A conv_in module in model_gen was trained.

[WuKe2002]

Awesome work!
And the conv_in module in model_gen was set to train, why do this? Could you do some ablation experiments of this?
Looking for your experience share, thks!

[cswry]

Hi there!

We actually haven't done a specific ablation study on the conv_in module yet. We included it in the optimization to ensure the input projection could align better with the LoRA updates, but it might not be strictly necessary.

It would be awesome if you could run a quick experiment to compare the results with and without training conv_in. We'd love to see your findings and update the training recipe based on your experience!