why my loss is negative and 0.000000

[fuyunfuyun666]

why my loss is negative and 0.000000,please help child!!@MichalBusta
like this:

[MichalBusta]

Dice loss is negative and bounded ( = it is ok )
zero loss ->

E2E-MLT/train.py

Line 471 in 8dd13e4

if step > 100000: #this is just extra augumentation step ... in early stage just slows down training

Hope it helps daddy.

[MichalBusta]

and read again readme - sec training

[fuyunfuyun666]

Dice loss is negative and bounded ( = it is ok )
zero loss ->

E2E-MLT/train.py

Line 471 in 8dd13e4

if step > 100000: #this is just extra augumentation step ... in early stage just slows down training
Hope it helps daddy.

what its meaning?please say clearly, thanks!

[MichalBusta]

my guess is that your data is not in expected format - try to run it with -debug=1

• you should see what is wrong ...

[kei6]

I've faced with the same issue in seg_loss, ctc_loss and reg:

Follow your above advice, I've run with -debug=1 and find some images, but this below example is confusing me:

Why does it just a part of my training sample?
The loss is still zero and negative after 2000 step. Is it right or what I need to do to fix this?

Thanks a lot, sir!

[MichalBusta]

On Fri, 17 May 2019 02:08 Kei_Cin, ***@***.***> wrote: I've faced with the same issue in seg_loss, ctc_loss and reg: [image: Screenshot from 2019-05-17 16-01-48] <https://user-images.githubusercontent.com/45931733/57916530-32ec2680-78bd-11e9-9e3d-a4723d533f03.png> Follow your above advice, I've run with -debug=1 and find some images, but this below example is confusing me: [image: Screenshot from 2019-05-17 16-05-38] <https://user-images.githubusercontent.com/45931733/57916791-c6255c00-78bd-11e9-9cbf-21b772b2639c.png> Why does it just a part of my training sample? The loss is still zero and negative after 2000 step. Is it right or what I need to do to fix this?

You have to fix this 1. input size should be at least 512 - otherwise a lot of words will be partially out of image = do not care see training parameters in Readme

…

Thanks a lot, sir! — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#29?email_source=notifications&email_token=AA7KHMAGXTNGX7JYQDGYNDLPVZY2HA5CNFSM4HIEMRQKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODVUGUNI#issuecomment-493382197>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA7KHMAXPP7DIWZMCMSKT4DPVZY2HANCNFSM4HIEMRQA> .