search

RobvanGastel / dinov2-finetune

Testing adaptation of the DINOv2 encoder for vision tasks with Low-Rank Adaptation (LoRA)
MIT License
82 stars 9 forks source link

Register Support and Training Info #4

Closed Oguzhanercan closed 1 month ago

Oguzhanercan commented 2 months ago

Hi thanks for your work. I want to ask that if this repo supports the register added version of dinov2. Also I want to ask that I am planning to do a training with 112x112 images, do you have any experiments with low resolution images for DinoV2 training?

And last question, in your experiments, how much time does it take to finetune the model (On which gpu also) ?

Thanks.

RobvanGastel commented 1 month ago

Hi! I also used the register DINOv2 weights for my experiments, it automatically removes these tokens in the DINOv2 encoder model.

https://github.com/RobvanGastel/dinov2-finetune/blob/e1815206165520b359e502a69fafa9c864303c41/main.py#L188-L193

I do not have access to the computer I trained on at the moment. But iirc it was a 3000 series GPU and I trained with batch size 8, it took quite long to train for 100 epochs. It could have been about a day. I forgot to log these details.

I do not have any experiments with low resolution images but it will help with training speed.

Oguzhanercan commented 1 month ago

Thank you