技术标签: 深度学习之TensorFlow入门、原理与进阶实战
9-23 yuyinchall
准备一批带有文字标注的语音样本,构建BiRNN网络,通过该语料样本进行训练,实现一个能够识别语音的神经网络模型
程序:
import numpy as np
import time
import tensorflow as tf
from tensorflow.python.ops import ctc_ops
from collections import Counter
#2 样本读取
## 自定义
yuyinutils = __import__("9-24 yuyinutils")
sparse_tuple_to_texts_ch = yuyinutils.sparse_tuple_to_texts_ch
ndarray_to_text_ch = yuyinutils.ndarray_to_text_ch
get_audio_and_transcriptch = yuyinutils.get_audio_and_transcriptch
pad_sequences = yuyinutils.pad_sequences
sparse_tuple_from = yuyinutils.sparse_tuple_from
get_wavs_lables = yuyinutils.get_wavs_lables
tf.reset_default_graph()
b_stddev = 0.046875
h_stddev = 0.046875
n_hidden = 1024
n_hidden_1 = 1024
n_hidden_2 = 1024
n_hidden_5 = 1024
n_cell_dim = 1024
n_hidden_3 = 2 * 1024
keep_dropout_rate = 0.95
relu_clip = 20
#使用3个1024节点的全连接层,然后是一个双向RNN,最后接上2个全连接层,并且都带有dropout层。这里使用的激活函数是带截断的Relu,截断值设为20。学习参数的初始化使用标准差为0.046875的random_normal。keep_dropout_rate为0.95.
def BiRNN_model(batch_x, seq_length, n_input, n_context, n_character, keep_dropout):
# batch_x_shape: [batch_size, n_steps, n_input + 2*n_input*n_context]
batch_x_shape = tf.shape(batch_x)
# 将输入转成时间序列优先
batch_x = tf.transpose(batch_x, [1, 0, 2])
# 再转成2维传入第一层
batch_x = tf.reshape(batch_x,
[-1, n_input + 2 * n_input * n_context]) # (n_steps*batch_size, n_input + 2*n_input*n_context)
# 使用clipped RELU activation and dropout.
# 1st layer
with tf.name_scope('fc1'):
b1 = variable_on_cpu('b1', [n_hidden_1], tf.random_normal_initializer(stddev=b_stddev))
h1 = variable_on_cpu('h1', [n_input + 2 * n_input * n_context, n_hidden_1],
tf.random_normal_initializer(stddev=h_stddev))
layer_1 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(batch_x, h1), b1)), relu_clip)
layer_1 = tf.nn.dropout(layer_1, keep_dropout)
# 2nd layer
with tf.name_scope('fc2'):
b2 = variable_on_cpu('b2', [n_hidden_2], tf.random_normal_initializer(stddev=b_stddev))
h2 = variable_on_cpu('h2', [n_hidden_1, n_hidden_2], tf.random_normal_initializer(stddev=h_stddev))
layer_2 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(layer_1, h2), b2)), relu_clip)
layer_2 = tf.nn.dropout(layer_2, keep_dropout)
# 3rd layer
with tf.name_scope('fc3'):
b3 = variable_on_cpu('b3', [n_hidden_3], tf.random_normal_initializer(stddev=b_stddev))
h3 = variable_on_cpu('h3', [n_hidden_2, n_hidden_3], tf.random_normal_initializer(stddev=h_stddev))
layer_3 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(layer_2, h3), b3)), relu_clip)
layer_3 = tf.nn.dropout(layer_3, keep_dropout)
# 双向rnn
with tf.name_scope('lstm'):
# Forward direction cell:
lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True)
lstm_fw_cell = tf.contrib.rnn.DropoutWrapper(lstm_fw_cell,
input_keep_prob=keep_dropout)
# Backward direction cell:
lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True)
lstm_bw_cell = tf.contrib.rnn.DropoutWrapper(lstm_bw_cell,
input_keep_prob=keep_dropout)
# `layer_3` `[n_steps, batch_size, 2*n_cell_dim]`
layer_3 = tf.reshape(layer_3, [-1, batch_x_shape[0], n_hidden_3])
outputs, output_states = tf.nn.bidirectional_dynamic_rnn(cell_fw=lstm_fw_cell,
cell_bw=lstm_bw_cell,
inputs=layer_3,
dtype=tf.float32,
time_major=True,
sequence_length=seq_length)
# 连接正反向结果[n_steps, batch_size, 2*n_cell_dim]
outputs = tf.concat(outputs, 2)
# to a single tensor of shape [n_steps*batch_size, 2*n_cell_dim]
outputs = tf.reshape(outputs, [-1, 2 * n_cell_dim])
with tf.name_scope('fc5'):
b5 = variable_on_cpu('b5', [n_hidden_5], tf.random_normal_initializer(stddev=b_stddev))
h5 = variable_on_cpu('h5', [(2 * n_cell_dim), n_hidden_5], tf.random_normal_initializer(stddev=h_stddev))
layer_5 = tf.minimum(tf.nn.relu(tf.add(tf.matmul(outputs, h5), b5)), relu_clip)
layer_5 = tf.nn.dropout(layer_5, keep_dropout)
with tf.name_scope('fc6'):
# 全连接层用于softmax分类
b6 = variable_on_cpu('b6', [n_character], tf.random_normal_initializer(stddev=b_stddev))
h6 = variable_on_cpu('h6', [n_hidden_5, n_character], tf.random_normal_initializer(stddev=h_stddev))
layer_6 = tf.add(tf.matmul(layer_5, h6), b6)
# 将2维[n_steps*batch_size, n_character]转成3维 time-major [n_steps, batch_size, n_character].
layer_6 = tf.reshape(layer_6, [-1, batch_x_shape[0], n_character])
# Output shape: [n_steps, batch_size, n_character]
return layer_6
"""
used to create a variable in CPU memory.
"""
def variable_on_cpu(name, shape, initializer):
# Use the /cpu:0 device for scoped operations
with tf.device('/cpu:0'):
# Create or get apropos variable
var = tf.get_variable(name=name, shape=shape, initializer=initializer)
return var
wav_path = 'F:/shendu/yuyinchall/wav/wav/train'
label_file = 'F:/shendu/yuyinchall/doc/doc/trans/train.word.txt'
wav_files, labels = get_wavs_lables(wav_path, label_file)
print(wav_files[0], labels[0])
# wav/train/A11/A11_0.WAV -> 绿 是 阳春 烟 景 大块 文章 的 底色 四月 的 林 峦 更是 绿 得 鲜活 秀媚 诗意 盎然
print("wav:", len(wav_files), "label", len(labels))
'''----------------------------------------------------------------------'''
#3 建立批次获取样本函数
# 字表
all_words = []
for label in labels:
# print(label)
all_words += [word for word in label]
counter = Counter(all_words)
words = sorted(counter)
words_size = len(words)
word_num_map = dict(zip(words, range(words_size)))
print('字表大小:', words_size)
n_input = 26 # 计算美尔倒谱系数的个数
n_context = 9 # 对于每个时间点,要包含上下文样本的个数
batch_size = 8
def next_batch(labels, start_idx=0, batch_size=1, wav_files=wav_files):
filesize = len(labels)
end_idx = min(filesize, start_idx + batch_size)
idx_list = range(start_idx, end_idx)
txt_labels = [labels[i] for i in idx_list]
wav_files = [wav_files[i] for i in idx_list]
(source, audio_len, target, transcript_len) = get_audio_and_transcriptch(None,
wav_files,
n_input,
n_context, word_num_map, txt_labels)
start_idx += batch_size
# Verify that the start_idx is not larger than total available sample size
if start_idx >= filesize:
start_idx = -1
# Pad input to max_time_step of this batch
source, source_lengths = pad_sequences(source) # 如果多个文件将长度统一,支持按最大截断或补0
sparse_labels = sparse_tuple_from(target)
return start_idx, source, source_lengths, sparse_labels
next_idx, source, source_len, sparse_lab = next_batch(labels, 0, batch_size)
print(len(sparse_lab))
print(np.shape(source))
# print(sparse_lab)
t = sparse_tuple_to_texts_ch(sparse_lab, words)
print(t[0])
# source已经将变为前9(不够补空)+本身+后9,每个26,第一个顺序是第10个的数据。
'''---------------------------------------------------------------------'''
#1 定义占位符
# shape = [batch_size, max_stepsize, n_input + (2 * n_input * n_context)]
# the batch_size and max_stepsize每步都是变长的。
input_tensor = tf.placeholder(tf.float32, [None, None, n_input + (2 * n_input * n_context)],
name='input') # 语音log filter bank or MFCC features
# Use sparse_placeholder; will generate a SparseTensor, required by ctc_loss op.
targets = tf.sparse_placeholder(tf.int32, name='targets') # 文本
# 1d array of size [batch_size]
seq_length = tf.placeholder(tf.int32, [None], name='seq_length') # 序列长
keep_dropout = tf.placeholder(tf.float32)
'''----------------------------------------------------------------------'''
#2 构建网络模型
# logits is the non-normalized output/activations from the last layer.
# logits will be input for the loss function.
# nn_model is from the import statement in the load_model function
logits = BiRNN_model(input_tensor, tf.to_int64(seq_length), n_input, n_context, words_size + 1, keep_dropout)
'''-----------'''
'''----------------------------------------------------------------------'''
#3 定义损失函数即优化器
# 调用ctc loss
avg_loss = tf.reduce_mean(ctc_ops.ctc_loss(targets, logits, seq_length))
# [optimizer]
learning_rate = 0.001
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(avg_loss)
'''----------------------------------------------------------------------'''
#4 定义解码,并评估模型节点
with tf.name_scope("decode"):
decoded, log_prob = ctc_ops.ctc_beam_search_decoder(logits, seq_length, merge_repeated=False)
with tf.name_scope("accuracy"):
distance = tf.edit_distance(tf.cast(decoded[0], tf.int32), targets)
# 计算label error rate (accuracy)
ler = tf.reduce_mean(distance, name='label_error_rate')
'''----------------------------------------------------------------------'''
#建立session并添加检查点处理
epochs = 100
savedir = "F:/shendu/yuyinchalltest/"
saver = tf.train.Saver(max_to_keep=1) # 生成saver
# create the session
sess = tf.Session()
# 没有模型的话,就重新初始化
sess.run(tf.global_variables_initializer())
kpt = tf.train.latest_checkpoint(savedir)
print("kpt:", kpt)
startepo = 0
if kpt != None:
saver.restore(sess, kpt)
ind = kpt.find("-")
startepo = int(kpt[ind + 1:])
print(startepo)
'''----------------------------------------------------------------------'''
#6 通过循环来迭代训练模型
# 准备运行训练步骤
section = '\n{0:=^40}\n'
print(section.format('Run training epoch'))
train_start = time.time()
for epoch in range(epochs): # 样本集迭代次数
epoch_start = time.time()
if epoch < startepo:
continue
print("epoch start:", epoch, "total epochs= ", epochs)
#######################run batch####
n_batches_per_epoch = int(np.ceil(len(labels) / batch_size))
print("total loop ", n_batches_per_epoch, "in one epoch,", batch_size, "items in one loop")
train_cost = 0
train_ler = 0
next_idx = 0
for batch in range(n_batches_per_epoch): # 一次batch_size,取多少次
# 取数据
next_idx, source, source_lengths, sparse_labels = \
next_batch(labels, next_idx, batch_size)
feed = {input_tensor: source, targets: sparse_labels, seq_length: source_lengths,
keep_dropout: keep_dropout_rate}
# 计算 avg_loss optimizer ;
batch_cost, _ = sess.run([avg_loss, optimizer], feed_dict=feed)
train_cost += batch_cost
'''----------------------------------------------------------------------'''
# 7 定期评估模型。输出模型解码结果
if (batch + 1) % 20 == 0:
print('loop:', batch, 'Train cost: ', train_cost / (batch + 1))
feed2 = {input_tensor: source, targets: sparse_labels, seq_length: source_lengths, keep_dropout: 1.0}
d, train_ler = sess.run([decoded[0], ler], feed_dict=feed2)
dense_decoded = tf.sparse_tensor_to_dense(d, default_value=-1).eval(session=sess)
dense_labels = sparse_tuple_to_texts_ch(sparse_labels, words)
counter = 0
print('Label err rate: ', train_ler)
for orig, decoded_arr in zip(dense_labels, dense_decoded):
# convert to strings
decoded_str = ndarray_to_text_ch(decoded_arr, words)
print(' file {}'.format(counter))
print('Original: {}'.format(orig))
print('Decoded: {}'.format(decoded_str))
counter = counter + 1
break
epoch_duration = time.time() - epoch_start
log = 'Epoch {}/{}, train_cost: {:.3f}, train_ler: {:.3f}, time: {:.2f} sec'
print(log.format(epoch, epochs, train_cost, train_ler, epoch_duration))
saver.save(sess, savedir + "yuyinch.cpkt", global_step=epoch)
train_duration = time.time() - train_start
print('Training complete, total duration: {:.2f} min'.format(train_duration / 60))
sess.close()
结果:
文章浏览阅读1.6k次。安装配置gi、安装数据库软件、dbca建库见下:http://blog.csdn.net/kadwf123/article/details/784299611、检查集群节点及状态:[root@rac2 ~]# olsnodes -srac1 Activerac2 Activerac3 Activerac4 Active[root@rac2 ~]_12c查看crs状态
文章浏览阅读1.3w次,点赞45次,收藏99次。我个人用的是anaconda3的一个python集成环境,自带jupyter notebook,但在我打开jupyter notebook界面后,却找不到对应的虚拟环境,原来是jupyter notebook只是通用于下载anaconda时自带的环境,其他环境要想使用必须手动下载一些库:1.首先进入到自己创建的虚拟环境(pytorch是虚拟环境的名字)activate pytorch2.在该环境下下载这个库conda install ipykernelconda install nb__jupyter没有pytorch环境
文章浏览阅读5.2k次,点赞19次,收藏28次。选择scoop纯属意外,也是无奈,因为电脑用户被锁了管理员权限,所有exe安装程序都无法安装,只可以用绿色软件,最后被我发现scoop,省去了到处下载XXX绿色版的烦恼,当然scoop里需要管理员权限的软件也跟我无缘了(譬如everything)。推荐添加dorado这个bucket镜像,里面很多中文软件,但是部分国外的软件下载地址在github,可能无法下载。以上两个是官方bucket的国内镜像,所有软件建议优先从这里下载。上面可以看到很多bucket以及软件数。如果官网登陆不了可以试一下以下方式。_scoop-cn
文章浏览阅读4.5k次,点赞2次,收藏3次。首先要有一个color-picker组件 <el-color-picker v-model="headcolor"></el-color-picker>在data里面data() { return {headcolor: ’ #278add ’ //这里可以选择一个默认的颜色} }然后在你想要改变颜色的地方用v-bind绑定就好了,例如:这里的:sty..._vue el-color-picker
文章浏览阅读640次。基于芯片日益增长的问题,所以内核开发者们引入了新的方法,就是在内核中只保留函数,而数据则不包含,由用户(应用程序员)自己把数据按照规定的格式编写,并放在约定的地方,为了不占用过多的内存,还要求数据以根精简的方式编写。boot启动时,传参给内核,告诉内核设备树文件和kernel的位置,内核启动时根据地址去找到设备树文件,再利用专用的编译器去反编译dtb文件,将dtb还原成数据结构,以供驱动的函数去调用。firmware是三星的一个固件的设备信息,因为找不到固件,所以内核启动不成功。_exynos 4412 刷机
文章浏览阅读2w次,点赞24次,收藏42次。Linux系统配置jdkLinux学习教程,Linux入门教程(超详细)_linux配置jdk
文章浏览阅读3.3k次,点赞5次,收藏19次。xlabel('\delta');ylabel('AUC');具体符号的对照表参照下图:_matlab微米怎么输入
文章浏览阅读119次。顺序读写指的是按照文件中数据的顺序进行读取或写入。对于文本文件,可以使用fgets、fputs、fscanf、fprintf等函数进行顺序读写。在C语言中,对文件的操作通常涉及文件的打开、读写以及关闭。文件的打开使用fopen函数,而关闭则使用fclose函数。在C语言中,可以使用fread和fwrite函数进行二进制读写。 Biaoge 于2024-03-09 23:51发布 阅读量:7 ️文章类型:【 C语言程序设计 】在C语言中,用于打开文件的函数是____,用于关闭文件的函数是____。
文章浏览阅读3.4k次,点赞2次,收藏13次。跟随鼠标移动的粒子以grid(SOP)为partical(SOP)的资源模板,调整后连接【Geo组合+point spirit(MAT)】,在连接【feedback组合】适当调整。影响粒子动态的节点【metaball(SOP)+force(SOP)】添加mouse in(CHOP)鼠标位置到metaball的坐标,实现鼠标影响。..._touchdesigner怎么让一个模型跟着鼠标移动
文章浏览阅读178次。项目运行环境配置:Jdk1.8 + Tomcat7.0 + Mysql + HBuilderX(Webstorm也行)+ Eclispe(IntelliJ IDEA,Eclispe,MyEclispe,Sts都支持)。项目技术:Springboot + mybatis + Maven +mysql5.7或8.0+html+css+js等等组成,B/S模式 + Maven管理等等。环境需要1.运行环境:最好是java jdk 1.8,我们在这个平台上运行的。其他版本理论上也可以。_基于java技术的停车场管理系统实现与设计
文章浏览阅读3.5k次。前言对于MediaPlayer播放器的源码分析内容相对来说比较多,会从Java-&amp;gt;Jni-&amp;gt;C/C++慢慢分析,后面会慢慢更新。另外,博客只作为自己学习记录的一种方式,对于其他的不过多的评论。MediaPlayerDemopublic class MainActivity extends AppCompatActivity implements SurfaceHolder.Cal..._android多媒体播放源码分析 时序图
文章浏览阅读2.4k次,点赞41次,收藏13次。java 数据结构与算法 ——快速排序法_快速排序法