categories of pretrained weight problem

[jeonjj1]

I am planning to use a pretrained weight file to fine-tune my custom dataset with transfer learning. My dataset has 5 categories, whereas I anticipate that the rt-detr weight file is designed for 80 categories. Therefore, I need to modify the weight file to accommodate only 5 categories, but I am not sure how to open the weight file. I would appreciate your assistance with this.

Additionally, I am curious to know if it is necessary to modify the weight file due to the difference in the number of categories.

[lyuwenyu]

You don't need to do this yourself. By default, mismatched weights will be discarded.

https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_pytorch/src/solver/solver.py#L142

[jeonjj1]

Thank you. However, I am curious if not modifying the pretrained weights will result in longer computation times.

Also, since there are only 5 categories, I wonder if I need to change all instances of num_classes = 80 to 5 in the code, and modify the category items as well. (I want to minimize computation as much as possible.)

[lyuwenyu]

Also, since there are only 5 categories, I wonder if I need to change all instances of num_classes = 80 to 5 in the code, and modify the category items as well. (I want to minimize computation as much as possible.)

Yes, you need to do so

[jeonjj1]

Thank you for your response.
I am wondering if there is a way to further reduce the model size. I am looking for methods that can significantly improve the speed, even if it means a slight decrease in accuracy.

[lyuwenyu]

You can try to use quantization technology. （ float32 -> float16 and float32 -> int8

---

ps. rtdetrv2 introduces a discrete deformable attention method to speedup inference. #179

[Caohz678]

Hello, I’d like to ask where exactly the code for the CCFF module is located. I haven’t been able to find it. Thank you very much!