search

dingfengshi / TriDet

[CVPR2023] Code for the paper, TriDet: Temporal Action Detection with Relative Boundary Modeling
MIT License
160 stars 13 forks source link

about GMACs #22

Open zuguang0425 opened 10 months ago

zuguang0425 commented 10 months ago

d6ccd2128f1db0d375919012a074eb0

我们在train utils.py 中 train_one_epoch函数中添加了测试GMACs代码，得到如下结果：

f1dfc92a18ef217e168d47325ffb6e8

请问论文中表6是如何计算得到的，希望得到你的帮助，非常感谢

5d9368b7977463fe1355f4538199cb8

dingfengshi commented 10 months ago

Hi, you inserted the GMAC evaluation code in the training loop. During training, the length of the temporal feature is not fixed and might be shorter than 2304. You can try inserting this part in the test part on THUMOS.

zuguang0425 commented 10 months ago

我们在train utils.py 中 valid_one_epoch函数中添加了测试GMACs代码，得到如下结果： 64cd327c5be6e485bee4d7fe6e2153a

6659caaa6bcb3efc7f40d81134b28ab 这样计算是否正确

dingfengshi commented 10 months ago

This result of MAC is too small. Can you check the input shape and the model setting, and try to evaluate each module inside the forward() function in meta_arch.py?