shachoi/RobustNet

Why there are max_epoch and max_iter?

EDDChang opened this issue · 4 comments

EDDChang commented

As title, why you use max_iter in training loop but still have max_epoch in args?

shachoi commented

Hi,
args.max_epoch is not used. we use args.max_iter. code clean-up is required.
Thank you!

fingerk28 commented

Hi,
I have the same question. In the default setting, the max_iter=40000, which is mean that the network just need to be trained 38 epoch? (number of images is 12388 in GTA, equal to 1032 batch (batch size=12), 40000/1032 = 38.75)
Thank you!

shachoi commented

Hi,
As you know, the optimal epoch can be different depending on your batch size, network architecture, datasets, and so on.
Therefore, training on GTAV may require more iterations as the number of GTAV images is much higher than for Cityscapes.
But, I simply conducted the entire experiment with a fixed iteration of 40K.
Our study does not need to find the optimal epoch, compares our method with other methods under fair conditions.

It is required to find optimal values ​​for various parameters in order to compete with other methods for the best performance on the leaderboard, but our study has not covered this competition.

Thank you!
Sungha.

fingerk28 commented

I got it. Thanks for your help. : )