Inquiry About EDC Loss Setting, YAML Configuration for Training from Scratch

[Fly-beep-256]

Thank you for your excellent work on this repository. The implementation has been very insightful, and I truly appreciate the effort you’ve put into making this project available for the community.

However, I’ve encountered a few questions while working with the code, and I was hoping to get some clarification.

1. EDC Loss Configuration:
   In the YAML configuration file, I noticed that the use_edc_loss_rir is set to false. Could you kindly explain the reasoning behind disabling this loss function in the current setup? Is there a specific scenario where you recommend turning this on, or is it omitted intentionally for better performance?

2. Training from Scratch:
   If we intend to train the model entirely from scratch, are there any additional changes to the YAML configuration that we should be aware of? Specifically, would you suggest any adjustments to the optimizer, learning rates, or loss functions when starting fresh?

3. Epoch Discrepancy:
   I also noticed a discrepancy between the repository’s implementation and the paper regarding the number of training epochs. In the paper, 200 epochs are mentioned, but the current setup doesn’t seem to correspond to this. Could you clarify if this is an intentional change in the code, and whether there are any recommendations on the number of epochs for best results?

Could you also share your specific configuration and techniques for training?

Thank you in advance for your help!

[yiduohao]

Same questions as well! I also wanted to train the model from scratch, but found that the YAML configuration file was a bit confusing. It seems that the author intentionally disabled the discriminator and EDC loss from this current setup, leaving only the MSE loss. Did you have any insight? How did you train this model eventually? Did you try to enable the discriminator and EDC loss?

Appreciate any information! :)

[PengyvWANG]

Same questions. Have you solved it? Many configurations are not the same with that in paper.

[yiduohao]

This repo seems to use a different dataset, making the setting different from the paper. I am wondering if you could reproduce the metrics by training from scratch on this repo? The released model that we could download seems to work fine.