的话矩阵大小会发生变化。根据w,我们知道卷积窗口大小是卷积核尺寸*词向量长度,由于输入矩阵宽度也是词向量长度,所以只能向下卷积,所以卷积之后的矩阵格式为[bath_size,in_height-卷积核尺寸+1,1,num_filter]即[64,最长句子单词数-2,1,64]。然后由于卷积核的个数,有多少个卷积核,返回多少个这种矩阵(feature map)。
然后进行激活函数。
不用过多解释了吧。 返回依然是[bath_size,in_height-卷积核尺寸+1,1,num_filter]=[bath_size,sequence_length -3or4or5+1,1,64]即[64,最长句子单词数-2or3or4,1,64]
然后进行池化操作,池化窗口尺寸是(sequence_length - filter_size + 1)*1,和我们返回的矩阵是一样的,这样的话池化之后返回的矩阵就是[batch_size,1,1,64]
然后把三次池化后的输出拼接起来,转化为一维向量[batch_size,64*3],然后dropout,然后又是重要的一层,全连接层:
- #全连接层
- with tf.name_scope("output"):
- W = tf.get_variable(
- "W",
- shape=[num_filters_total, num_classes],#[64*3,2]
- initializer=tf.contrib.layers.xavier_initializer())
- b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name="b")
- #l2正则化
- l2_loss += tf.nn.l2_loss(W)
- l2_loss += tf.nn.l2_loss(b)
- self.scores = tf.nn.softmax(tf.nn.xw_plus_b(self.h_drop, W, b, name="scores"))
- #最终预测值
- self.predictions = tf.argmax(self.scores, 1, name="predictions")
-
- # Calculate mean cross-entropy loss
- with tf.name_scope("loss"):
- 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"):
- correct_predictions = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
- self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy")
经过softmax之后返回的scores是一个[batch_size,2]的矩阵,和真实值input_y比较可以得到损失值losses,然后我们分别计算平均损失值和准确率。到此就结束了。
再回头看就很简单了,什么优化器啊,什么计算梯度啊,什么保存模型啊,等等就没啥可说的了,都是常规操作。
对了,附上代码链接~
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。