def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
print(f, '->', label, ", score->", score)
return label, score
def task_verify(wav_url, person_id):
start_time = time.time()
print('开始时间：',
time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(start_time)))
m = ModelInterface.load(model)
if person_id not in m.features:
return 'fail', 'current user not trained', ''
# 下载训练语音文件，
current_time = time.strftime("%Y%m%d%H%I%S", time.localtime(time.time()))
dest_wav = verify_voice_dir + current_time + '_' + person_id + '.wav'
utils.download_file(wav_url, dest_wav)
for f in glob.glob(os.path.expanduser(dest_wav)):
fs, signal = utils.read_wav(f)
probability = m.verify(fs, signal, person_id)
print(probability)
if probability > -48:
print(f, '-> 匹配成功 ：', person_id)
return 'success', '', 'yes'
else:
print(f, '->未匹配成功')
return 'success', '', 'no'
end_time = time.time()
print('结束时间：', time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(end_time)))
print('共耗时', end_time - start_time)
def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
fs, signal = read_wav(input_files)
label, score = m.predict(fs, signal)
print("label", '->', label, ", score->", score)
result = [label, score]
return result
def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
if os.path.exists(input_files):
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
filepath = "http://sh.illegalfm.com:4881/record/" + os.path.basename(
input_files)
with DB(host='47.92.33.19',
user='******',
passwd='1qazxsw2',
db='database_fm') as db:
# db.execute("INSERT INTO database_fm (id,radio_file_path,sound_markup) VALUES (null,'{}','{}')".format(f,label))
db.execute(
"UPDATE fm_t_scan_record SET sound_markup = '{}' WHERE radio_file_path = '{}'"
.format(label, filepath))
print(filepath, '->', label, ", score->", score)
os.remove(f)
else:
filepath = "http://sh.illegalfm.com:4881/record/" + os.path.basename(
input_files)
with DB(host='47.92.33.19',
user='******',
passwd='1qazxsw2',
db='database_fm') as db:
db.execute(
"UPDATE fm_t_scan_record SET sound_markup = 'Exception' WHERE radio_file_path = '{}'"
.format(filepath))
def OrderEnroll():
m = ModelInterface.load("model.out")
fs, signal = read_wav("./GUI/TotalRecording/18082020202755.wav")
m.enroll("18082020202755", fs, signal)
m.train()
m.CheckEnroll()
m.dump("mo1.out")
# def task_predictgui(path, input_model):
# m = ModelInterface.load(input_model)
# f=glob.glob(path)
# fs, signal = read_wav(f[0])
# label, score = m.predict(fs, signal)
# return label
#
# if __name__ == "__main__":
# # global args
# # args = get_args()
# #
# # task = args.task
# #
# # if task == 'enroll':
# # task_enroll(args.input, args.model)
# #
# #
# # elif task == 'predict':
# # task_predict(args.input, args.model)
# # task_predict("datatest/*.wav", "model1.out")
# task_enroll("./Train/*","model.out")
#
# # task_predict("./Test", "model.out")
# Predict_ByFile("./GUI/TotalRecording/18082020202755.wav", "D:/doantotnghiep/Speaker_recognition/model.out")
# OrderEnroll()
def load_model(self, MainWindow):
global m
fileName = QtWidgets.QFileDialog().getOpenFileName(
MainWindow, "Load Model", "", "Model File (*.out)")
print(fileName[0])
self.ln_model.setText(fileName[0])
m = ModelInterface.load(fileName[0])
def task_check_status(person_id):
m = ModelInterface.load(model)
if person_id not in m.features:
return 'success', '', 'no'
else:
return 'success', '', 'yes'
def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
g = open("Test_results.txt", "a")
print(f, '->', label, ", score->", score, file=g)
g.close()
def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
str = "label = {l2} score = {l3}".format(l2=label, l3=score)
return [str, label]
def load(self):
fname = QFileDialog.getOpenFileName(self, "Open Data File:", "", "")
if fname:
try:
self.backend = ModelInterface.load(fname)
except Exception as e:
self.warn(str(e))
else:
self.status("Loaded from file: " + fname)
def load(self):
fname = QFileDialog.getOpenFileName(self, "Open Data File:", "", "")
if fname:
try:
self.backend = ModelInterface.load(fname)
except Exception as e:
self.warn(str(e))
else:
self.status("Loaded from file: " + fname)
def __init__(self):
self.backend = ModelInterface.load(self.INPUT_MODEL)
try:
fs, signal = read_wav(self.BG)
self.backend.init_noise(fs, signal)
except:
print "file not found!"
self.pub = rospy.Publisher('/speaker',String,queue_size = 10)
self.sub = rospy.Subscriber('/wav',numpy_msg(Floats),self.task_predict)
def __init__(self):
self.backend = ModelInterface.load(self.INPUT_MODEL)
try:
fs, signal = read_wav(self.BG)
self.backend.init_noise(fs, signal)
except:
print "file not found!"
self.pub = rospy.Publisher('/speaker', String, queue_size=10)
self.sub = rospy.Subscriber('/wav', numpy_msg(Floats),
self.task_predict)
def Predict_ByFile(file, input_model):
print("start")
m = ModelInterface.load(input_model)
fs, signal = read_wav(file)
print(fs)
print(signal)
label, score = m.predict(fs, signal)
strPath = os.path.realpath(file)
y_true = os.path.basename(os.path.dirname(strPath))
print(label)
print(score)
return label
def feature_re_extract():
#pdb.set_trace()
test_class = ['FAML_S', 'FDHH_S', 'FEAB_S', 'FHRO_S',
'FJAZ_S', 'FMEL_S', 'FMEV_S', 'FSLJ_S', 'FTEJ_S',
'FUAN_S', 'MASM_S', 'MCBR_S', 'MFKC_S', 'MKBP_S',
'MLKH_S', 'MMLP_S', 'MMNA_S', 'MNHP_S', 'MOEW_S',
'MPRA_S', 'MREM_S', 'MTLS_S']
m = ModelInterface.load('model/model.out')
# construct train set
train_set = []
for c in test_class:
for i in m.features[c]:
train_set.append(i)
# construct autoencoder
train_data = T.dmatrix('train_data')
x = T.dmatrix('x') # the data is presented as rasterized images
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 10))
da = dA(
numpy_rng=rng,
theano_rng=theano_rng,
input=x,
n_visible=39,
n_hidden=100
)
cost, updates = da.get_cost_updates(
corruption_level=0.,
learning_rate=0.4
)
train_da = theano.function(
[train_data],
cost,
updates=updates,
givens={
x: train_data
}
)
# train autoencoder
training_epochs = 100
c1 = []
for epoch in xrange(training_epochs):
c1.append(train_da(numpy.asarray(train_set)))
print 'Training epoch %d, cost ' % epoch, c1[len(c1)-1]
for c in test_class:
m.features[c] = da.get_hidden_values(m.features[c]).eval()
m.train()
m.dump('model/model_da.out')
with open('model/da.out', 'w') as f:
pickle.dump(da, f, -1)
def evaluate(eval_data_dir, model_path):
m = ModelInterface.load(model_path)
files = [f for f in os.listdir(eval_data_dir) if re.search(r"\.wav", f)]
total, n_correct = 0, 0
for f in files:
total += 1
label, _ = f.split("_")
file = os.path.join(eval_data_dir, f)
fs, signal = read_wav(file)
pred, _ = m.predict(fs, signal)
logger.info("Input: {}, Output: {}".format(file, pred))
if label == pred:
n_correct += 1
logger.info("Accuracy: {}".format(n_correct / total))
def task_predict(input_files, input_model):
# 把输入的多个模型目录字符串分离为目录列表
input_models = [os.path.expanduser(k) for k in input_model.strip().split()]
# 把各个目录下的模型列表解压出来组合成一个迭代器
models = itertools.chain(*(glob.glob(m) for m in input_models))
# 生成并加载包括所有模型文件的列表
models = [ModelInterface.load(m) for m in models]
# 定义统计准确率的变量
right = 0
wrong = 0
num = 0
# 对每个预测音频文件提取特征并与每个模型匹配得到TOP结果
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
print(f)
feat = get_feature(fs, signal)
predict_result = []
# 每个音频文件分别匹配每个模型组并得出分数放到总列表
for model in models:
#print(model)
#m = ModelInterface.load(model)
results = model.predict(feat)
for result in results:
predict_result.append(result)
#print("predict_result:",predict_result)
# 对预测结果按分数作高到底排序
predict_result = sorted(predict_result,
key=operator.itemgetter(1),
reverse=True)
#print("sort_predict_result:", predict_result)
# 微信语音数据集的label格式
label = os.path.basename(f).split('_')[0] #[6:11]
#label=os.path.basename(f).split('(')[0]#[6:11]
# AISHELL数据集的label格式
#label=os.path.basename(f)[6:11]
predict = predict_result[0][0]
#print('Top:',predict_result[:10])
# 统计准确率
if label in predict:
right += 1
print('label:', label, ' predict:', predict, ' right')
else:
wrong += 1
print('label:', label, ' predict:', predict, ' wrong')
num += 1
print('All:', num, ' right:', right, ' wrong:', wrong, ' acc:',
right / num)
def task_predict(input_files, input_model):
start_time = time.time()
print('开始时间：',
time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(start_time)))
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = utils.read_wav(f)
label, probability = m.predict(fs, signal)
#print(probability)
if probability > -48:
print(f, '->', label)
else:
print(f, '->未识别到说话人')
end_time = time.time()
print('结束时间：', time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(end_time)))
print('共耗时', end_time - start_time)
def task_predict():
m = ModelInterface.load('data.bin')
predict_sound_path = os.path.join(os.getcwd(), 'predictSounds')
dirs = os.listdir(predict_sound_path)
wavs = []
if len(dirs) == 0:
print('No wav files found')
else:
for d in dirs:
ext = os.path.splitext(d)[-1].lower()
if ext == '.wav':
wavs.append(d)
for w in wavs:
sample_rate, signal = read_wav(os.path.join(predict_sound_path, w))
label = os.path.splitext(w)[0]
label2, score = m.predict(sample_rate, signal)
print(label, '->', label2, '->', score)
def task_predict(input_files, input_model):
total = 0
acc = 0
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
total += 1
fs, signal = read_wav(f)
label = m.predict(fs, signal)
print(f, '->', label, end=''),
if f.split('/')[-2] == label:
print("√")
acc += 1
else:
print('×')
acc = acc * 1.0 / total
print(acc)
def test( up_bound, lower_bound ):
m = ModelInterface.load('model/model_da.out')
with open('model/da.out', 'r') as f:
da = pickle.load(f)
count = 0;
test_dir = 'data/test/'
test_class = ['FAML_S', 'FDHH_S', 'FEAB_S', 'FHRO_S',
'FJAZ_S', 'FMEL_S', 'FMEV_S', 'FSLJ_S', 'FTEJ_S',
'FUAN_S', 'MASM_S', 'MCBR_S', 'MFKC_S', 'MKBP_S',
'MLKH_S', 'MMLP_S', 'MMNA_S', 'MNHP_S', 'MOEW_S',
'MPRA_S', 'MREM_S', 'MTLS_S']
file_name = ['1.wav', '2.wav']
for c in test_class:
for n in file_name:
fs, signal = wavfile.read(test_dir + c + n)
if(predict(m, fs, signal[:80000], da, up_bound, lower_bound) == c):
count = count + 1
print 'accuracy is:', (100.0*count)/(len(test_class)*len(file_name)), '%'
def feature_re_extract():
test_class = ['FAML_S', 'FDHH_S', 'FEAB_S', 'FHRO_S',
'FJAZ_S', 'FMEL_S', 'FMEV_S', 'FSLJ_S', 'FTEJ_S',
'FUAN_S', 'MASM_S', 'MCBR_S', 'MFKC_S', 'MKBP_S',
'MLKH_S', 'MMLP_S', 'MMNA_S', 'MNHP_S', 'MOEW_S',
'MPRA_S', 'MREM_S', 'MTLS_S']
m = ModelInterface.load('model/model.out')
# construct train set
train_set = []
up_bound = []
lower_bound = []
for c in test_class:
for i in m.features[c]:
train_set.append(i)
# put all values into -1~1
up_bound = []
lower_bound = []
for j in xrange(len(train_set[0])):
up_bound.append(train_set[0][j])
lower_bound.append(train_set[0][j])
for i in xrange(len(train_set)):
for j in xrange(len(train_set[0])):
up_bound[j] = max(up_bound[j], train_set[i][j])
lower_bound[j] = min(lower_bound[j], train_set[i][j])
for i in xrange(len(train_set)):
for j in xrange(len(train_set[0])):
train_set[i][j] = 2*((train_set[i][j]-lower_bound[j]) / (up_bound[j]-lower_bound[j]))-1
# construct stacked autoencoder
sda = mSdA(
layers = [39, 100]
)
sda.setMinMax(up_bound, lower_bound)
sda.train(train_set, 500) # use 500 as the batch size
for c in test_class:
m.features[c] = sda.get_hidden_values(m.features[c])
m.train()
m.dump('model/model_sda.out')
sda.dump('model/sda.out')
def task_predict(path, input_model):
m = ModelInterface.load(input_model)
files = []
sum, true = 0, 0
# r=root, d=directories, f = files
for r, d, f in os.walk(path):
for file in f:
if '.wav' in file:
files.append(os.path.join(r, file))
for f in files:
sum += 1
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
strPath = os.path.realpath(f)
y_true = os.path.basename(os.path.dirname(strPath))
if (label == y_true):
true += 1
print(f, '->', label, ", score->", score)
print('So file du doan dung: ', true)
print('Tong so file: ', sum)
print('accuracy: ', true / sum * 100, '%')
def test():
m = ModelInterface.load('model/model_sda.out')
sda = mSdA.load('model/sda.out')
count = 0
allsum = 0
test_dir = 'data/test/'
test_class = ['FAML_S', 'FDHH_S', 'FEAB_S', 'FHRO_S',
'FJAZ_S', 'FMEL_S', 'FMEV_S', 'FSLJ_S', 'FTEJ_S',
'FUAN_S', 'MASM_S', 'MCBR_S', 'MFKC_S', 'MKBP_S',
'MLKH_S', 'MMLP_S', 'MMNA_S', 'MNHP_S', 'MOEW_S',
'MPRA_S', 'MREM_S', 'MTLS_S']
file_name = ['1.wav', '2.wav']
for c in test_class:
for n in file_name:
fs, signal = wavfile.read(test_dir + c + n)
signal_size = 40000
for indx in xrange(len(signal)/signal_size):
allsum = allsum + 1
if(predict(m, fs, signal[indx*signal_size:(indx+1)*signal_size], sda) == c):
count = count + 1
print 'accuracy is:', (100.0*count)/(allsum), '%'
def task_train_single(wav_url, person_id):
if os.path.exists(model):
m = ModelInterface.load(model)
else:
m = ModelInterface()
if person_id in m.features:
return 'fail', 'aleady exist'
#下载训练语音文件
dest_dir = train_voice_dir + person_id
if not os.path.exists(dest_dir):
os.makedirs(dest_dir)
current_time = time.strftime("%Y%m%d%H%I%S", time.localtime(time.time()))
dest_wav = dest_dir + '/' + current_time + '_' + person_id + '.wav'
print(wav_url)
print(dest_wav)
utils.download_file(wav_url, dest_wav)
#获取下载好的训练语音文件
wavs = glob.glob(dest_dir + '/*.wav')
if len(wavs) == 0:
return 'fail', 'no wav files under this dir'
#train the wavs
for wav in wavs:
try:
fs, signal = utils.read_wav(wav)
m.enroll(person_id, fs, signal)
print("wav %s has been enrolled" % (wav))
except Exception as e:
print(wav + " error %s" % (e))
m.train_single(person_id)
m.dump(model)
return 'success', ''
def task_realtime_predict(input_model):
print('start')
m = ModelInterface.load(input_model)
# set recording parameter
CHUNK = 1024
FORMAT = pyaudio.paInt16
CHANNELS = 1
RATE = 16000
INTERVAL = 1
INITLEN = 2
p = pyaudio.PyAudio()
stream = p.open(format=FORMAT,
channels=CHANNELS,
rate=RATE,
input=True,
frames_per_buffer=CHUNK)
print("* recording")
frames = []
# fulfill the frame
for i in range(0, int(RATE / CHUNK * INITLEN)):
data = np.fromstring(stream.read(CHUNK), dtype=np.int16).tolist()
frames.append(data)
while True:
for i in range(0, int(RATE / CHUNK * INTERVAL)):
# 添加新的时间窗数据
frames.append(
np.fromstring(stream.read(CHUNK), dtype=np.int16).tolist())
# 去掉最老的时间窗数据
frames.remove(frames[0])
framesjoin = utils.flat_array(frames)
framesjoin = np.array(framesjoin)
label, probability = m.predict(16000, framesjoin)
print('当前说话人->', label)
required=True,
help="Path to model file")
parser.add_argument("--task",
required=True,
choices=["enroll", "predict"],
help='Task to do. Either "enroll" or "predict"')
args = parser.parse_args()
if args.task == "predict" and not os.path.isfile(args.model_path):
raise ValueError("Please provide valid model path")
r = sr.Recognizer()
r.pause_threshold = 1.0
if args.task == "enroll":
if os.path.isfile(args.model_path) and not args.overwrite_model:
model = ModelInterface.load(args.model_path)
else:
model = ModelInterface()
print("***** Enroll sound data for one speaker *****")
name = input("Enter your name: ")
name = name.strip()
print(
f"Hello {name}. Please input your voice {args.num_samples} times")
with tempfile.TemporaryDirectory() as tempdir:
i = 1
while i <= args.num_samples:
with sr.Microphone() as source:
audio = r.listen(source)
# Generate random filename
filename = os.path.join(
tempdir, name + "_" + str(uuid.uuid1()) + ".wav")
def task_predict(input_files, input_model, isDynamic):
"""
Predict the speaker from the given file(s)
Args:
input_files (string): full path to the speaker file
input_model (string): model trained to give the solution
"""
# Loads the model object and retrieve the number of speaker #
m = ModelInterface.load(input_model)
n_label = m.get_n_label()
# Computes the threshold (dynamic or static) #
if (isDynamic):
dyn_thrsh = m.get_dyn_threshold()
else:
threshold = 1 / n_label
# Creates an Evaluation object to save the results #
ev = Evaluation()
# Starts the prediction process #
print(input_files)
for f in glob.glob(os.path.expanduser(input_files)):
try:
start_time = time.time()
fs, signal = read_wav(f)
signal = signal / max(abs(signal))
# Extracts the features and predicts the label using the higher score within all possible speaker #
label, score = m.predict(fs, VAD_process(signal))
except Exception as e:
print(f + ' error %s' % (e))
# Retrieves the expected label from the directory (evaluation not real time only) #
root = os.path.split(f)
if (input_files[-9:] == "*/*/*.wav"):
root = os.path.split(root[0])
speaker = os.path.basename(root[0])
# Recognition process : If the given score is higher than the threshold, the label is correct #
# Else the speaker is not recognize #
if (isDynamic):
threshold = dyn_thrsh[label]
recog = (score > threshold)
# recog = True
if not (recog):
print(speaker, ' not recognize. ->', label, 'Score->', score)
else:
print(speaker, '->', label, ', score->', score)
# Adds the speaker and its results to the evaluation object #
ev.new(speaker, label, recog)
# Retrieves the Database label used and prints the accuracy #
path = os.path.split(root[0])[0]
DB_name = os.path.split(path)[0]
DB_name = os.path.basename(os.path.split(DB_name)[1])
print('Accuracy : ', ev.accuracy(), '\n')
ev.save(os.path.basename(path), n_label, DB_name,
(time.time() - start_time))
def task_predict(input_files, input_model):
m = ModelInterface.load(input_model)
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label, score = m.predict(fs, signal)
print ('You are', label, " GMM score of ", score)
)
for i in range(0, 3, 1):
if not (os.path.exists(input_model)):
input_model = input(
"Model does not exists ! " + str(3 - i) +
" attempts left. Enter relative path of the model (ex.: ./model/Pers_DB.out) : "
)
else:
break
if os.path.exists(input_model):
# Loading the model and starting the while loop #
print("Model found ! Starting real-time recognition process")
m = ModelInterface.load(input_model)
speaker = input(
"Write the name of the speaker (for evaluation purposes) :")
start_time = time.time()
while tmp < 5:
count += 1
buffer.record(chunk_size=sampling_rate) # 1 second of record
data = buffer.get_data()
data = np.frombuffer(data, 'int16')
# Predicting every 3 loop #
# Recording at 16000 Hz as sampling rate, (1 * 3) sec as buffer size and converting data in int16 type #
if count >= 3:
print(" + Predicting")
input_model = "model.out"
input_files = "./tmp.wav"
# m = ModelInterface.load(input_model)
m = model
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
label = m.predict(fs, signal)
# print max(label, key=operator.itemgetter(1))
if abs(label[0][1] - label[1][1]) > .0000002:
speakers_detected = [x for x in label if x[1] >= -.1]
if len(speakers_detected) > 0:
print max(speakers_detected, key=operator.itemgetter(1))
else:
print "___ not similar enough to know speaker"," Best guess:", max(label, key=operator.itemgetter(1))
else:
print "... predictions too similar"," Best guess:", max(label, key=operator.itemgetter(1))
mods = ModelInterface()
mods = mods.load("./model.out")
see_model(mods)
import operator
# find_user()
print "\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n"
for i in range(1,60):
find_user(mods)
def feature_re_extract():
#pdb.set_trace()
test_class = ['FAML_S', 'FDHH_S', 'FEAB_S', 'FHRO_S',
'FJAZ_S', 'FMEL_S', 'FMEV_S', 'FSLJ_S', 'FTEJ_S',
'FUAN_S', 'MASM_S', 'MCBR_S', 'MFKC_S', 'MKBP_S',
'MLKH_S', 'MMLP_S', 'MMNA_S', 'MNHP_S', 'MOEW_S',
'MPRA_S', 'MREM_S', 'MTLS_S']
m = ModelInterface.load('model/model.out')
# construct train set
train_set = []
up_bound = []
lower_bound = []
for c in test_class:
for i in m.features[c]:
train_set.append(i)
'''
# put all values into -1~1
up_bound = []
lower_bound = []
for j in xrange(len(train_set[0])):
up_bound.append(train_set[0][j])
lower_bound.append(train_set[0][j])
for i in xrange(len(train_set)):
for j in xrange(len(train_set[0])):
up_bound[j] = max(up_bound[j], train_set[i][j])
lower_bound[j] = min(lower_bound[j], train_set[i][j])
for i in xrange(len(train_set)):
for j in xrange(len(train_set[0])):
train_set[i][j] = 2*((train_set[i][j]-lower_bound[j]) / (up_bound[j]-lower_bound[j]))-1
'''
# construct autoencoder
train_data = T.dmatrix('train_data')
x = T.dmatrix('x') # the data is presented as rasterized images
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 10))
da = dA(
numpy_rng=rng,
theano_rng=theano_rng,
input=x,
n_visible=39,
n_hidden=30
)
cost, updates = da.get_cost_updates(
corruption_level=0.,
learning_rate=0.4
)
train_da = theano.function(
[train_data],
cost,
updates=updates,
givens={
x: train_data
}
)
# train first autoencoder
training_epochs = 20
c1 = []
for epoch in xrange(training_epochs):
c1.append(train_da(numpy.asarray(train_set)))
print 'Training 1st ae epoch %d, cost ' % epoch, c1[len(c1)-1]
# train second autoencoder
train_set2 = da.get_hidden_values(train_set).eval()
train_data = T.dmatrix('train_data')
x = T.dmatrix('x') # the data is presented as rasterized images
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 10))
da2 = dA(
numpy_rng=rng,
theano_rng=theano_rng,
input=x,
n_visible=30,
n_hidden=20
)
cost, updates = da2.get_cost_updates(
corruption_level=0.,
learning_rate=0.4
)
train_da2 = theano.function(
[train_data],
cost,
updates=updates,
givens={
x: train_data
}
)
training_epochs = 20
c1 = []
for epoch in xrange(training_epochs):
c1.append(train_da2(numpy.asarray(train_set2)))
print 'Training 2nd ae epoch %d, cost ' % epoch, c1[len(c1)-1]
for c in test_class:
m.features[c] = da2.get_hidden_values(da.get_hidden_values(m.features[c]).eval()).eval()
m.train()
m.dump('model/model_sda.out')
with open('model/da1.out', 'w') as f:
pickle.dump(da, f, -1)
with open('model/da2.out', 'w') as f:
pickle.dump(da2, f, -1)
return up_bound, lower_bound
def record_predict(input_file, input_model):
m = ModelInterface.load(input_model)
fs, signal = read_wav(input_file)
label = m.predict(fs, signal)
return label
def task_predict(input_files, input_model):
# 把输入的多个模型目录字符串分离为目录列表
input_models = [os.path.expanduser(k) for k in input_model.strip().split()]
# 把各个目录下的模型列表解压出来组合成一个迭代器
models = itertools.chain(*(glob.glob(m) for m in input_models))
# 生成并加载包括所有模型文件（skgmm.GMMSet object）的列表
models = [ModelInterface.load(m) for m in models]
if len(models) == 0:
print("No model file found in %s" % input_model)
sys.exit(1)
# 定义统计准确率的变量
right = 0
right1 = 0
wrong = 0
wrong1 = 0
num = 0
# 对每个预测音频文件提取特征并与每个模型匹配得到TOP结果
for f in glob.glob(os.path.expanduser(input_files)):
fs, signal = read_wav(f)
print(f)
feat = get_feature(fs, signal)
predict_result = []
# 每个音频文件分别匹配每个模型组并得出分数放到总列表
for model in models:
#print(model)
#m = ModelInterface.load(model)
results = model.predict_proba(feat)
print(results)
#print(results)
for result in results:
predict_result.append(result)
#print("predict_result:",predict_result)
# 对预测结果按分数作高到底排序
predict_result = sorted(predict_result,
key=operator.itemgetter(1),
reverse=True)
#print("sort_predict_result:", predict_result)
# 微信语音数据集的label格式
label = os.path.basename(f).split('_')[0] #[6:11]
#label=os.path.basename(f).split('(')[0]#[6:11]
# AISHELL数据集的label格式
# label=os.path.basename(f)[6:11]
predict = predict_result[0][0]
predict_score = predict_result[0][1]
# #print('Top:',predict_result[:10])
# 统计top1准确率
if label in predict:
right1 += 1
print('label:', label, ' predict:', predict, ' score:',
predict_score, ' top1 right')
else:
wrong1 += 1
print('label:', label, ' predict:', predict, ' score:',
predict_score, ' top1 wrong')
# 统计Top10准确率
predicts = []
predict_scores = []
for pre in predict_result[:10]:
predicts.append(pre[0])
predict_scores.append(pre[1])
if label in predicts:
right += 1
print('label:', label, ' predicts:', predicts, ' scores:',
predict_scores, ' top10 Right')
else:
wrong += 1
print('label:', label, ' predicts:', predicts, ' scores:',
predict_scores, ' top10 Wrong')
num += 1
print('top1:', num, ' right:', right1, ' wrong:', wrong1, ' top1 acc:',
right1 / num)
print('top10:', num, ' right:', right, ' wrong:', wrong, ' top10 acc:',
right / num)
def task_predict(input_files, input_model):
# 把输入的多个模型目录字符串分离为目录列表
input_models = [os.path.expanduser(k) for k in input_model.strip().split()]
# 把各个目录下的模型列表解压出来组合成一个迭代器
models = itertools.chain(*(glob.glob(m) for m in input_models))
# 生成并加载包括所有模型文件（skgmm.GMMSet object）的列表
models = [ModelInterface.load(m) for m in models]
if len(models) == 0:
print("No model file found in %s" % input_model)
sys.exit(1)
# 定义统计准确率的变量
right = 0
right1 = 0
wrong = 0
wrong1 = 0
num = 0
# 对每个预测音频文件提取特征并与每个模型匹配得到TOP结果
for f in glob.glob(os.path.expanduser(input_files)):
start_time = time.time()
fs, signal = read_wav(f)
print(f)
feat = get_feature(fs, signal)
#print("Get feature ", time.time() - start_time, " seconds")
predict_result = []
f_models = [(feat, m) for m in models]
#print(models)
# 每个音频文件分别匹配每个模型组并得出分数放到总列表
# for model in models:
# #start_time1 = time.time()
# #print(model)
# # 模型文件是一个元组：(label,gmm)
# score = model[1].score(feat)
# label=model[0]
# result=(label,score)
# #print(results)
# predict_result.append(result)
#print("Get one score ", time.time() - start_time1, " seconds")
pool = ThreadPool(2)
predict_result = pool.map(get_score, f_models)
pool.close()
pool.join()
#print(results)
#print("Get score ", time.time() - start_time, " seconds")
proba = GMMSet.softmax([i[1] for i in predict_result])
predict_result = [(predict_result[i][0], proba[i])
for i in range(len(proba))]
#print("predict_result:",predict_result)
# 对预测结果按分数作高到底排序
predict_result = sorted(predict_result,
key=operator.itemgetter(1),
reverse=True)
#print("sort_predict_result:", predict_result)
# 微信语音数据集的label格式
label = os.path.basename(f).split('_')[0] #[6:11]
#label=os.path.basename(f).split('(')[0]#[6:11]
# AISHELL数据集的label格式
# label=os.path.basename(f)[6:11]
predict = predict_result[0][0]
predict_score = predict_result[0][1]
print("Predict ", time.time() - start_time, " seconds")
# #print('Top:',predict_result[:10])
# 统计top1准确率
if label in predict:
right1 += 1
print('label:', label, ' predict:', predict, ' score:',
predict_score, ' top1 right')
else:
wrong1 += 1
print('label:', label, ' predict:', predict, ' score:',
predict_score, ' top1 wrong')
# 统计Top10准确率
predicts = []
predict_scores = []
for pre in predict_result[:10]:
predicts.append(pre[0])
predict_scores.append(pre[1])
if label in predicts:
right += 1
print('label:', label, ' predicts:', predicts, ' scores:',
predict_scores, ' top10 Right')
else:
wrong += 1
print('label:', label, ' predicts:', predicts, ' scores:',
predict_scores, ' top10 Wrong')
num += 1
print('top1:', num, ' right:', right1, ' wrong:', wrong1, ' top1 acc:',
right1 / num)
print('top10:', num, ' right:', right, ' wrong:', wrong, ' top10 acc:',
right / num)
