Do you have any interst to support Lora for fast adaption to small datasets from big COCO？

[LYMDLUT]

Lora ，the most adavenced method for fast adapation，with extremely low parameter and training epochs，using the pretrained COCO ckpt，achieving compareable performance of fintuing or even better.
http://arxiv.org/abs/2106.09685

[hhaAndroid]

@LYMDLUT This is an interesting topic. Does it also work in the visual domain and especially in YOLO algorithms?

[LYMDLUT]

https://github.com/LYMDLUT/Lora_with_ResBnLn/blob/main/lora1.py
I'll provide a simple example, and maybe we can make some fair comparisons based on it.
model.requires_grad_(False)
lora_layers = warp_with_lora(model, 20)
lora_layers.requires_grad_(True)