Confusion of test result?(about test validation set split)
Liudzz opened this issue · comments
Firstly, thanks for sharing so many models on cifar100 which is very friendly to beginner ! I have learned CNN for 2 years and your project can always inspire me !
I'd like you to ask you a question. In your implement train code , do you use test set to detemine your best model? If so, all the results got from best train model ? I have learned that datasets alway need to be divide into ,"train/test/val",3 folds.
And what also confused me is the accs on Imagenet .Do they use val set to detemine best model and then test ?
In my mind , you use test fold to test after train and that's your best acc . I also use this thought in my project and learning.
I've searched on Internet and still have't find out . It would be great if you can response me ! : )
Bty , In line 171 of your train.py code , you use time as file name , if we train start from a certain epoch , we can't get a continuous train line . In my project , I use runs as a default name . And it seems that I can get a continuous line ?(Maybe I saved checkpoint and trained from a certrain epoch?)
I've noticed this project was been updated 5 month ago , and I'd like to thank you again !