tensorflow/tensorboard

Isolate point

bonseong11 opened this issue · 4 comments

The actual label and image look different. When the Isolate point button is pressed.
스크린샷 2024-05-17 14-14-24

As I mention in #6849, this seems specifically an issue with your model and/or data set, do you have a reason to believe this is a bug in the tool?

I'm afraid we don't have bandwidth to help troubleshoot issues other than bugs and features in the tool itself. I would encourage you to try other forums like stackoverflow.

If you think this is a bug in the tool we would need a way to reproduce (ideally a minimal example), so we can investigate.

I checked the dataset and proceeded, but when I bring up all the data, the label values and images match, but when I press the button, the label and sprite images are not imported properly from the label and sprite image importers, is it because our dataset is in a different order?

I'm afraid I'm a bit unfamiliar with the projector plugin. The people who originally worked on it have left the team, so we're going to need some help to troubleshoot this.

When you mention the "label and sprite image importers", are you referring to a specific piece of code? Is that code from TB? Or code that you own?

AFAIU, from the examples I can see at the standalone projector demo (which should be almost the same implementation, AFAIU), the list on the right should be "close by" points, so the "2" is not the image that corresponds to the selected "9" label (if that's what you meant), but rather a data point that is considered "close" to it.

However, it does seem like the "2" image is associated with a "3" label, from what I can appreciate from the screenshot. I'm not very sure how that happens. Again, I would imagine this is more related to your particular data set.

Would you be able to share a reproducible example? and/or a dataset that we can use to reproduce? Otherwise, I can't really do much to help you from that screenshot.

I was able to reproduce by running the standalone version with the example data sets, as described in https://github.com/tensorflow/tensorboard/blob/ae7d0b9250f5986dd0f0c238fcaf3c8d7f4312ca/tensorboard/plugins/projector/README.md

As mentioned, the "2" shown is not meant to correspond to the "9" label, but I do think the images should match the numbers shown on the right. In the screenshot above, it should show a "3", I believe.

In the screenshot below, I believe it should show images with the number 6.

image

Unfortunately, I don't know when we could prioritize looking into this. Is this affecting your work in a particularly disruptive way @bonseong11 ?