affine transform issue (divided by zero)

[AtousaTorabi]

Hi,
Thanks for great work to do augmentation without server. I get divided error by zero when training at affine function in this line:
scale_size = config.transform_params.target_dist / (scale_self * self.scale)
seems scale might be zero sometimes.

Thank you!

[abinjoabraham]

Even I am facing the same issue now. I will post if I find a solution. @AtousaTorabi

[murrayLuke]

Interesting that could explain the NaN and inf values I've been seeing.

[abinjoabraham]

Still I am stuck in this situation .

[murrayLuke]

I don't think it would be too hard to write a one liner and change the scale size if it is 0.

[abinjoabraham]

I have wrote a simple if statement to skip the 0 of the scale size. But not sure what value I need to give to scale size if it is 0.

[murrayLuke]

Try 1 since that should give no scaling whatsoever.

…

On Wed, May 2, 2018 at 2:55 PM abinjoabraham ***@***.***> wrote: I have wrote a simple if statement to skip the 0 of the scale size. But not sure what value I need to give to scale size if it is 0. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#28 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ASesbArmcJ_GIrijCdmLBGqTFlUirp-Kks5tugE7gaJpZM4TcLPI> .

[abinjoabraham]

Tried the same and it is running. But the loss is going nan.

[anatolix]

neither scale_self nor self.scale shouldn't be zero. scale_self taken from hdf5, check dataset generation, may be we need to put assert there.

[abinjoabraham]

Issue is not anymore for me. I put an assert in the dataset generation file and also changed the first learning rate of the epochs.

[YicongHong]

@abinjoabraham I tried to add an if condition in the data generation file to avoid divide by zero:
if (img_anns[p]["bbox"][3]) == 0: pers["scale_provided"] = 1
but the loss is going nan, I am wondering how exactly should I modify the dataset generation file and change the first learning rate?

[abinjoabraham]

@YicongHong Exact lines are below

    if pers["scale_provided"]==0:      
        pers["scale_provided"] = 0.2

Also regarding the epoch learning rate, I reverted back to the original line of codes. As it was not making any issues.