Training loss value is very large about ivus-segmentation-icsm2018

[Kulbear]

Hi if i remember correctly, I didnt get the mean loss so that's the total loss over all pixels * batch size. to get the normal loss you can divide the number by pixel number in the image * batch size.

…

On Sat., Jun. 22, 2019, 20:43 gzYou, ***@***.***> wrote: Hi， I got your code and ran it. But the training loss value is always very large and it keeps 190 000 levels. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3?email_source=notifications&email_token=ADRSL3DZWF55WBC2JZJCFQTP33PL7A5CNFSM4H2X6N7KYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4G3D32WA>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADRSL3ASVMMC2QCMNAWLJBLP33PL7ANCNFSM4H2X6N7A> .

[gzYou]

Thx, I'll try it.

[gzYou]

Sorry, another question. I found there are three labels file in the Test_Set. They are "LABELS_obs1","LABELS_obs2_v1" and "LABELS_obs2_v2". So what's the meaning of the file name and which is the true label?

[Kulbear]

@gzYou There are multiple (4, I think) annotators and they may annotate the frame multiple times.

The ground truth can be the average across all the annotators.