search

Res2Net / Res2Net-PretrainedModels

(ImageNet pretrained models) The official pytorch implemention of the TPAMI paper "Res2Net: A New Multi-scale Backbone Architecture"
https://mmcheng.net/res2net/
1.07k stars 215 forks source link

How to determine the width parameter #66

Closed drcege closed 2 years ago

drcege commented 2 years ago

May I ask how to determine the width parameter w?

For example in Table 3, Res2Net-50 (Preserved complexity) of settings 48wx2s, 26w×4s, and 14w×8s have the same FLOPs as ResNet-50 of 64w. How did you get the values 48, 26, and 14? Can you share the calculation formula?

gasvn commented 2 years ago

We adjuct the width of models with each scale to make the overall parameters and FLOPS constant with resnet-50. It's easy to keep the Flops as the overall structure is not changed.