环境：ubuntu14.04+python3.5+sklearn
本帖训练一个简单的神经网络模型，用来判断声音是男是女。
本帖数据集取自voice-gender项目，这个项目使用了n种分类模型，并比较了准确率，但是它没有使用神经网络模型，本帖算是一个补充。
数据集
这个数据集是经过R语言处理过的，它提取出了.WAV文件的一些声音属性。如果你想自己从wav文件中提取声音属性，参看voice-gender项目中一个叫sound.R源码文件。
数据集字段：”meanfreq”,”sd”,”median”,”Q25″,”Q75″,”IQR”,”skew”,”kurt”,”sp.ent”,”sfm”,”mode”,”centroid”,”meanfun”,”minfun”,”maxfun”,”meandom”,”mindom”,”maxdom”,”dfrange”,”modindx”,”label”。最后一个字段标记了是男声还是女声，前面字段是声音属性。
import os
import requests
import pandas as pd
import numpy as np
import random
import tensorflow as tf # 0.12
from sklearn.model_selection import train_test_split
# 下载数据集
if not os.path.exists('voice.csv'):
url = 'http://blog.topspeedsnail.com/wp-content/uploads/2016/12/voice.csv'
data = requests.get(url).content
with open('voice.csv', 'wb') as f:
f.write(data)
voice_data = pd.read_csv('voice.csv')
#print(voice_data.head())
#print(voice_data.tail())
voice_data = voice_data.values
# 分离声音特性和分类
voices = voice_data[:, :-1]
labels = voice_data[:, -1:] # ['male'] ['female']
# 把分类转为one-hot
labels_tmp = []
for label in labels:
tmp = []
if label[0] == 'male':
tmp = [1.0, 0.0]
else: # 'female'
tmp = [0.0, 1.0]
labels_tmp.append(tmp)
labels = np.array(labels_tmp)
# shuffle
voices_tmp = []
lables_tmp = []
index_shuf = [i for i in range(len(voices))]
random.shuffle(index_shuf)
for i in index_shuf:
voices_tmp.append(voices[i])
lables_tmp.append(labels[i])
voices = np.array(voices_tmp)
labels = np.array(lables_tmp)
train_x, test_x, train_y, test_y = train_test_split(voices, labels, test_size=0.1)
banch_size = 64
n_banch = len(train_x) // banch_size
X = tf.placeholder(dtype=tf.float32, shape=[None, voices.shape[-1]]) # 20
Y = tf.placeholder(dtype=tf.float32, shape=[None, 2])
# 3层（feed-forward）
def neural_network():
w1 = tf.Variable(tf.random_normal([voices.shape[-1], 512], stddev=0.5))
b1 = tf.Variable(tf.random_normal([512]))
output = tf.matmul(X, w1) + b1
w2 = tf.Variable(tf.random_normal([512, 1024],stddev=.5))
b2 = tf.Variable(tf.random_normal([1024]))
output = tf.nn.softmax(tf.matmul(output, w2) + b2)
w3 = tf.Variable(tf.random_normal([1024, 2],stddev=.5))
b3 = tf.Variable(tf.random_normal([2]))
output = tf.nn.softmax(tf.matmul(output, w3) + b3)
return output
# 训练神经网络
def train_neural_network():
output = neural_network()
cost = tf.reduce_mean(tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits=output, labels=Y)))
lr = tf.Variable(0.001, dtype=tf.float32, trainable=False)
opt = tf.train.AdamOptimizer(learning_rate=lr)
var_list = [t for t in tf.trainable_variables()]
train_step = opt.minimize(cost, var_list=var_list)
#saver = tf.train.Saver(tf.global_variables())
#saver.restore(sess, tf.train.latest_checkpoint('.'))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
#summary_writer = tf.train.SummaryWriter('voices')
for epoch in range(500):
sess.run(tf.assign(lr, 0.001 * (0.97 ** epoch)))
for banch in range(n_banch):
voice_banch = train_x[banch*banch_size:(banch+1)*(banch_size)]
label_banch = train_y[banch*banch_size:(banch+1)*(banch_size)]
_, loss = sess.run([train_step, cost], feed_dict={X: voice_banch, Y: label_banch})
print(epoch, banch, loss)
# 准确率
prediction = tf.equal(tf.argmax(output, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(prediction, dtype=tf.float32))
accuracy = sess.run(accuracy, feed_dict={X: test_x, Y: test_y})
print("准确率", accuracy)
#prediction = sess.run(output, feed_dict={X: test_x})
train_neural_network()
原文链接：http://blog.topspeedsnail.com/archives/10729
