Use Chinese dataset to train expert lip-sync discriminator

[Amber-Believe]

I use Chinese data set to train expert lip-sync discriminator, and the train loss remains at 0.69. Do you have such a situation?How should this situation be resolved

[ghost]

did you filter your data?

[Amber-Believe]

您过滤了您的数据了吗？

What do we mean by filtering? So far I have processed the data to fps 25 audio with a sampling rate of 16khz

[Amber-Believe]

did you filter your data?

The current data is in Chinese, including 50 different people and 7700 videos; The English data contains 40 people and 5,500 videos. Is it due to the small amount of data or other reasons. Our data has not been processed by syncnet_python. We found that the model syncnet_v2. model is not available. Can you provide it? That's the problem right now

[einsqing]

did you filter your data?

The current data is in Chinese, including 50 different people and 7700 videos; The English data contains 40 people and 5,500 videos. Is it due to the small amount of data or other reasons. Our data has not been processed by syncnet_python. We found that the model syncnet_v2. model is not available. Can you provide it? That's the problem right now

syncnet_v2 模型是英文模型，中文不适用，你需要训练一个中文的

[ghost]

how's about your config? lr, bs?

[Amber-Believe]

how's about your config? lr, bs?

All Settings are set by default, such as lr=1e-3 when training expert lip-sync discriminator
hparams.txt

[ghost]

1e-3 is too large, you can choose 1e-4 or 1e-5

[Amber-Believe]

1e-3 is too large, you can choose 1e-4 or 1e-5
Okay, I'll try

[Amber-Believe]

1e-3 is too large, you can choose 1e-4 or 1e-5

Thank you very much for your advice. After the lr is adjusted to 1e-5, the loss begins to decrease. What is the appropriate learning rate for wav2lip training? 1e-4?

[ghost]

1e-4 is good

[Amber-Believe]

Thank you!

[Nyquist0]

Hi @Amber-Believe.
May I ask what dataset are you using? Is it CMLR? LRS-1000? Or a private one?

[ChengsongLu]

Hi @Amber-Believe. May I ask what dataset are you using? Is it CMLR? LRS-1000? Or a private one?

@Amber-Believe
BTW, is the eval loss below 0.3 after changing lr from 1e-3 to 1e-5? And is there anything else you have done to achieve that?

[Nyquist0]

Hi @primepake
I was re-directed to this page from #97

But I still did not figure out my question.
The dataset I am using is LRS2. Because the official wav2lip algorithm use that for training. So I am assuming that should be filtered. And I also randomly checked some audios and videos of the dataset. The wav files are in 16khz sample rate and video files are in 25 fps.

And I would like to ask when do you think the syncnet training would be converged... Will it still be stuck on 0.69 for a long time? (110k steps for me currently..)

Look for your reply. Thanks.

[ghost]

again, how's about your lr? bs? num of gpus?

[Nyquist0]

LR is 1e-4, and BS is 64, 1 RTX A6000 GPU.

[Nyquist0]

@primepake Greetings!
I am trying to following the pipeline you proposed here

And may I ask how could you pre-process the video data? Are you using the preprocess code from the official wav2lip code?

[MarwanAj]

did you filter your data?

The current data is in Chinese, including 50 different people and 7700 videos; The English data contains 40 people and 5,500 videos. Is it due to the small amount of data or other reasons. Our data has not been processed by syncnet_python. We found that the model syncnet_v2. model is not available. Can you provide it? That's the problem right now

how much is the average length per video of the videos of the datasets ?