Problem about TEs

[sky-creater]

Hello, Thank you very much for your work.

I would like to ask whether TEs need to be trained separately first. If so, how many epochs would be required?

Thank you!

[bstandaert]

Hello,

If you start from random initialization, it’s generally preferable to first train only the TEs. This has been observed to improve overall performance. You can use the same configuration, but simply remove the GAFFE by replacing it with the identity module.
On the other hand, if you start from pretrained weights, it’s not necessary to train the TEs separately first, you can directly train with both TE + GAFFE.

Hope this helps! Let me know if you have any further questions.