吕文玉

You can use `torch.profiler ` https://pytorch.org/tutorials/recipes/recipes/profiler_recipe.html

1. RTDETR-R18 has 60 FLOPs 2. only need once forward pass

见主页文档 @20231211

Not yet. And welcome to submit relevant PR

I think you should choose `score threshold` carefully.

You can reference this issue/discussion https://github.com/lyuwenyu/RT-DETR/issues/13

没太理解具体什么需求 --- 或许这些地方是你需要的 1. https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_pytorch/src/solver/solver.py#L96 2. https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_pytorch/src/nn/backbone/presnet.py#L142 3. https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_pytorch/configs/rtdetr/include/optimizer.yml#L18

see discussion: https://github.com/lyuwenyu/RT-DETR/issues/95 and some useful code piece: https://github.com/lyuwenyu/RT-DETR/tree/main/benchmark

你论文在做什么方向啊；反正实时目标检测这快一般不需要；都是用

确实有点奇怪；可以把用来回复一下；另外可以给他列一下的计算复杂度；（比如模型里如果有Conv、Attention一类的） --- 复杂度的数据和分析方法可以从其他论文里找一下找（搜关键字：complexity） - Attention Is All You Need - Deformable DETR: Deformable Transformers for End-to-End Object Detection 例如：