论文Convolutional Naural Networks for Sentence Classification阅读（一）----模型实现篇

之前把模型介绍了，也自己写了代码实现，但一直拖着没有吧博客写出来。最近要做句子相似度相关的项目，想要自己写一个模型，想到了这个简单的模型，才惊觉博客一直没写，所以现在补课，也是一个复习TensorFlow实现模型的过程。本文主要参考了Denny Britz的文章和代码。

1.数据及预处理
实验所用的数据集是Movie Review data from Rotten Tomatoes，即MR电影评论数据，其中包含10662条评论，一半正面评论，一半负面。共包含18758个单词（vocab_size），最长的评论有56个单词（Padding，sequence_len），保存在data目录下的rt-polarity.neg和rt-poliarity.pos文件中。这里我们使用10%作为验证集，剩下的作为训练集。数据预处理部分写在data_helpers.py文件中，其实代码很简单，这里仅对load_data_and_labels函数进行介绍：

def load_data_and_labels(positive_data_file, negative_data_file):
    #从文件中读取数据
    # 将file里所有的句子转化成一个list，每一个句子是list的一个元素
    positive_examples = list(open(positive_data_file, "r", encoding="utf-8").readlines())
    # 将每个句子两端的转义字符比如换行和空格等删除，仍然是一个list
    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]

    #positive和negative两个列表合成一个
    x_text = positive_examples + negative_examples
    x_text = [clean_str(sent) for sent in x_text]

    #正面评价的label是[0,1],负面评价的label是[1,0]
    positive_labels = [[0, 1] for _ in positive_examples]
    negative_labels = [[1, 0] for _ in negative_examples]
    #y是二维数组的形式是这样的：前正例数量的label是[0,1]，后负例数量的label是[1,0]
    #这样y中label就和x_text中的句子一一对应了
    y = np.concatenate([positive_labels, negative_labels], 0)
    return [x_text, y]

然后我们将数据载入，构建vocabulary，并且将每个句子转化为单词的索引列表。这部分代码在train.py中实现：

# Load data
print("loading data....")
x_text, y = data_helpers.load_data_and_labels(FLAGS.positive_data_file, FLAGS.negative_data_file)

# build vocabulary
# 获取句子最大长度，用于padding，这里的值是56
max_document_length = max([len(x.split(" ")) for x in x_text])
# 调用tf内部函类VocabularyProcessor，其会读取x_text，按照词语出现顺序构建vocabulary，并给每个单词以索引
# 例如第一个出现的单词编号为1，第二个出现的为2，以此类推。用0进行padding
# 然后返回的x是形如[[1,2,3,4...],...,[55,66,777...]]的嵌套列表。
# 内列表代表每个句子，其值是评论中每个词在vocabulary中的索引。所以x是10662*56维
# 注意vocabulary和x并不是一件事情
vocab_processor = learn.preprocessing.VocabularyProcessor(max_document_length)
x = np.array(list(vocab_processor.fit_transform(x_text)))
print("x::", x)

# randomly shuffle data
np.random.seed(10)
# np.arange创建等差数组。默认起点0，终点len(y)，步长1
# permutation会返回一个根据随机数打乱顺序的新数组，不改变原数组
shuffle_indices = np.random.permutation(np.arange(len(y)))
# shuffle后的xy仍然是以一一对应的
x_shuffled = x[shuffle_indices]
y_shuffled = y[shuffle_indices]

# split train/test set 构建训练集和测试集,10%作为验证集，剩下的作为训练集
# -1取负号是从数组最后开始倒着取的意思
dev_sample_index = -1 * int(FLAGS.dev_sample_percentage * float(len(y)))
print(len(y))
print(dev_sample_index)
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:]

del x, y, x_shuffled, y_shuffled

print("Vocabulary Size: {:d}".format(len(vocab_processor.vocabulary_)))
print("Train/Dev split: {:d}/{:d}".format(len(y_train), len(y_dev)))

2.构建CNN网络模型
代码在text_cnn.py中，大家可以参考上一篇模型介绍的博客。 因为代码中的注释已经比较明确了 ，所以这里不做过多的介绍，大家进去看看注释就好~

class TextCNN(object):
    """
    embedding layer , followed by a convolutional ,max-pooling and softmax layer
    """
    def __init__(
            # 每个句子的长度sequence_length=56,
            # 类别数num_classes=2
            # 词汇表大小vocab_size=18758
            # 词向量维度embedding_siza=128
            # 卷积核纵向长度filter_size，也就是一个卷积核一次所能覆盖的单词数量，这里是[3,4,5]三个
            # 卷积核个数num_filters=128
            # 正则化强度l2_reg_lambda默认0
            self, sequence_length, num_classes, vocab_size,
            embedding_size, filter_sizes, num_filters, l2_reg_lambda=0.0):

        # 输入 输出 dropout 的 占位符
        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")
        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
        # 根据输入x的单词在vocabulary中的索引lookup得到每个单词对应的词向量
        # 这里没有使用Word2Vec，而是随机初始化的。也就是论文中提到的non-static
        with tf.device('/cpu:0'), tf.name_scope("embedding"):
            self.W = tf.Variable(
                tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0),
                name="W")
            # input_x是None*56维的，lookup时是对56个单词都去查找一个词向量
            # 因此生成的embedded_chars是None*56*128
            self.embedded_chars = tf.nn.embedding_lookup(self.W, self.input_x)
            # 增加一个维度，-1表示最后一维，默认增加维度中的数值是1
            # 因为卷积操作conv2d()需要输入的是四维数据，分别代表着批处理大小、宽度、高度、通道数。
            # 而embedded_chars只有前三维，所以需要添加一维，设为1。变为：[None, sequence_length, embedding_size, 1]
            # 注意这里embedded_chars_expanded 是 None*56*128*1
            self.embedded_chars_expanded = tf.expand_dims(self.embedded_chars, -1)

        # 为每一个filter_size构建卷积层和最大池化层，将最后的结果合并成一个大的特征向量
        pooled_outputs = []
        for i, filter_size in enumerate(filter_sizes):
            with tf.name_scope("conv-maxpool-%s" % filter_size):
                # 构建卷积核尺寸 输入和输出channel分别为1和num_filters(128)
                # filter_shape = [3/4/5, 128, 1, 128]
                filter_shape = [filter_size, embedding_size, 1, num_filters]
                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")
                # 对于sequence_length个句子组成的图片而言，
                # 每个经过卷积层输出的形状[1, sequence_length - filter_size + 1, 1, 1]
                conv = tf.nn.conv2d(
                    self.embedded_chars_expanded,
                    W,
                    strides=[1, 1, 1, 1],
                    padding="VALID",
                    name="conv"
                )
                # 非线性激活函数Relu ：relu(W*x + b)
                h = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu")
                # 池化
                # 经过池化层输出的shape:[None, 1, 1, num_filters]
                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最终是一个长度是3的列表，每个元素都是[None, 1, 1, 128]的tensor张量
                pooled_outputs.append(pooled)

        # Combine all the pooled features
        num_filters_total = num_filters * len(filter_sizes)
        # 将pooled_outputs在第四个维度上进行合并，变成一个[None,1,1,384]Tensor张量
        self.h_pool = tf.concat(pooled_outputs, 3)
        # 展成二维tensor [None, 384]
        self.h_pool_flat = tf.reshape(self.h_pool, [-1, num_filters_total])

        # Add dropout
        with tf.name_scope("drop_out"):
            self.h_drop = tf.nn.dropout(self.h_pool_flat, self.dropout_keep_prob)

        # Final (unnormalized) scores and predictions
        # 全连接层计算输出向量(w*h+b)和预测(scores向量中的最大值即为预测结果)
        with tf.name_scope("output"):
            W = tf.get_variable(
                "W",
                shape=[num_filters_total, num_classes],
                initializer=tf.contrib.layers.xavier_initializer()
            )
            b = tf.Variable(tf.constant(0.1, shape=[num_classes]), name="b")
            l2_loss += tf.nn.l2_loss(W)
            l2_loss += tf.nn.l2_loss(b)
            self.scores = tf.nn.xw_plus_b(self.h_drop, W, b, name="scores")
            self.predictions = tf.argmax(self.scores, 1, name="prediction")

        # Calculate mean cross-entropy loss 计算scores和input_y的交叉熵损失函数
        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 计算准确度，预测和真实标签相同即为正确
        with tf.name_scope("accuracy"):
            # 返回正确或者错误的布尔值
            correct_predictions = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
            # cast将布尔值转化为float ,True是1
            self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"), name="accuracy")

对上述代码中使用的常用函数做个总结：
1，tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, data_format=None,
name=None)：执行卷积操作

input是一个四维Tensor：[batch, in_height, in_width, in_channels]
filter是一个四维Tensor：[filter_height, filter_width, in_channels,
out_channels]
strides卷积步长四维Tensor：必须满足[1, stride, stride, 1]的格式，每一位代表在输入上每维移动的步长。
padding：“SAME”或者“VALID”，意味着宽卷积和窄卷积。

2，tf.random_uniform(shape, minval=0, maxval=None, dtype=tf.float32, seed=None, name=None)输出服从均匀分布[minval, maxval)的随机初始化函数。shape为要输出结果尺寸

3，tf.truncated_normal(shape, mean=0.0, stddev=1.0, dtype=tf.float32, seed=None, name=None)输出服从截断正态随机分布的随机初始化函数。生成的值会遵循一个指定了平均值和标准差的正态分布，只保留两个标准差以内的值，超出的值会被弃掉重新生成。

3.训练代码
代码里的注释已经比较详细了，所以这里不做过多介绍

with tf.Graph().as_default():
    session_conf = tf.ConfigProto(
        # 允许TensorFlow回退到特定设备，并记录程序运行的设备信息
        # log_device_placement: 是否打印设备分配日志
        # allow_soft_placement：如果你指定的设备不存在，允许TF自动分配设备
        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],
            num_classes=y_train.shape[1],
            vocab_size=len(vocab_processor.vocabulary_),
            embedding_size=FLAGS.embedding_dim,
            filter_sizes=list(map(int, FLAGS.filter_sizes.split(","))),
            num_filters=FLAGS.num_filters,
            l2_reg_lambda=FLAGS.l2_reg_lambda
        )
        # Define Training procedure
        # trainable=False表明该参数虽然是Variable但并不属于网络运行参数，无需计算梯度并更新
        global_step = tf.Variable(0, name="global_step", trainable=False)
        # 优化器的学习速率是0.001
        optimizer = tf.train.AdamOptimizer(1e-3)
        # compute_gradients返回一个各个变量对于loss的梯度的列表
        grads_and_vars = optimizer.compute_gradients(cnn.loss)
        # 把梯度Apply到变量上面去。其实就是按照梯度下降的方式加到上面去
        train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)

        # 将一些想要在tensorboard中展示的网络参数记录下来，保存到summary中
        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)
                # tf.nn.zero_fraction返回0在value中的小数比例
                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
        loss_summary = tf.summary.scalar("loss", cnn.loss)
        acc_summary = tf.summary.scalar("accuracy", cnn.accuracy)

        # Train summaries
        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
        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 assume this directory already exists so we need to create it
        checkpoint_dir = os.path.abspath(os.path.join(out_dir, "checkpoint"))
        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)

        # 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
            }
            # sess.run()运行一次网络优化，并将相应信息输出
            _, 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)

        # 验证集效果检测
        def dev_step(x_batch, y_batch, writer=None):
            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
        # zip 将x_train和y_train做成一个一一对应的列表
        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) #将batch解压
            train_step(x_batch, y_batch)
            current_step = tf.train.global_step(sess, global_step)
            print("current_step:", current_step)
            if current_step % FLAGS.evaluate_every == 0:
                print("\nEvaluation:")
                dev_step(x_dev, y_dev, writer=dev_summary_writer)
                print("")
            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))