def create_tf_record(self, *, prefix, subset_size, parallel=True):
counter = 0
p = multiprocessing.Pool(multiprocessing.cpu_count())
for i in range(0, len(self.clean_filenames), subset_size):
tfrecord_filename = 'D:/downloads/dataset/' + prefix + '_' + str(
counter) + '.tfrecords'
if os.path.isfile(tfrecord_filename):
print(f"Skipping {tfrecord_filename}")
counter += 1
continue
writer = tf.io.TFRecordWriter(tfrecord_filename)
clean_filenames_sublist = self.clean_filenames[i:i + subset_size]
print(f"Processing files from: {i} to {i + subset_size}")
if parallel:
out = p.map(self.parallel_audio_processing,
clean_filenames_sublist)
else:
out = [
self.parallel_audio_processing(filename)
for filename in clean_filenames_sublist
]
for o in out:
noise_stft_magnitude = o[0]
clean_stft_magnitude = o[1]
noise_stft_phase = o[2]
noise_stft_mag_features = prepare_input_features(
noise_stft_magnitude, numSegments=8, numFeatures=129)
noise_stft_mag_features = np.transpose(noise_stft_mag_features,
(2, 0, 1))
clean_stft_magnitude = np.transpose(clean_stft_magnitude,
(1, 0))
noise_stft_phase = np.transpose(noise_stft_phase, (1, 0))
noise_stft_mag_features = np.expand_dims(
noise_stft_mag_features, axis=3)
clean_stft_magnitude = np.expand_dims(clean_stft_magnitude,
axis=2)
for x_, y_, p_ in zip(noise_stft_mag_features,
clean_stft_magnitude, noise_stft_phase):
y_ = np.expand_dims(y_, 2)
example = get_tf_feature(x_, y_, p_)
writer.write(example.SerializeToString())
counter += 1
writer.close()
def __init__(self, data_dir, data_type, transform=None):
"""
Args:
data_dir (string): Path to data directory of audio files.
data_type (string): 'training', 'test' or 'validation'
transform (callable, optional): Optional transform to be applied on a sample.
"""
self.data_dir = data_dir
self.transform = transform
clean_data = os.path.join(self.data_dir, "clean")
noisy_data = os.path.join(self.data_dir, "noisy")
clean_files = os.listdir(os.path.join(clean_data, data_type))
noisy_files = os.listdir(os.path.join(noisy_data, data_type))
# check if the # of noisy files = # of clean files
assert len(clean_files) == len(noisy_files)
clean_files_full_path = [
os.path.join(clean_data, data_type, filename)
for filename in clean_files
]
noisy_files_full_path = [
os.path.join(noisy_data, data_type, filename)
for filename in noisy_files
]
print("Start processing", data_type, "data")
for idx in range(len(clean_files)):
clean_new_array = np.load(clean_files_full_path[idx])
noisy_new_array = np.load(noisy_files_full_path[idx])
# Do some of the preprocessing here
noisy_new_array, clean_new_array = prepare_input_features(
noisy_new_array, clean_new_array, 8, 129)
if idx == 0:
noisy_final = np.copy(noisy_new_array)
clean_final = np.copy(clean_new_array)
else:
noisy_final = np.concatenate((noisy_final, noisy_new_array))
clean_final = np.concatenate((clean_final, clean_new_array))
self.noisy_data = torch.from_numpy(noisy_final).float()
self.clean_data = torch.from_numpy(clean_final).float()
def create_tf_record(self, *, prefix, subset_size, parallel=True):
counter = 0
# 多进程处理方式，多核同时处理，提高速度
#p = multiprocessing.Pool(multiprocessing.cpu_count())
# 假设有10000条语音，每一批2000个的话就从0到1999个为第一个训练集tfrecords，一共5个训练集
print(len(self.clean_filenames))
print(subset_size)
# 遍历传入的每一个数据文件，步长是subset_size,训练集设置为2000步长，测试集是1000步长
for i in range(0, len(self.clean_filenames), subset_size):
#记录每个训练集的名称
tfrecord_filename = './records/' + prefix + '_' + str(
counter) + '.tfrecords'
#如果文件存在则跳过，counter计数器加一
if os.path.isfile(tfrecord_filename):
print(f"Skipping {tfrecord_filename}")
counter += 1
continue
#创建tfrecord文件,等待写入
writer = tf.io.TFRecordWriter(tfrecord_filename)
# 从第i个到第i+subset_size个为一个语音批次
clean_filenames_sublist = self.clean_filenames[i:i + subset_size]
print(f"Processing files from: {i} to {i + subset_size}")
#if parallel:
# out = p.map(self.parallel_audio_processing, clean_filenames_sublist)
# else:
# out是将一个批次的语音处理后的列表（重点部分）
# self.parallel_audio_processing(filename)是将语音文件加噪后傅里叶变换之后的数据(核心部分)
out = [
self.parallel_audio_processing(filename)
for filename in clean_filenames_sublist
]
print(len(out))
for o in out:
noise_stft_magnitude = o[0]
clean_stft_magnitude = o[1]
noise_stft_phase = o[2]
noise_stft_mag_features = prepare_input_features(
noise_stft_magnitude, numSegments=8, numFeatures=129)
noise_stft_mag_features = np.transpose(noise_stft_mag_features,
(2, 0, 1))
clean_stft_magnitude = np.transpose(clean_stft_magnitude,
(1, 0))
noise_stft_phase = np.transpose(noise_stft_phase, (1, 0))
noise_stft_mag_features = np.expand_dims(
noise_stft_mag_features, axis=3) # 再加一维shape=(201,129,8,1)
clean_stft_magnitude = np.expand_dims(
clean_stft_magnitude, axis=2) # 再加一维shape=(201,129,1)
for x_, y_, p_ in zip(noise_stft_mag_features,
clean_stft_magnitude, noise_stft_phase):
y_ = np.expand_dims(y_, 2)
example = get_tf_feature(x_, y_, p_)
# 写入数据
writer.write(example.SerializeToString())
# 计数器加一，关闭文件
counter += 1
writer.close()
1j * phase) # that fixes the abs() ope previously done
features = np.transpose(features, (1, 0))
return noiseAudioFeatureExtractor.get_audio_from_stft_spectrogram(
features)
#return noiseAudioFeatureExtractor.get_audio_from_stft_spectrogram_GL(np.abs(features))
noisyPhase = np.angle(noise_stft_features)
print(noisyPhase.shape)
noise_stft_features = np.abs(noise_stft_features)
mean = np.mean(noise_stft_features)
std = np.std(noise_stft_features)
noise_stft_features = (noise_stft_features - mean) / std
predictors = prepare_input_features(noise_stft_features, numSegments,
numFeatures)
predictors = np.reshape(
predictors,
(predictors.shape[0], predictors.shape[1], 1, predictors.shape[2]))
predictors = np.transpose(predictors, (3, 0, 1, 2)).astype(np.float32)
print('predictors.shape:', predictors.shape)
STFTFullyConvolutional = model.predict(predictors)
print(STFTFullyConvolutional.shape)
denoisedAudioFullyConvolutional = revert_features_to_audio2(
STFTFullyConvolutional, noisyPhase, mean, std)
print("Min:", np.min(denoisedAudioFullyConvolutional), "Max:",
np.max(denoisedAudioFullyConvolutional))
# ipd.Audio(data=denoisedAudioFullyConvolutional, rate=fs) # load a local WAV file