def iden(testfile,fa_data_dir,iden_model,max_sec, step_sec, frame_step):
# 读入测试数据、标签
print("Use {} for test".format(testfile))
iden_list = np.loadtxt(testfile, str,delimiter=",")
labels = np.array([int(i[1]) for i in iden_list])
voice_list = np.array([os.path.join(fa_data_dir, i[0]) for i in iden_list])
# Load model
print("Load model form {}".format(iden_model))
model = load_model(iden_model)
print("Start identifying...")
total_length = len(voice_list)
res, p_labels = [], []
buckets = build_buckets(max_sec, step_sec, frame_step)
for c, ID in enumerate(voice_list):
if c % 1000 == 0: print('Finish identifying for {}/{}th wav.'.format(c, total_length))
specs = get_fft_spectrum(ID, buckets)
v = model.predict(specs.reshape(1, *specs.shape, 1))
v = np.squeeze(v)
p_labels.append(np.argmax(v))
p_labels = np.array(p_labels)
compare = (labels == p_labels)
counts = sum(compare==True)
acc = counts/total_length
print(acc)
def get_fft_features_from_list_file(list_file, max_sec):
buckets = build_buckets(max_sec, c.BUCKET_STEP, c.FRAME_STEP)
result = pd.read_csv(list_file, delimiter=",")
result['features'] = result['filename'].apply(
lambda x: get_fft_spectrum(x, buckets))
# iterate over rows with iterrows
dataX = []
dataY = []
for index, row in result.iterrows():
x = row['features']
y = row['speaker']
# print(row['speaker'])
if isinstance(y, str):
y = y.replace(c.SPEAKER_PREFIX, '')
y = int(y, 10)
dataY.append(y)
values = x.reshape(*x.shape, 1)
# print("here", values.shape)
if (values.shape[1] == c.MAX_SEC * 100): #same time
dataX.append(values)
dataX = np.stack((dataX), axis=0)
dataY = np.asarray(dataY)
# print(dataY)
print("X.shape: {}, Y.shape{}".format(dataX.shape, dataY.shape))
return dataX, dataY
def score(testfile, fa_data_dir, test_model_path, max_sec, step_sec,
frame_step, metric):
print("Use {} for test".format(testfile))
verify_list = np.loadtxt(testfile, str)
verify_lb = np.array([int(i[0]) for i in verify_list])
list1 = np.array([os.path.join(fa_data_dir, i[1]) for i in verify_list])
list2 = np.array([os.path.join(fa_data_dir, i[2]) for i in verify_list])
total_list = np.concatenate((list1, list2))
unique_list = np.unique(total_list)
# Load model
print("Load model form {}".format(test_model_path))
model = load_model(test_model_path)
model = Model(inputs=model.layers[0].input,
outputs=model.layers[34].output) # 取 fc7 层的输出（1024,）
print("Start testing...")
total_length = len(unique_list) # 4715
feats, scores, labels = [], [], []
buckets = build_buckets(max_sec, step_sec, frame_step)
for c, ID in enumerate(unique_list):
if c % 50 == 0:
print('Finish extracting features for {}/{}th wav.'.format(
c, total_length))
specs = get_fft_spectrum(ID, buckets)
v = model.predict(specs.reshape(1, *specs.shape, 1))
feats += [v]
feats = np.array(feats)
# 计算相似度
for c, (p1, p2) in enumerate(zip(list1, list2)):
ind1 = np.where(unique_list == p1)[0][0]
ind2 = np.where(unique_list == p2)[0][0]
v1 = feats[ind1, 0, 0]
v2 = feats[ind2, 0, 0]
# print(v1 - v2)
# print(np.sum(v1*v1))
# print(np.sum(v2*v2))
scores += [np.sum(v1 * v2)]
labels += [verify_lb[c]]
# print("scores: ",scores)
# print("labels: ",labels)
# if c>0:
# break
scores = np.array(scores)
labels = np.array(labels)
eer = calculate_eer(labels, scores)
print("EER: {}".format(eer))
def get_embeddings_from_list_file(model, list_file, max_sec):
buckets = build_buckets(max_sec, c.BUCKET_STEP, c.FRAME_STEP)
result = pd.read_csv(list_file, delimiter=",")
result['features'] = result['filename'].apply(
lambda x: get_fft_spectrum(x, buckets))
result['embedding'] = result['features'].apply(
lambda x: np.squeeze(model.predict(x.reshape(1, *x.shape, 1))))
return result[['filename', 'speaker', 'embedding']]
def get_ids_random_feature(self, Id):
id_directory = os.path.join(self.directory, Id)
id_random_speech = os.path.join(
id_directory, random.choice(os.listdir(id_directory)))
id_random_utterance_path = os.path.join(
id_random_speech, random.choice(os.listdir(id_random_speech)))
fft = get_fft_spectrum(id_random_utterance_path, self.buckets)
return fft
def get_train_list(path):
buckets = build_buckets(c.MAX_SEC, c.BUCKET_STEP, c.FRAME_STEP)
read_csv = pd.read_csv(path, delimiter=",")
print("Preprocessing voice data...")
read_csv["voice"] = read_csv["filename"].apply(
lambda x: get_fft_spectrum(c.FA_DIR + x, buckets))
read_csv["lable"] = read_csv["speaker"].apply(lambda x: to_one_hot(x - 1))
return read_csv
def get_embeddings_from_file(model, file_path, max_sec):
buckets = build_buckets(max_sec, c.BUCKET_STEP, c.FRAME_STEP)
result = pd.DataFrame({'filename': [file_path]})
result['features'] = result['filename'].apply(
lambda x: get_fft_spectrum(x, buckets))
result['embedding'] = result['features'].apply(
lambda x: np.squeeze(model.predict(x.reshape(1, *x.shape, 1))))
# result['speaker'] = 19
return result[['filename', 'embedding']]
def get_embeddings_from_list_file(model, list_file, max_sec):
buckets = build_buckets(max_sec, c.BUCKET_STEP, c.FRAME_STEP)
result = pd.read_csv(list_file, delimiter=",")
result['features'] = result['filename'].apply(
lambda x: get_fft_spectrum(x, buckets)
) # Get as much fft as possible from the buckets 'size' for each filename
result['embedding'] = result['features'].apply(
lambda x: np.squeeze(model.predict(x.reshape(1, *x.shape, 1)))
) # Squeeze function removes all single-dimensionnal entries from the shape of an array
return result[['filename', 'speaker', 'embedding']]
def get_ids_features(self, Id, count=np.iinfo(np.int32).max):
yielded = 0
features = []
for path in self.__file_iterator__(Id):
if yielded == count:
break
feature = get_fft_spectrum(path, self.buckets)
features.append(feature)
yielded += 1
return features
def get_embeddings_from_list_file(model, list_file, max_sec): buckets = build_buckets(max_sec, c.BUCKET_STEP, c.FRAME_STEP) result = pd.read_csv(list_file, delimiter=",") # print(result) # time.sleep(10) result['features'] = result['filename'].apply(lambda x: get_fft_spectrum(c.FA_DIR+x, buckets)) # 获取每个人的语音特征 矩阵 # print(result['features']) # time.sleep(10) result['embedding'] = result['features'].apply(lambda x: np.squeeze(model.predict(x.reshape(1,*x.shape,1))))# to input (1, 512, 1000, 1) per voice(feature) # print(result['embedding']) # time.sleep(10) return result[['filename','speaker','embedding']]
def load_wave(wave_file):
buckets = build_buckets(10, 1, 0.01)
data = get_fft_spectrum(wave_file, buckets)
if data.shape[1] == 300:
pass
else:
start = np.random.randint(0, data.shape[1] - 300)
data = data[:, start:start + 300]
data = np.expand_dims(data, -1)
return data
def iden(testfile, fa_data_dir, iden_model, max_sec, step_sec, frame_step, dim,
batch_size, n_classes, epoch):
# 读入测试数据、标签
print("Use {} for test".format(testfile))
iden_list = np.loadtxt(testfile, str, delimiter=",")
voice_list = np.array([os.path.join(fa_data_dir, i[0]) for i in iden_list])
total_length = len(voice_list)
device = torch.device('cuda')
labels = torch.tensor([int(i[1]) for i in iden_list]).to(device)
criterion = nn.CrossEntropyLoss()
# Load model
print("Load model form {}".format(iden_model))
model = torch.load(iden_model).to(device)
print("Start identifying...")
acc = 0
loss = 0
model.eval()
for num, ID in enumerate(voice_list):
if num % 100 == 0:
print('Finish identifying for {}/{}th wav.'.format(
num, total_length))
b_data = torch.tensor(
get_fft_spectrum(ID,
build_buckets(max_sec, step_sec, frame_step),
mode="test").tolist()).to(device)
b_data = b_data.unsqueeze(0).to(device)
with torch.no_grad():
eval_predict, tmp_eval_accuracy = model(b_data.unsqueeze(0),
labels[num].unsqueeze(0),
1, n_classes)
tmp_eval_loss = criterion(eval_predict,
labels[num].unsqueeze(0).long()).to(device)
loss += tmp_eval_loss.item()
acc += tmp_eval_accuracy
loss /= total_length
acc /= total_length
viz.line(torch.tensor([loss]), [epoch],
update='append',
win='test_loss',
opts={'title': 'test_loss'})
viz.line(torch.tensor([acc]), [epoch + 1],
update='append',
win='test_acc',
opts={'title': 'test_acc'})
print("eval_loss:", loss)
print("eval_acc:", acc)
def _gene_Data(self, list_IDs_temp, indexes):
'得到频谱数组和类标签，以输入模型进行训练'
b_data = np.empty((self.batch_size,) + self.dim)
b_labels = np.empty((self.batch_size,), dtype=int)
for i, ID in enumerate(list_IDs_temp):
b_data[i, :, :, 0] = get_fft_spectrum(ID, self.buckets)
b_labels[i] = self.labels[indexes[i]] # 0~n-1
# b_labels[i] = self.labels[indexes[i]] - 1 # 1~n
b_labels = keras.utils.to_categorical(b_labels, num_classes=self.n_classes)
b_labels = b_labels.reshape(self.batch_size, 1, 1, self.n_classes)
# os.system("pause")
return b_data, b_labels
def get_np_list(file_list, buckets):
voice = np.empty()
for pt in range(len(file_list)):
np.concatenate(get_fft_spectrum(c.FA_DIR + pt, buckets))
print(voice.shape)
def get_embedding(model, wav_file, max_sec):
buckets = build_buckets(max_sec, c.BUCKET_STEP, c.FRAME_STEP)
signal = get_fft_spectrum(wav_file, buckets)
embedding = np.squeeze(model.predict(signal.reshape(1, *signal.shape, 1)))
return embedding
def give_vggvox_input(input_file_path):
buckets = build_buckets(c.MAX_SEC, c.BUCKET_STEP, c.FRAME_STEP)
file = pd.read_csv(input_file_path, delimiter=",")
output = file['filename'].apply(lambda x: get_fft_spectrum(x, buckets))
return output
def give_vggvox_input_simple(input_file_path):
buckets = build_buckets(c.MAX_SEC, c.BUCKET_STEP, c.FRAME_STEP)
output = get_fft_spectrum(input_file_path, buckets)
return output
