一、TensorFlow-word2vec

连续词袋模型（CBOW）：根据词的上下文词汇来预测目标词汇，例如上下文词汇是“今天早餐吃__”，要预测的目标词汇可能
是“面包“；

Skip-Gram模型：Skip-Gram模型刚好和CBOW相反，它是通过目标词汇来预测上下文词汇。例如目标词汇是“早餐”
，上下文词汇可能是“今天”和“吃面包”：

Word2vec模型
训练Word2vec模型我们通常可以选择使用噪声对比估计（Noise Contrastive Estimation）：NCE使用的方法是把上下文h对应地正确的目标词汇标记为正样本（D=1），然后再抽取一些错误的词汇作为负样本（D=0）。然后最大化目标函数的值：

当真实的目标单词被分配到较高的概率，同时噪声单词的概率很低时，目标函数也就达到最大值了。计算这个函数时，只需要计算挑选出来的k个噪声单词，而不是整个语料库。所以训练速度会很快。
Word2Vec图形化:

案例代码：

# encoding=utf8  
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import collections
import math
import os
import random
import zipfile

import numpy as np
from six.moves import urllib
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf

# Step 1: Download the data.
url = 'http://mattmahoney.net/dc/'

# 下载数据集
def maybe_download(filename, expected_bytes):
    """Download a file if not present, and make sure it's the right size."""
    if not os.path.exists(filename):
        filename, _ = urllib.request.urlretrieve(url + filename, filename)
    # 获取文件相关属性
    statinfo = os.stat(filename)
    # 比对文件的大小是否正确
    if statinfo.st_size == expected_bytes:
        print('Found and verified', filename)
    else:
        print(statinfo.st_size)
        raise Exception(
            'Failed to verify ' + filename + '. Can you get to it with a browser?')
    return filename

filename = maybe_download('text8.zip', 31344016)

# Read the data into a list of strings.
def read_data(filename):
    """Extract the first file enclosed in a zip file as a list of words"""
    with zipfile.ZipFile(filename) as f:
        data = tf.compat.as_str(f.read(f.namelist()[0])).split()
    return data

# 单词表
words = read_data(filename)

# Data size
print('Data size', len(words))

# Step 2: Build the dictionary and replace rare words with UNK token.
# 只留50000个单词，其他的词都归为UNK
vocabulary_size = 50000

def build_dataset(words, vocabulary_size):
    count = [['UNK', -1]]
    # extend追加一个列表
    # Counter用来统计每个词出现的次数
    # most_common返回一个TopN列表，只留50000个单词包括UNK  
    # c = Counter('abracadabra')
    # c.most_common()
    # [('a', 5), ('r', 2), ('b', 2), ('c', 1), ('d', 1)]
    # c.most_common(3)
    # [('a', 5), ('r', 2), ('b', 2)]
    # 前50000个出现次数最多的词
    count.extend(collections.Counter(words).most_common(vocabulary_size - 1))
    # 生成 dictionary，词对应编号, word:id(0-49999)
    # 词频越高编号越小
    dictionary = dict()
    for word, _ in count:
        dictionary[word] = len(dictionary)
    # data把数据集的词都编号
    data = list()
    unk_count = 0
    for word in words:
        if word in dictionary:
            index = dictionary[word]
        else:
            index = 0  # dictionary['UNK']
            unk_count += 1
        data.append(index)
    # 记录UNK词的数量
    count[0][1] = unk_count
    # 编号对应词的字典
    reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
    return data, count, dictionary, reverse_dictionary

# data 数据集，编号形式
# count 前50000个出现次数最多的词
# dictionary 词对应编号
# reverse_dictionary 编号对应词
data, count, dictionary, reverse_dictionary = build_dataset(words, vocabulary_size)
del words  # Hint to reduce memory.
print('Most common words (+UNK)', count[:5])
print('Sample data', data[:10], [reverse_dictionary[i] for i in data[:10]])

data_index = 0

# Step 3: Function to generate a training batch for the skip-gram model.
def generate_batch(batch_size, num_skips, skip_window):
    global data_index
    assert batch_size % num_skips == 0
    assert num_skips <= 2 * skip_window
    
    batch = np.ndarray(shape=(batch_size), dtype=np.int32)
    labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)
    
    span = 2 * skip_window + 1  # [ skip_window target skip_window ]
    buffer = collections.deque(maxlen=span)
    # [ skip_window target skip_window ]
            # [ skip_window target skip_window ]
                    # [ skip_window target skip_window ]
            
#     [0 1 2 3 4 5 6 7 8 9 ...]
#            t     i  
    # 循环3次
    for _ in range(span):
        buffer.append(data[data_index])
        data_index = (data_index + 1) % len(data)
    # 获取batch和labels
    for i in range(batch_size // num_skips):
        target = skip_window  # target label at the center of the buffer
        targets_to_avoid = [skip_window]
        # 循环2次，一个目标单词对应两个上下文单词
        for j in range(num_skips):
            while target in targets_to_avoid:
                # 可能先拿到前面的单词也可能先拿到后面的单词
                target = random.randint(0, span - 1)
            targets_to_avoid.append(target)
            batch[i * num_skips + j] = buffer[skip_window]
            labels[i * num_skips + j, 0] = buffer[target]
        buffer.append(data[data_index])
        data_index = (data_index + 1) % len(data)
    # Backtrack a little bit to avoid skipping words in the end of a batch
    # 回溯3个词。因为执行完一个batch的操作之后，data_index会往右多偏移span个位置
    data_index = (data_index + len(data) - span) % len(data)
    return batch, labels

# 打印sample data
batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1)
for i in range(8):
    print(batch[i], reverse_dictionary[batch[i]],
        '->', labels[i, 0], reverse_dictionary[labels[i, 0]])

# Step 4: Build and train a skip-gram model.
batch_size = 128
# 词向量维度
embedding_size = 128  # Dimension of the embedding vector.
skip_window = 1       # How many words to consider left and right.
num_skips = 2         # How many times to reuse an input to generate a label.

# We pick a random validation set to sample nearest neighbors. Here we limit the
# validation samples to the words that have a low numeric ID, which by
# construction are also the most frequent.
valid_size = 16     # Random set of words to evaluate similarity on.
valid_window = 100  # Only pick dev samples in the head of the distribution.
# 从0-100抽取16个整数，无放回抽样
valid_examples = np.random.choice(valid_window, valid_size, replace=False) 
# 负采样样本数
num_sampled = 64    # Number of negative examples to sample.

graph = tf.Graph()
with graph.as_default():
    # Input data.
    train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
    train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
    valid_dataset = tf.constant(valid_examples, dtype=tf.int32)

    # Ops and variables pinned to the CPU because of missing GPU implementation
#     with tf.device('/cpu:0'):
        # 词向量
        # Look up embeddings for inputs.
    embeddings = tf.Variable(
        tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
    # embedding_lookup(params,ids)其实就是按照ids顺序返回params中的第ids行
    # 比如说，ids=[1,7,4],就是返回params中第1,7,4行。返回结果为由params的1,7,4行组成的tensor
    # 提取要训练的词
    embed = tf.nn.embedding_lookup(embeddings, train_inputs)

    # Construct the variables for the noise-contrastive estimation(NCE) loss
    nce_weights = tf.Variable(
        tf.truncated_normal([vocabulary_size, embedding_size],
                        stddev=1.0 / math.sqrt(embedding_size)))
    nce_biases = tf.Variable(tf.zeros([vocabulary_size]))

    # Compute the average NCE loss for the batch.
    # tf.nce_loss automatically draws a new sample of the negative labels each
    # time we evaluate the loss.
    loss = tf.reduce_mean(
        tf.nn.nce_loss(weights=nce_weights,
                       biases=nce_biases,
                       labels=train_labels,
                       inputs=embed,
                       num_sampled=num_sampled,   
                       num_classes=vocabulary_size))

    # Construct the SGD optimizer using a learning rate of 1.0.
    optimizer = tf.train.GradientDescentOptimizer(1).minimize(loss)

    # Compute the cosine similarity between minibatch examples and all embeddings.
    norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))
    normalized_embeddings = embeddings / norm
    # 抽取一些常用词来测试余弦相似度
    valid_embeddings = tf.nn.embedding_lookup(
        normalized_embeddings, valid_dataset)
    # valid_size == 16
    # [16,1] * [1*50000] = [16,50000]
    similarity = tf.matmul(
        valid_embeddings, normalized_embeddings, transpose_b=True)

    # Add variable initializer.
    init = tf.global_variables_initializer()

# Step 5: Begin training.
num_steps = 100001
final_embeddings = []

with tf.Session(graph=graph) as session:
    # We must initialize all variables before we use them.
    init.run()
    print("Initialized")

    average_loss = 0
    for step in xrange(num_steps):
        # 获取一个批次的target，以及对应的labels，都是编号形式的
        batch_inputs, batch_labels = generate_batch(
            batch_size, num_skips, skip_window)
        feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}

        # We perform one update step by evaluating the optimizer op (including it
        # in the list of returned values for session.run()
        _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
        average_loss += loss_val

        # 计算训练2000次的平均loss
        if step % 2000 == 0:
            if step > 0:
                average_loss /= 2000
            # The average loss is an estimate of the loss over the last 2000 batches.
            print("Average loss at step ", step, ": ", average_loss)
            average_loss = 0
    
        # Note that this is expensive (~20% slowdown if computed every 500 steps)
        if step % 20000 == 0:
            sim = similarity.eval()
            # 计算验证集的余弦相似度最高的词
            for i in xrange(valid_size):
                # 根据id拿到对应单词
                valid_word = reverse_dictionary[valid_examples[i]]
                top_k = 8  # number of nearest neighbors
                # 从大到小排序，排除自己本身，取前top_k个值
                nearest = (-sim[i, :]).argsort()[1:top_k + 1]
                log_str = "Nearest to %s:" % valid_word
                for k in xrange(top_k):
                    close_word = reverse_dictionary[nearest[k]]
                    log_str = "%s %s," % (log_str, close_word)
                print(log_str)
    # 训练结束得到的词向量
    final_embeddings = normalized_embeddings.eval()

# Step 6: Visualize the embeddings.

def plot_with_labels(low_dim_embs, labels, filename='tsne.png'):
    assert low_dim_embs.shape[0] >= len(labels), "More labels than embeddings"
    # 设置图片大小
    plt.figure(figsize=(15, 15))  # in inches
    for i, label in enumerate(labels):
        x, y = low_dim_embs[i, :]
        plt.scatter(x, y)
        plt.annotate(label,
                 xy=(x, y),
                 xytext=(5, 2),
                 textcoords='offset points',
                 ha='right',
                 va='bottom')

    plt.savefig(filename)

try:
    from sklearn.manifold import TSNE
    import matplotlib.pyplot as plt

    tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000, method='exact')# mac：method='exact'
    # 画500个点
    plot_only = 300
    low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only, :])
    labels = [reverse_dictionary[i] for i in xrange(plot_only)]
    plot_with_labels(low_dim_embs, labels)

except ImportError:
    print("Please install sklearn, matplotlib, and scipy to visualize embeddings.")

二、TensorFlow-CNN解决文本分类：

CNN应用于NLP的任务，处理的往往是以矩阵形式表达的句子或文本。矩阵中的每一行对应于一个分词元素，一般是一个单词，也可以是一个字符。也就是说每一行都是一个词或者字符的向量（比如前面说到的word2vec）。假设我们一共有10个词，每个词都用128维的向量来表示，那么我们就可以得到一个10×128维的矩阵。

参考项目：https://github.com/dennybritz/cnn-text-classification-tf

三、TensorFlow在语音方面的应用案例

声谱图：语音被分为很多帧，每帧语音都对应于一个频谱（通过FFT计算得到），频谱表示频率与能量的关
系；
频谱图：先将一帧语音的频谱通过坐标表示出来，如左图。再将图旋转90度，如中间的图。然后把这些
幅度映射到一个灰度级表示；

spectrogram声谱图：

分离包络和频谱的细节：

Mel频率分析：人类听觉感知实验表明，人类的听觉的感知只聚焦在某些特定的区域，而不是整个频谱包络。Mel频率分析就是基于人类听觉感知实验的。人耳就像一个滤波器组，它只关注某些特定频率的分量，也就是说它只让某些频率的信号通过。并且在低频区域由很多的滤波器，分布比较密集，在高频区域，滤波器比较少，也比较稀疏。

人的听觉系统是一个特殊的非线性系统，它响应不同频率信号的灵明度是不同的。在语音特征的提取上，人类的听觉系统非常好，它不仅能提取出语义信息，而且能提取出说话人的个人特征。所以语音识别系统中能模拟人类听觉感知处理的特点，就有可能提高语音的识别率。MFCC考虑到了人类的听觉特征，将线性频谱映射到基于听觉感知的Mel非线性频谱中。
语音处理流程：

声音分类代码案例：

# coding: utf-8
# 数据集：https://serv.cusp.nyu.edu/projects/urbansounddataset/  
# 
# librosa：https://github.com/librosa/librosa  
# 
# 分类：  
# 0 = air_conditioner  
# 1 = car_horn  
# 2 = children_playing  
# 3 = dog_bark  
# 4 = drilling  
# 5 = engine_idling  
# 6 = gun_shot  
# 7 = jackhammer  
# 8 = siren  
# 9 = street_music  

# In[1]:

import os
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
import time
import librosa # pip install librosa
from tqdm import tqdm # pip install tqdm
import random

# Parameters
# ==================================================

# Data loading params
# validation数据集占比
tf.flags.DEFINE_float("dev_sample_percentage", .2, "Percentage of the training data to use for validation")
# 父目录
tf.flags.DEFINE_string("parent_dir", "audio/", "Data source for the data.")
# 子目录
tf.flags.DEFINE_string("tr_sub_dirs", ['fold1/','fold2/','fold3/'], "Data source for the data.")

# Model Hyperparameters
# 第一层输入，MFCC信号
tf.flags.DEFINE_integer("n_inputs", 40, "Number of MFCCs (default: 40)")
# cell个数
tf.flags.DEFINE_string("n_hidden", 300, "Number of cells (default: 300)")
# 分类数
tf.flags.DEFINE_integer("n_classes", 10, "Number of classes (default: 10)")
# 学习率
tf.flags.DEFINE_integer("lr", 0.005, "Learning rate (default: 0.005)")
# dropout参数
tf.flags.DEFINE_float("dropout_keep_prob", 0.5, "Dropout keep probability (default: 0.5)")

# Training parameters
# 批次大小
tf.flags.DEFINE_integer("batch_size", 50, "Batch Size (default: 50)")
# 迭代周期
tf.flags.DEFINE_integer("num_epochs", 100, "Number of training epochs (default: 100)")
# 多少step测试一次
tf.flags.DEFINE_integer("evaluate_every", 50, "Evaluate model on dev set after this many steps (default: 50)")
# 多少step保存一次模型
tf.flags.DEFINE_integer("checkpoint_every", 500, "Save model after this many steps (default: 500)")
# 最多保存多少个模型
tf.flags.DEFINE_integer("num_checkpoints", 2, "Number of checkpoints to store (default: 2)")

# 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("")


# 获得训练用的wav文件路径列表  
def get_wav_files(parent_dir,sub_dirs): 
    wav_files = []  
    for l, sub_dir in enumerate(sub_dirs):
        wav_path = os.path.join(parent_dir, sub_dir)
        for (dirpath, dirnames, filenames) in os.walk(wav_path):  
            for filename in filenames:  
                if filename.endswith('.wav') or filename.endswith('.WAV'):  
                    filename_path = os.sep.join([dirpath, filename])  
                    wav_files.append(filename_path)  
    return wav_files  

# 获取文件mfcc特征和对应标签
def extract_features(wav_files):
    inputs = []
    labels = []
    
    for wav_file in tqdm(wav_files):
        # 读入音频文件
        audio,fs = librosa.load(wav_file)

        # 获取音频mfcc特征
        # [n_steps, n_inputs]
        mfccs = np.transpose(librosa.feature.mfcc(y=audio, sr=fs, n_mfcc=FLAGS.n_inputs), [1,0]) 
        inputs.append(mfccs.tolist()) 
    #获取label
    for wav_file in wav_files:
        label = wav_file.split('/')[-1].split('-')[1]
        labels.append(label) 
    return inputs, np.array(labels, dtype=np.int)

# 获得训练用的wav文件路径列表 
wav_files = get_wav_files(FLAGS.parent_dir,FLAGS.tr_sub_dirs)
# 获取文件mfcc特征和对应标签
tr_features,tr_labels = extract_features(wav_files)

np.save('tr_features.npy',tr_features)
np.save('tr_labels.npy',tr_labels)

# tr_features=np.load('tr_features.npy')
# tr_labels=np.load('tr_labels.npy')

#(batch,step,input)
#(50,173,40)

# 计算最长的step
wav_max_len = max([len(feature) for feature in tr_features])
print("max_len:",wav_max_len)

# 填充0
tr_data = []
for mfccs in tr_features:  
    while len(mfccs) < wav_max_len: #只要小于wav_max_len就补n_inputs个0
        mfccs.append([0] * FLAGS.n_inputs) 
    tr_data.append(mfccs)

tr_data = np.array(tr_data)


# In[6]:

# Randomly shuffle data
np.random.seed(10)
shuffle_indices = np.random.permutation(np.arange(len(tr_data)))
x_shuffled = tr_data[shuffle_indices]
y_shuffled = tr_labels[shuffle_indices]

# Split train/test set
# TODO: This is very crude, should use cross-validation
# 数据集切分为两部分
dev_sample_index = -1 * int(FLAGS.dev_sample_percentage * float(len(y_shuffled)))
train_x, test_x = x_shuffled[:dev_sample_index], x_shuffled[dev_sample_index:]
train_y, test_y = y_shuffled[:dev_sample_index], y_shuffled[dev_sample_index:]


# In[7]:

# placeholder
x = tf.placeholder("float", [None, wav_max_len, FLAGS.n_inputs])
y = tf.placeholder("float", [None])
dropout = tf.placeholder(tf.float32)
# learning rate
lr = tf.Variable(FLAGS.lr, dtype=tf.float32, trainable=False)

# 定义RNN网络
# 初始化权制和偏置
weights = tf.Variable(tf.truncated_normal([FLAGS.n_hidden, FLAGS.n_classes], stddev=0.1))
biases = tf.Variable(tf.constant(0.1, shape=[FLAGS.n_classes]))

# 多层网络
num_layers = 3
def grucell():
    cell = tf.contrib.rnn.GRUCell(FLAGS.n_hidden)
#     cell = tf.contrib.rnn.LSTMCell(FLAGS.n_hidden)
    cell = tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=dropout)
    return cell
cell = tf.contrib.rnn.MultiRNNCell([grucell() for _ in range(num_layers)])


outputs,final_state = tf.nn.dynamic_rnn(cell,x,dtype=tf.float32)

# 预测值
prediction = tf.nn.softmax(tf.matmul(final_state[0],weights) + biases)

# labels转one_hot格式
one_hot_labels = tf.one_hot(indices=tf.cast(y, tf.int32), depth=FLAGS.n_classes)

# loss
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction,labels=one_hot_labels))

# optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=lr).minimize(cross_entropy)

# Evaluate model
correct_pred = tf.equal(tf.argmax(prediction,1), tf.argmax(one_hot_labels,1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))


# In[8]:

def batch_iter(data, batch_size, num_epochs, shuffle=True):
    """
        Generates a batch iterator for a dataset.
    """
    data = np.array(data)
    data_size = len(data)
    # 每个epoch的num_batch
    num_batches_per_epoch = int((len(data) - 1) / batch_size) + 1
    print("num_batches_per_epoch:",num_batches_per_epoch)
    for epoch in range(num_epochs):
        # Shuffle the data at each epoch
        if shuffle:
            shuffle_indices = np.random.permutation(np.arange(data_size))
            shuffled_data = data[shuffle_indices]
        else:
            shuffled_data = data
        for batch_num in range(num_batches_per_epoch):
            start_index = batch_num * batch_size
            end_index = min((batch_num + 1) * batch_size, data_size)
            yield shuffled_data[start_index:end_index]

# Initializing the variables
init = tf.global_variables_initializer()
# 定义saver
saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(init) 

    # Generate batches
    batches = batch_iter(list(zip(train_x, train_y)), FLAGS.batch_size, FLAGS.num_epochs)

    for i,batch in enumerate(batches):
        i = i + 1
        x_batch, y_batch = zip(*batch)
        sess.run([optimizer], feed_dict={x: x_batch, y: y_batch, dropout: FLAGS.dropout_keep_prob})
        
        # 测试
        if i % FLAGS.evaluate_every == 0:
            sess.run(tf.assign(lr, FLAGS.lr * (0.99 ** (i // FLAGS.evaluate_every))))
            learning_rate = sess.run(lr)
            tr_acc, _loss = sess.run([accuracy, cross_entropy], feed_dict={x: train_x, y: train_y, dropout: 1.0})
            ts_acc = sess.run(accuracy, feed_dict={x: test_x, y: test_y, dropout: 1.0})
            print("Iter {}, loss {:.5f}, tr_acc {:.5f}, ts_acc {:.5f}, lr {:.5f}".format(i, _loss, tr_acc, ts_acc, learning_rate))

        # 保存模型
        if i % FLAGS.checkpoint_every == 0:
            path = saver.save(sess, "sounds_models/model", global_step=i)