下面的代码遍历音频文件的目录(约50k),并将它们转换为声谱图图像,并将它们保存在同一顶级目录中。
def plot_and_save(denoised_data, f_name):
fig, ax = plt.subplots()
i = 0
# Add this line to show plots else ignore warnings
# plt.ion()
ax.imshow(denoised_data)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
fig.set_size_inches(10, 10)
fig.savefig(
f"{f_name}" + "_{:04d}.png".format(i),
dpi=80,
bbox_inches="tight",
quality=95,
pad_inches=0.0)
ax.draw_artist(ax.xaxis)
ax.draw_artist(ax.yaxis)
i += 1
def standardize_and_plot(sampling_rate, file_path_image):
logger.info(f"All files will be resampled to {sampling_rate}Hz")
output_image_folder = "PreProcessed_image/"
for dirs, subdirs, files in os.walk(file_path_image):
for i, file in enumerate(files):
if file.endswith(('.wav', '.WAV')):
logger.info(f"Pre-Processing file: {file}")
data, sr = librosa.core.load(
os.path.join(dirs, file), sr=sampling_rate, res_type='kaiser_fast')
target_path = os.path.join(output_image_folder, dirs)
pcen_S = apply_per_channel_energy_norm(data, sr)
denoised_data = wavelet_denoising(pcen_S)
work_dir = os.getcwd()
if not os.path.exists(target_path):
os.makedirs(target_path)
os.chdir(target_path)
f_name, _ = os.path.splitext(os.path.basename(file))
plot_and_save(denoised_data, f_name)
os.chdir(work_dir)
if __name__ == '__main__':
chunkSize = 3
sampling_rate = 44100
file_path_audio = 'Recordings'
file_path_audio = "data/"
output_audio_folder = "PreProcessed_audio/"
file_path_image = os.path.join(output_audio_folder, file_path_audio)
standardize_and_plot(sampling_rate, file_path_image)
如何优化plot_and_save()方法?将大量图像保存在磁盘上需要花费大量时间(约2-3 fps)。我正在为此目的使用Google Colab。
保存这些图像大约半小时后,colab崩溃,表明已使用所有RAM。有没有办法知道内存泄漏在哪里?还是有办法加快打印和保存过程?