使用scikit-learn进行音乐分类

使用scikit-learn进行音乐分类

之前简要说了下Yaafe和Essentia的安装使用，自己也专门对这两个框架做了封装，实现了各个特征提取的接口，并将特征值存到本地文件中方便训练时使用。下面这张图是对应提取的特征数据，一共35种特征，每种特征维度不一样。例如MFCC最大值维度为13维。



这里的音乐分类主要是按情感进行分类，一共分成7种情感：happy，sad，passion，excited，quiet，nostalgic，relax。

一、得到训练数据

def getTrainData(train_path, train_mode, mode_name, featureList):
num = []
train_data = []
train_target = []
introFeature = []

for i in range(len(train_mode)):
allFeature = []
# print len(train_mode),len(train_path),i,train_path[0],train_path[1]
# print tool.getFileNum(train_path[i])
num.append(tool.getFileNum(train_path[i]))
for j in tool.getAllFile(train_path[i]):
yl.startEngine(j)
el.startLoader(j)
oneFeature, introFeature = getFeature(yl, el, featureList)

allFeature.append(oneFeature)
# print allFeature

# 判断是否有歌曲的特征被提取出
if len(allFeature) == 0:
print '无音频文件'
else:
wf.setMode(train_mode[i])  # 设置类别和目录．输出
wf.setFeaturePrintPath(train_path[i])  # 音乐目录/feature_num/train_mode[i].txt  readme.txt
wf.writeFeature(allFeature, introFeature)
tool.mkdir('dataSystem/' + mode_name)
wf.writeDetial('dataSystem/' + mode_name + '/feature.txt', featureList)
for singleFeature in allFeature:
train_data.append(singleFeature)  # 设置训练数据
train_target.append(i)  # 设置训练目标

return train_data, train_target

train_path和train_mode分别表示传入的一类音乐文件夹路径，该类别标签，例如train_path=‘/home/chenming/Music/Happy’表示happy类型音乐文件夹，train_mode=‘happy’表示该组训练数据类型为happy。featureList是传入的需要提取的特征，例如featureList=[1,2,3]表示提取FEATURES中对应1,2,3位置的特征（上面第一张图片）。

train_data = []和 train_target = []分别是得到的训练数据以及该数据对应的标签。最终特征数据存到/home/chenming/Music/Happy下feature_num/happy.txt中

二、训练

def train(train_data, train_target, train_mode, mode_name):
# 数据标准化
train_data = np.array(train_data)
train_target = np.array(train_target)

# sklearn提供的数据处理方法
# train_data = preprocessing.scale(train_data)
# test_data = preprocessing.scale(test_data)

(meanData, stdData) = tool.trainDataStandardNor(train_data)

# 存储
np.savetxt('dataSystem/' + mode_name + '/meanData.txt', meanData)
np.savetxt('dataSystem/' + mode_name + '/stdData.txt', stdData)
wf.writeDetial('dataSystem/' + mode_name + '/modeList.txt', train_mode)
#np.savetxt('dataSystem/' + mode_name + '/modeList.txt', train_mode)
'''
c=np.arange(1,10,1)
g=np.arange(0,1,0.1)
tuned_parameters = [{'kernel': ['rbf'], 'gamma': g,'C': c}]
#print classifierType,type(classifierType)
'''
clf = svm.SVC()
if classifierType == 'svm':
clf = svm.SVC()
elif classifierType == 'knn':
clf = neighbors.KNeighborsClassifier()  # knn K-临近算法
elif classifierType == 'tree':
clf = tree.DecisionTreeClassifier()  # 决策树

# clf = grid_search.GridSearchCV(svm.SVC(), tuned_parameters, cv=3) #svm支持向量机
print train_data
print train_target

clf.fit(train_data, train_target)
# clf.fit(train_data)
# clf.score(test_data, test_target)
# showStyleDetail(test_data,test_target,clf)
from sklearn.externals import joblib

tool.mkdir('svcSystem/' + mode_name)
joblib.dump(clf, 'svcSystem/' + mode_name + '/svc1.pkl')
# clf2=joblib.load('svc/svc1.pkl')
# print clf.score(test_data, test_target)
# showDetail(test_data,test_target,clf2)
# print clf.best_estimator_

return '训练成功,模型' + mode_name + '存储成功,'+'位置:svcSystem/'+mode_name +'/svc1.pkl'

(meanData, stdData) = tool.trainDataStandardNor(train_data)是对数据进行归一化处理，代码如下：

def trainDataStandardNor(featureMatrix):

#print 'train_feature',featureMatrix
meanData=featureMatrix.mean(axis=0) #平均值
stdData=featureMatrix.std(axis=0) #标准差
for data in featureMatrix:
for i in range(len(data)):
if stdData[i]==0:
data[i]=0
else:
print data[i]
data[i]=(data[i]-meanData[i])/stdData[i]

return (meanData,stdData)

http://www.tuicool.com/articles/qeIzI3F 这篇博客介绍了Scikit-Learn的使用。

三、测试

def test(test_data, test_target, test_mode, mode_name):
test_data = np.array(test_data)
test_target = np.array(test_target)

print 'test_data:', test_data
print 'test_target:', test_target
print 'test_mode:', test_mode

path1 = 'dataSystem/' + mode_name + '/meanData.txt'
path2 = 'dataSystem/' + mode_name + '/stdData.txt'
path3 = 'svcSystem/' + mode_name + '/svc1.pkl'
path4 = 'dataSystem/' + mode_name + '/modeList.txt'

print '数据路径:' + 'dataSystem/' + mode_name + '和svcSystem/' + mode_name

meanData = np.loadtxt(path1)
stdData = np.loadtxt(path2)
modeList = wf.readFileTxt(path4)
mode = wf.readFileTxt('dataSystem/' + mode_name + '/modeList.txt')

tool.testDataStandardNor(test_data, meanData, stdData)

from sklearn.externals import joblib

clf = joblib.load(path3)

s = showDetail(clf, test_data, test_target, modeList)

return s

showDetail函数展示结果：

def showDetail(clf, test_data, test_target, mode):
modeNum = []

expected = test_target
predicted = clf.predict(test_data)
# print expected
# print predicted
# print test_mode

out = '''
*******准确率=被识别为该分类的正确分类记录数/被识别为该分类的记录数
*******召回率=被识别为该分类的正确分类记录数/测试集中该分类的记录总数
*******F1-score=2（准确率 * 召回率）/（准确率 + 召回率），F1-score是F-measure（又称F-score）beta=1时的特例
*******support=测试集中该分类的记录总数\n
'''
for d in mode:
print 'd', d
out = out + str(metrics.classification_report(expected, predicted, target_names=mode))
print (metrics.classification_report(expected, predicted, target_names=mode))
result = metrics.confusion_matrix(expected, predicted)
out = out + '\n预测数目:\t'
print  '预测数目:\t',
for j in range(len(mode)):
out = out + str(mode[j]) + '\t'
print mode[j], '\t',
out = out + '\n\n'
print '\n'
for i in range(len(result)):
out = out + '测试数据' + str(i + 1) + ':\t'
print '测试数据' + str(i + 1) + ':\t',
num = 0
for p in range(len(result[i])):
out = out + str(result[i][p]) + '\t'
print result[i][p], '\t',
num = num + result[i][p]
out = out + '\n'
print '\n'
modeNum.append([result[i][i], num])
out = out + '\n'
for k in range(len(modeNum)):
out = out + str(mode[k]) + ':\t命中/总数:' + str(modeNum[k][0]) + '/' + str(modeNum[k][1]) + '\t正确率:' + str(
tool.ratio(modeNum[k][0], modeNum[k][1])) + '\n'
print mode[k], ':\t命中/总数:', modeNum[k][0], '/', modeNum[k][1], '\t正确率:', tool.ratio(modeNum[k][0],
modeNum[k][1])

return out

以下是测试结果测试50首happy歌曲音乐，得到47首正确，3首识别错误: