python3 train_syncnet_sam.py

[einsqing]

Step 3355 | out_of_sync_distance: 0.50673163 | Loss: 0.56647605 | Elapsed: 0.77767 Step 3356 | out_of_sync_distance: 0.50098807 | Loss: 0.56791657 | Elapsed: 0.76504 Step 3357 | out_of_sync_distance: 0.50698453 | Loss: 0.56805406 | Elapsed: 0.72830 Step 3358 | out_of_sync_distance: 0.53348744 | Loss: 0.56974050 | Elapsed: 0.77868 Step 3359 | out_of_sync_distance: 0.48923862 | Loss: 0.56817545 | Elapsed: 0.75280

The expert discriminator's eval loss should go down to ~0.25 and the Wav2Lip eval sync loss should go down to ~0.2 to get good results.

Does the first step of loss also need to be trained to 0.25？Step 2 to 0.2？

[shahidmuneer]

I have reached a loss of ~ 0.29 it will go further down to 0.20, and then I will start training wav2lip_sam. So you should wait till the loss goes down to 0.2. Thanks

[fzuqxt]

@einsqing 你好请问下您这个口型判别器是怎么跑起来的，我卡在0epoch动不了

[fzuqxt]

我加载无反应

[zhuxiu1234]

我加载也无反应，有解决吗？

我加载无反应

[ashishjamarkattel]

may be its due to batch size being large enough or your dataloader is not sending data. Mostly its problem of dataloader.

[fqhwwlzzbh]

我加载也无反应，有解决吗？

我加载无反应

源代码里面有很多坑，你首先要看一下你的程序有没有读取到文件，没有读取到文件的话，会一直循环下去，作者给报错隐藏了，所以你也看不见报错。

[eminem171333491]

我加载无反应 我查出来是librosa版本高了 librosa==0.7.0，numba==0.48就行了

[jibingyangsf]

请问这套384 有在不改网络结构和损失函数情况下 收敛的很好的吗