Travis tests sometimes don't pass due to travis slowness

[jcarreira]

It seems sometimes the travis tests fail because not enough time is given for the test to converge.

Should we increase the threshold? @andrewmzhang @yangryan0

[andrewmzhang]

I think that's a good quick fix.

To fix it more completely maybe we should sum up the total number of updates and compare that to the loss. Although this would require investigating some sort of updates to loss relationships.

[jcarreira]

That's too complex. I suggest we relax the threshold.

[yangryan0]

Yes I think increasing the threshold should work. I added that small change to the softmax model PR I added this weekend, since I didn't think I needed to create another PR for such a small change.

…

On Tue, Jul 31, 2018, 9:51 AM João Carreira ***@***.***> wrote: That's too complex. I suggest we relax the threshold. — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#71 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/Abwst2XSknNeVpVZ3DnaFVD38zjYGBhVks5uMIsGgaJpZM4VlOby> .

[jcarreira]

Ok, that sounds good to me.

[jcarreira]

Can we close this?