Not match the pretrained ckpt

[hyysam]

I use VCTK-DEMAND to train the model (120 epoch)，but when I use the 120 epoch ckpt to inference the VCTK-DEMAND noisy data，the denoising effect is not as good as the pretrained ckpt. Did you use extract dataset to train the model?

[hyysam]

Can you show me your training loss tensorboard?

[ruizhecao96]

Can you show me your training loss tensorboard?

Sorry I didn't record the loss on tensorboard, but we have use the early stop method since it gets overfitting when epoch number is too large, usually 80 epochs would be enough.

[hyysam]

I find that the --loss_weights", default=[0.1, 0.9, 0.2, 0.05] is not same as the paper default, and let loss_gan be smaller than the other loss, finally the GAN is not work

[hyysam]

Do you use extract dataset to train the pretrained ckpt?

[ruizhecao96]

Since the GAN loss is much large than other losses, we add the weights to scale them to the same level. We didn't use any other dataset to train the model.

[hyysam]

Thanks your answer