search

InhwanBae / ENet-SAD_Pytorch

Pytorch implementation of "Learning Lightweight Lane Detection CNNs by Self Attention Distillation (ICCV 2019)"
MIT License
76 stars 18 forks source link

Confused about the training time #12

Closed byeciao closed 2 years ago

byeciao commented 3 years ago

image Is it normal that it takes me 30min to train an epoch on 1080ti? image and the test time is slow too.

byeciao commented 3 years ago

why the test result is so bad after training tusimple for 60k iteration? 1630574014(1) 1630573986(1) 1630574004(1)

InhwanBae commented 2 years ago

Hi @byeciao Thanks for your interest in my implementation.

In my case, It took 10 minutes per epoch with 8* 2080ti GPU. Also, it seems to be overfitted on the training set with the respect to your val_loss_seg graph. The training result of TuSimple (also CuLane) is somewhat affected by the batch size. 12~16 is suitable.

Note that the result for the TuSimple dataset was reproduced in previous version. rewritten code also works exactly the same, but check it out!