Negative samples outside of batch
kudkudak opened this issue · comments
For some reason we generate negative samples within-batch. That's arguably weird, not sure what it does.
Try outside of batch generation, and see how it impacts models for both negative argsim and normal split.
It is really bad for small batch sizes - makes negative samples much harder actually.
That was a specific thing that was done in Li et al
…On Tue, Nov 28, 2017 at 8:49 PM Stanislaw Jastrzebski < ***@***.***> wrote:
It is really bad for small batch sizes - makes negative samples much
harder actually.
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#52 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ADO_QQ5kz7di4d9FowDRVVG3JsUv3vUbks5s7LgTgaJpZM4Qt3RN>
.
I know. That was a self-note :)
2017-11-29 15:09 GMT-05:00 Michael <notifications@github.com>:
… That was a specific thing that was done in Li et al
On Tue, Nov 28, 2017 at 8:49 PM Stanislaw Jastrzebski <
***@***.***> wrote:
> It is really bad for small batch sizes - makes negative samples much
> harder actually.
>
> —
> You are receiving this because you are subscribed to this thread.
> Reply to this email directly, view it on GitHub
> <https://github.com/kudkudak/common-sense-prediction/
issues/52#issuecomment-347725757>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/ADO_
QQ5kz7di4d9FowDRVVG3JsUv3vUbks5s7LgTgaJpZM4Qt3RN>
> .
>
—
You are receiving this because you were assigned.
Reply to this email directly, view it on GitHub
<#52 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ABf8aXWj-elmf-edq221A0P1ept0eKllks5s7boGgaJpZM4Qt3RN>
.
Not important ATM, closing