Poor Training Results Compared to Provided Pretrained Model

[whu-wkn]

Hi there,

I recently cloned the repository and tried to train the model following the instructions. However, I'm getting significantly worse results compared to the provided pretrained model250000.pt.

Here are some details about what I did:

• I used the list file (list.txt) provided by @sakmalh.
• I used the same training configuration as described in the README.

Despite all this, my trained model performs much worse than the released pretrained model. I suspect there might be something missing in the training process or data preprocessing.

Could anyone help me with:

1. What might cause such a performance gap?
2. Are there any specific training tricks or hyperparameters that need to be adjusted?
3. How can I verify if my training pipeline is correct?

Any help or suggestions would be greatly appreciated!

Thanks in advance!