When I retrain the model by models/DeepFilterNet2/config.ini, the result is strange~

[Vinttt]

Thanks for your awesome work!
I'm trying to recurrence the training through models/DeepFilterNet2/config.ini, but when I use DeepFilterNet/df/enhance.py to test the retrained model, output wavs is nearly to silence.
This is my retraining step:Thanks for your awesome work!
I'm trying to recurrence the training through models/DeepFilterNet2/config.ini, but when I use DeepFilterNet/df/enhance.py to test the retrained model, output wavs is nearly to silence.
This is my retraining step:
(1) copy the config.ini of DeepFilterNet2

    cd DeepFilterNet
    mkdir base_dir
    cp ../models/DeepFilterNet2/config.ini ./base_dir

(2) preparing data
take VocalSet_48kHz_mono_000_NA_NA.tar.bz2 as clean data
take noise_fullband/datasets_fullband.noise_fullband.audioset_000.tar.bz2 as noise data

    find ../scripts/out/datasets_fullband/clean_fullband/VocalSet_48kHz_mono -iname "*.wav" > training_set_speech.txt
    python df/scripts/prepare_data.py --sr 48000 speech training_set_speech.txt ./data_dir/TRAIN_SET_SPEECH.hdf5
    find ../scripts/out/datasets_fullband/noise_fullband -iname "*.wav" > training_set_noise.txt
    python df/scripts/prepare_data.py --sr 48000 noise training_set_noise.txt ./data_dir/TRAIN_SET_NOISE.hdf5

this is my dataset.cfg：
{
"train": [
[
"TRAIN_SET_SPEECH.hdf5",
1.0
],
[
"TRAIN_SET_NOISE.hdf5",
1.0
],
[
"TRAIN_SET_RIR.hdf5",
0.0
]
],
"valid": [
[
"TRAIN_SET_SPEECH.hdf5",
0.2
],
[
"TRAIN_SET_NOISE.hdf5",
0.2
],
[
"TRAIN_SET_RIR.hdf5",
0.0
]
],
"test": [
[
"TRAIN_SET_SPEECH.hdf5",
0.2
],
[
"TRAIN_SET_NOISE.hdf5",
0.2
],
[
"TRAIN_SET_RIR.hdf5",
0.0
]
]
}

(3) start retraining
python df/train.py dataset.cfg ./data_dir/ ./base_dir/

(4) test retrained model

python DeepFilterNet/df/enhance.py -m DeepFilterNet/base_dir2 --output-dir DeepFilterNet/test_dir_out/ DeepFilterNet/test_dir_48k/interview_48k.wav

but the output of the reatrained model is just like this

I’m not sure which step is wrong or missed, have you ever met the same problem?

[Rikorose]

If you are only training with vocal set, I am not surprised that the results are not decent. Starting point for debugging would be full training logs, and some samples in the summaries folder.

[BWMa]

how do you download the clean_fullband dataset? excute download_process_dns4.sh ?

[Rikorose]

You should ask this at the dns repo

[Vinttt]

If you are only training with vocal set, I am not surprised that the results are not decent. Starting point for debugging would be full training logs, and some samples in the summaries folder.

I finally made it, thks a lot! The model performance is excellent!

And a small bug. The vocalset is set to 48000 hz when it's real samplerate is 16000 hz, in split_hdf5.py you correct the samplerate to 16000. But in summaries folder, the examples wavs are still in wrong samplerate.

[Vinttt]

I finally made it, thks a lot! The model performance is excellent!

And a small bug. The vocalset is set to 48000 hz when it's real samplerate is 16000 hz, in split_hdf5.py you correct the samplerate to 16000. But in summaries folder, the examples wavs are still in wrong samplerate.

[Penguin168]

I finally made it, thks a lot! The model performance is excellent!

And a small bug. The vocalset is set to 48000 hz when it's real samplerate is 16000 hz, in split_hdf5.py you correct the samplerate to 16000. But in summaries folder, the examples wavs are still in wrong samplerate.

Could I ask how did you solve it?

I take Voicebank+Demand as my training set, but the output of the reatrained model is just the noise, and I realized all the samples of that dataset are 16khz now, so I might take my cue from your method.

[Vinttt]

I finally made it, thks a lot! The model performance is excellent!
And a small bug. The vocalset is set to 48000 hz when it's real samplerate is 16000 hz, in split_hdf5.py you correct the samplerate to 16000. But in summaries folder, the examples wavs are still in wrong samplerate.

Could I ask how did you solve it?

I take Voicebank+Demand as my training set, but the output of the reatrained model is just the noise, and I realized all the samples of that dataset are 16khz now, so I might take my cue from your method.

In df/scripts/prepare_data.py, the following code

if meta.sample_rate != self.sr:
    # Load as normalized float32 and resample
    x, sr = torchaudio.load(file, normalize=True)
    x = resample(x, sr, new_sr=self.sr, method="kaiser_best")
else:
    x, sr = torchaudio.load(file, normalize=False)

if don't resample the signal, normalize = False;
if resample the signal, normalize=True, this may leading to error training output, you can change the code to

if meta.sample_rate != self.sr:
    # Load as normalized float32 and resample
    x, sr = torchaudio.load(file, normalize=True)
    x = resample(x, sr, new_sr=self.sr, method="kaiser_best")
    x = torch.tensor(x * 32768, dtype=torch.int16)
else:
    x, sr = torchaudio.load(file, normalize=False)

Hope to solve your problem.

[github-actions]

This issue is stale because it has been open 90 days with no activity. Remove stale label or comment or this will be closed in 7 days.

[QQQQQQQQY]

我终于成功了，非常感谢！模型性能出色！还有一个小虫子。当实际采样率为 48000 Hz 时，人声集设置为 16000 Hz，split_hdf5.py将采样率更正为 16000。但是在摘要文件夹中，示例波形仍然处于错误的采样率。

我能问你是怎么解决的吗？
我以 Voicebank+Demand 作为我的训练集，但重新训练模型的输出只是噪声，我意识到该数据集的所有样本现在都是 16khz，所以我可能会从你的方法中得到提示。

在 df/scripts/prepare_data.py 中，以下代码

if meta.sample_rate != self.sr:
    # Load as normalized float32 and resample
    x, sr = torchaudio.load(file, normalize=True)
    x = resample(x, sr, new_sr=self.sr, method="kaiser_best")
else:
    x, sr = torchaudio.load(file, normalize=False)

如果不对信号进行重新采样，则归一化 = False;如果对信号进行重采样，归一化=True，这可能会导致错误训练输出，您可以将代码更改为

if meta.sample_rate != self.sr:
    # Load as normalized float32 and resample
    x, sr = torchaudio.load(file, normalize=True)
    x = resample(x, sr, new_sr=self.sr, method="kaiser_best")
    x = torch.tensor(x * 32768, dtype=torch.int16)
else:
    x, sr = torchaudio.load(file, normalize=False)

我终于成功了，非常感谢！模型性能出色！还有一个小虫子。当实际采样率为 48000 Hz 时，人声集设置为 16000 Hz，split_hdf5.py将采样率更正为 16000。但是在摘要文件夹中，示例波形仍然处于错误的采样率。

我能问你是怎么解决的吗？
我以 Voicebank+Demand 作为我的训练集，但重新训练模型的输出只是噪声，我意识到该数据集的所有样本现在都是 16khz，所以我可能会从你的方法中得到提示。

在 df/scripts/prepare_data.py 中，以下代码

if meta.sample_rate != self.sr:
    # Load as normalized float32 and resample
    x, sr = torchaudio.load(file, normalize=True)
    x = resample(x, sr, new_sr=self.sr, method="kaiser_best")
else:
    x, sr = torchaudio.load(file, normalize=False)

如果不对信号进行重新采样，则归一化 = False;如果对信号进行重采样，归一化=True，这可能会导致错误训练输出，您可以将代码更改为

if meta.sample_rate != self.sr:
    # Load as normalized float32 and resample
    x, sr = torchaudio.load(file, normalize=True)
    x = resample(x, sr, new_sr=self.sr, method="kaiser_best")
    x = torch.tensor(x * 32768, dtype=torch.int16)
else:
    x, sr = torchaudio.load(file, normalize=False)

希望能解决您的问题。

你好，我也是遇到这个问题，按照你所说的更改依旧没有改变，请问是否还有其他解决措施，期待您的回复，感谢

[QQQQQQQQY]

I finally made it, thks a lot! The model performance is excellent!
And a small bug. The vocalset is set to 48000 hz when it's real samplerate is 16000 hz, in split_hdf5.py you correct the samplerate to 16000. But in summaries folder, the examples wavs are still in wrong samplerate.

Could I ask how did you solve it?

I take Voicebank+Demand as my training set, but the output of the reatrained model is just the noise, and I realized all the samples of that dataset are 16khz now, so I might take my cue from your method.

Hello, I also encountered this problem, according to the changes you said remains unchanged, may I ask whether there are other measures to solve the problem, look forward to your reply, thank you!