基于深度学习的语音识别声学模型研究（一）

  经历了两周的bug，终于把全连接神经网络搭好，并把数据成功的喂入到神经网络中。下面我我对我的网络进行简单的介绍：

  首先通过tensorboard‘画图’，网络的结构如下图所示：

  这是一侧五层的全连接神经网路：神经元等超参数的设定在之后的脚本中有详细的解释，具体见后面的代码。

  之后训练200轮之后绘制出测试集和训练集的曲线以及loss曲线如下图所示：

  通过上面的分析，我们发现，当网络堆叠到5层的时候，处理80w的声学特征的准确率也只有在40%左右，无法达到更高的效果，这个和网络的本身的结构有关，之后准备尝试使用RNN,CNN等不同结构的网络进行试验，下面附上代码：

#-*- coding:utf-8 -*-
#author : zhangwei

import numpy as np
import tensorflow as tf

filename_01 = '/home/zhangwei/data/train_mfcc_800000.txt'
filename_02 = '/home/zhangwei/data/train_label_800000.txt'
filename_03 = '/home/zhangwei/data/test_mfcc.txt'
filename_04 = '/home/zhangwei/data/test_label.txt'
X_train = np.loadtxt(filename_01)
Y_train = np.loadtxt(filename_02)
X_test = np.loadtxt(filename_03)
Y_test = np.loadtxt(filename_04)

batch_size = 100
n_epoch = 200

def bn(in_size,out_size):
    fc_mean , fc_var = tf.nn.moments(in_size , axes=[0])
    scale = tf.Variable(tf.ones([out_size]))
    shift = tf.Variable(tf.zeros([out_size]))
    epsilon = 0.001
    Wx_plus_b = tf.nn.batch_normalization(in_size , fc_mean , fc_var , shift , scale , epsilon)
    return  Wx_plus_b

with tf.Graph().as_default():
    with tf.name_scope('Input'):
        x = tf.placeholder(tf.float32 , [batch_size , 39])
        y = tf.placeholder(tf.int64 , [batch_size , 219])
        keep_prob = tf.placeholder(tf.float32)

    with tf.name_scope('Layer_1'):
        W1 = tf.Variable(tf.truncated_normal([39,100] , stddev=0.1))
        tf.summary.histogram('layer1_W' , W1)
        b1 = tf.Variable(tf.zeros([100]) + 0.1)
        tf.summary.histogram('layer1_b' , b1)
        l1 = tf.add(tf.matmul(x , W1) ,b1)
        l1_bn = bn(l1 , 100)
        h1 = tf.nn.relu(l1_bn)
        h1_prob = tf.nn.dropout(h1 , keep_prob)

    with tf.name_scope('Layer_2'):
        W2 = tf.Variable(tf.truncated_normal([100  , 256] , stddev=0.1))
        tf.summary.histogram('layer2_W', W2)
        b2 = tf.Variable(tf.zeros([256])+0.1)
        tf.summary.histogram('layer2_b', b2)
        l2 = tf.matmul(h1_prob , W2) + b2
        l2_bn = bn(l2 , 256)
        h2 = tf.nn.relu(l2_bn)
        h2_drop = tf.nn.dropout(h2 , keep_prob)

    with tf.name_scope('Layer_3'):
        W3 = tf.Variable(tf.truncated_normal([256  , 512] , stddev=0.1))
        tf.summary.histogram('layers_w' , W3)
        b3 = tf.Variable(tf.zeros([512])+0.1)
        tf.summary.histogram('layer3_b' , b3)
        l3 = tf.matmul(h2 , W3) + b3
        l3_bn = bn(l3 , 512)
        h3 = tf.nn.relu(l3_bn)
        h3_drop = tf.nn.dropout(h3 , keep_prob)

    with tf.name_scope('Layer_4'):
        W4 = tf.Variable(tf.truncated_normal([512  , 219] , stddev=0.1))
        tf.summary.histogram('layer4_w' , W4)
        b4 = tf.Variable(tf.zeros([219])+0.1)
        tf.summary.histogram('layer4_b' , b4)
        l4= tf.matmul(h3_drop , W4) + b4
        prediction = tf.nn.relu(l4)

    with tf.name_scope('Loss'):
        loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction , labels=y))
        tf.summary.scalar('loss' , loss_op)

    with tf.name_scope('Train'):
        train_op = tf.train.AdamOptimizer(0.001).minimize(loss_op)

    with tf.name_scope('Accuracy'):
        correct_prediction = tf.equal(tf.argmax(prediction , 1) , tf.argmax(y , 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction , tf.float32))
        tf.summary.scalar('Accuracy' , accuracy)

    merged = tf.summary.merge_all()
    init = tf.global_variables_initializer()

    with tf.Session() as sess:
        sess.run(init)
        train_writer = tf.summary.FileWriter('/home/zhangwei/tensorboard/train' , sess.graph)
        test_writer = tf.summary.FileWriter('/home/zhangwei/tensorboard/test' , sess.graph)
        
                    test_acc = sess.run(accuracy , feed_dict=feed_dict)
                    summary = sess.run(merged , feed_dict=feed_dict)
                    test_writer.add_summary(summary , j)

            #         result = sess.run(prediction , feed_dict=feed_dict)
            # print 'Iter : ' + str(j) + '; result' + str(result)
            print 'Iter : ' + str(j) + ' ; Loss : ' + str(loss) + ' ; training_Accuracy : ' + str(train_acc) +\
                  ' ; testing accuracy : ' + str(test_acc)