Fine-tuning questions

[scarbain]

Hi,

Thank you for the release, this is a great project, congrats!

I have some questions about training:

• The run_1024_finetune.sh uses the file finetune_infilling.py but it isn't in the repo, is that okay to simply replace this with finetune_accessory.py ?
• What minimum VRAM will be needed to start training at resolution of 1024 ?
• Do you think it would be easy to implement validation samples during training ?

Thank you very much :)

[RuoyiDu]

1. Yes, thanks for your feedback!
2. The Lumina-Accessory training was conducted on 8*A100 GPUs. We haven't yet tested the model's resource requirements on GPUs with smaller VRAM or fewer GPU units. We expect GPUs with >40G VRAM should provide reasonably smooth fine-tuning.
3. This shouldn't be particularly challenging -- just like training any other model.

[Feynman1999]

1. Yes, thanks for your feedback!
2. The Lumina-Accessory training was conducted on 8*A100 GPUs. We haven't yet tested the model's resource requirements on GPUs with smaller VRAM or fewer GPU units. We expect GPUs with >40G VRAM should provide reasonably smooth fine-tuning.
3. This shouldn't be particularly challenging -- just like training any other model.

Hello, I have a few questions about the training:

1. Are all the tasks mentioned in the README completed in a single training run?
2. Approximately how many iterations were trained, and what was the batch size?
3. I noticed that the finetune_accessory.py code does not include the restoration task. Is this because the released code is not the final version?