search

Bartzi / see

Code for the AAAI 2018 publication "SEE: Towards Semi-Supervised End-to-End Scene Text Recognition"
GNU General Public License v3.0
573 stars 147 forks source link

About the train_fsns.py #84

Open feitiandemiaomi opened 4 years ago

feitiandemiaomi commented 4 years ago

@Bartzi Thanks for your good job When I try to train the fsns according to readme, I find the loss are unusual, which shows as follows: image ps: my train dataset is the original fsns test ,val dataset is the original val I do not change code ,so it trains in default ,and I have two gusses,one is the my configuration of the fsns dataset has error ,the final gt shows ： image Is it right? And the other guess is that the lr is too big, as the number of iterations increases, the lr becomes larger . I will go on training. Looking forward to your reply

Bartzi commented 4 years ago

Hmm looks like your starting learning rate is too high. My guess is you tried to train with a learning rate of 1e-3 you should set it to 1e-4. It is totally normal that the printed learning rate increases this is because chainer calculates some crazy things for the Adam optimizer.

feitiandemiaomi commented 4 years ago

@Bartzi I will try, Thanks a lot!