train with custom dataset

[liuxiaoyu1104]

If I have custom camera poses (which may be labeled differently from those in your dataset), what should I consider if I want to train using your code? Specifically, should I recalculate the mean and variance for the data, and is it necessary to adjust the variance of the noise added during training and sampling?

[robincourant]

Hi,

If you plan to train the model on a different dataset, I highly recommend recomputing the mean and variance and storing them there - this should ensure things work correctly.
As for the diffusion hyperparameters, I wouldn't touch them, as the EDM framework handles that well.
However, you might need to tune some training hyperparameters, such as the learning rate.

Keep me in touch if you face any troubles!

[liuxiaoyu1104]

Thank you so much for your reply. I would like to ask, do you have any suggestions for tuning some training hyperparameters? My dataset consists of approximately 3K videos.

[robincourant]

Could you give me more details to help you with this?
Have you already tried launching a training? Do you have the loss curves? What are the current results?