# 2022/08/23 Meeting ## 測量Alphapose Flops 利用 fvcore 計算 FLOP,param | model | FLOPs | param | | ------------- | ----- | ----- | | fastpose(our) | 23.78G | 40.57M | | fastpose 4096 | 60.27G | 134M | | fastpose 3x3 | 31.03G | 78.33M | | fastpose 5x5 | 26.87G | 112M | ## fastpose: ![](https://i.imgur.com/sdriozS.png) ## fastpose4096 ![](https://i.imgur.com/Ck0xEwS.png) ## fastpose 3x3 Conv Layer ![](https://i.imgur.com/qSvfYjR.png) ## fastpose 5x5 Conv Layer ![](https://i.imgur.com/wsu8nju.png) # 改良空間 別人做的alphapose參數 (輸入大小為320x256): ![](https://i.imgur.com/grOdntH.png) *FLOPs/2 is the number of FLOPs divided by two to be similar to the number of MACs. https://github.com/osmr/imgclsmob/blob/master/tensorflow2/README.md 有人在alphapose上做過類似的teacher student model ![](https://i.imgur.com/gd8gU9O.png) ![](https://i.imgur.com/N5qt8uP.png) ![](https://i.imgur.com/fTvb5Si.png =500x) https://arxiv.org/pdf/2008.10060.pdf https://github.com/HaoyiZhu/Multi-person-Full-Body-Pose-Estimation