training the encoder only

[cakirogluozan]

Hi everyone,

First of all, thank you for the incredible work on this project — I’m really enjoying experimenting with it! 💃

I have a question regarding training: I’m considering training the encoder (and possibly the decoder) on my own complex industrial dataset. Since I don’t have any flow-related labels, my goal is to adapt the encoder to better fit my domain.

Do you have any insights on this strategy?

Thanks in advance!

[aharley]

Thanks for the nice message.

Sounds like a good plan, but my first instinct might be to freeze the encoder and only train the decoder (whereas it sounds like you're describing the opposite).

I wonder if you can share a picture or video of your domain, to give a better sense of why finetuning is necessary.