dataset question

[zhfqd1]

Thank you very much for your reply, sir. I still have some questions about dataset preparation:
1、Do ori_wav and codec_wav correspond to high-quality audio and low-quality audio respectively? Is my understanding correct?
2、Why does the configuration file use HDF5, while the training code reads WAV audio? What is the difference between these two?
3、What is the final format of the audio data fed into the model? Is it HDF5 or WAV?
4、Is there any requirement for the sampling rate?
5、Is it necessary to use the Apollo-data-preprocess code to process into HDF5 format before training the model?

[JusperLee]

Thank you for your questions! Let me address each of them:

1. ori_wav and codec_wav: Yes, your understanding is correct. ori_wav corresponds to high-quality audio and codec_wav corresponds to low-quality audio.

2. HDF5 vs WAV format: The reason we use HDF5 format in the configuration is because our company's storage is not local, and HDF5 provides faster read speeds for remote storage access. You can absolutely use WAV format if you prefer. It's important to note that HDF5 is not an audio format itself - it's a data container format that stores the audio data efficiently.

3. Input data format: The HDF5 files store audio segments that have been processed by VAD (Voice Activity Detection). The actual audio data inside can be accessed and used by the model.

4. Sample rate: Currently, our default sample rate is 44.1kHz.

5. Apollo-data-preprocess requirement: No, it's not mandatory to use Apollo-data-preprocess to convert data to HDF5 format. You can modify the datasets file to directly work with WAV files if that better suits your workflow.

Feel free to adapt the data loading pipeline to your needs.

[zhfqd1]

Thank you very much for your reply, sir. Can I understand it this way? 1、In the file musdb_moisesdb_datamodule.py, can the part processed by h5py be modified to the reading method required for my current audio format? 2、Is this code also not mandatory? def codec_simu(wav, sr=16000, options={'bitrate':'random','compression':'random', 'complexity':'random', 'vbr':'random'}):     if options['bitrate'] == 'random':         options['bitrate'] = random.choice([24000, 32000, 48000, 64000, 96000, 128000])     compression = int(options['bitrate']//1000)     param = {'format': "mp3", "compression": compression}     wav_encdec = apply_codec(wav, sr, **param)     if wav_encdec.shape[-1] >= wav.shape[-1]:         wav_encdec = wav_encdec[...,:wav.shape[-1]]     else:         wav_encdec = torch.cat([wav_encdec, wav[..., wav_encdec.shape[-1]:]], -1)     tau = match2(wav, wav_encdec)      wav_encdec = torch.roll(wav_encdec, -tau[0], -1)     return wav_encdec  3、Are there any requirements for high-quality audio and low-quality audio? 4、Can ./hdf5_datas in the configuration file also be in WAV or FLAC format?

…

------------------ 原始邮件 ------------------ 发件人: "JusperLee/Apollo" ***@***.***>; 发送时间: 2025年6月4日(星期三) 下午4:37 ***@***.***>; ***@***.******@***.***>; 主题: Re: [JusperLee/Apollo] dataset question (Issue #26) JusperLee left a comment (JusperLee/Apollo#26) Thank you for your questions! Let me address each of them: ori_wav and codec_wav: Yes, your understanding is correct. ori_wav corresponds to high-quality audio and codec_wav corresponds to low-quality audio. HDF5 vs WAV format: The reason we use HDF5 format in the configuration is because our company's storage is not local, and HDF5 provides faster read speeds for remote storage access. You can absolutely use WAV format if you prefer. It's important to note that HDF5 is not an audio format itself - it's a data container format that stores the audio data efficiently. Input data format: The HDF5 files store audio segments that have been processed by VAD (Voice Activity Detection). The actual audio data inside can be accessed and used by the model. Sample rate: Currently, our default sample rate is 44.1kHz. Apollo-data-preprocess requirement: No, it's not mandatory to use Apollo-data-preprocess to convert data to HDF5 format. You can modify the datasets file to directly work with WAV files if that better suits your workflow. Feel free to adapt the data loading pipeline to your needs. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[zhfqd1]

    Hey expert, I have another question. When executing python train.py --conf_dir=configs/apollo.yaml, an error occurs at the step ori_data, codec_data = batch: [rank0]: ValueError: too many values to unpack (expected 2). Does this indicate any requirements for the dataset?

…

------------------ 原始邮件 ------------------ 发件人: "JusperLee/Apollo" ***@***.***>; 发送时间: 2025年6月4日(星期三) 下午4:37 ***@***.***>; ***@***.******@***.***>; 主题: Re: [JusperLee/Apollo] dataset question (Issue #26) JusperLee left a comment (JusperLee/Apollo#26) Thank you for your questions! Let me address each of them: ori_wav and codec_wav: Yes, your understanding is correct. ori_wav corresponds to high-quality audio and codec_wav corresponds to low-quality audio. HDF5 vs WAV format: The reason we use HDF5 format in the configuration is because our company's storage is not local, and HDF5 provides faster read speeds for remote storage access. You can absolutely use WAV format if you prefer. It's important to note that HDF5 is not an audio format itself - it's a data container format that stores the audio data efficiently. Input data format: The HDF5 files store audio segments that have been processed by VAD (Voice Activity Detection). The actual audio data inside can be accessed and used by the model. Sample rate: Currently, our default sample rate is 44.1kHz. Apollo-data-preprocess requirement: No, it's not mandatory to use Apollo-data-preprocess to convert data to HDF5 format. You can modify the datasets file to directly work with WAV files if that better suits your workflow. Feel free to adapt the data loading pipeline to your needs. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[JusperLee]

Sorry for the late reply. Please avoid submitting multiple issues.

1. You can directly use audio as input without any problem.
2. This code is for calculating the Codec. If you already have pre-generated results, you don’t need to use this code.
3. FLAC and WAV are essentially the same in terms of usage here.