Add --samples_per_plugin to specify explicit sampling counts

[ghost]

Add TensorBoard flag samples_per_plugin that allows for custom specification of samples to keep per summary type.

Previously, TensorBoard always downsampled summaries for OOM reasons but some users wanted the ability to keep all their summaries of some types.

With this flag, --samples_per_plugin='histogram=2,images=0'
keeps TensorBoard defaults for other summaries, restricts the number of histograms to 2, and keeps all image summaries.

[linzhiqiu]

If using tensorboard-pytorch, how can one upgrade the tensorboard to 1.9.0 in order to use this feature?

[nfelt]

@linzhiqiu We have no affiliation with tensorboard-pytorch, but as far as I can tell it looks like tensorboard-pytorch primarily offers a PyTorch-compatible way to write summary data to event files, which you'd then view with regular TensorBoard. So however you're getting TensorBoard now, you would just update it. If you're getting TensorBoard as part of a TensorFlow installation via pip then you would just pip install tensorboard==1.9.0.

[jperl]

Thank you so much for this feature. I was about to manually export all the images, this is 100x better. 😍

[ghost]

Aw, thank you for telling me that! :)

…

On Wed, Jul 25, 2018 at 4:41 PM Jon Perl ***@***.***> wrote: Thank you so much for this feature. I was about to manually export all the images, this is 100x better. 😍 — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1138 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/Ab4bt0beKouQ9ys7GzMdwjtY0vFNWLKGks5uKQJFgaJpZM4TXmlH> .