Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Projector: default compute PCA issue #1204

Open
csarron opened this issue May 21, 2018 · 3 comments
Open

Projector: default compute PCA issue #1204

csarron opened this issue May 21, 2018 · 3 comments
Labels
plugin:projector stat:contributions welcome theme:performance Performance, scalability, large data sizes, slowness, etc. theme:ui-polish Features or fixes that make core UI more pleasant.

Comments

@csarron
Copy link

csarron commented May 21, 2018

When clicking on the projector tab, tensorboard starts loading tensor values and computes the PCA.
The problem is that if the tensor is very large (e.g. large embeddings), default computing the PCA will take super long time, and you can do nothing but wait a few minutes at least.

I believe this is flaw for the default logic, a better option would be to let the user decide whether to compute the PCA(T-SNE, etc), for example, a button saying visualize.

  • TensorBoard version: 1.8.0
  • OS Platform and version: Ubuntu 16.04
  • Python version: 3.5
@ashudeep
Copy link

This happens even if you the tensor is already Nx2 sized!

@nfelt
Copy link
Contributor

nfelt commented Oct 2, 2018

I agree that there should be some control when launching the projector about which type of analysis to do, and one of the options should be "no analysis" if the values are already 2 or 3 dimensional as described in #1192. (I think the custom panel can be used to achieve this for 2 dims today, but I'm not entirely sure). It could either be a splash screen that has you select the analysis to perform, or it could just choose one by default (like PCA today) but make it interruptible if that's not the one you want.

We may not have the bandwidth to work on this in the near future, but contributions are welcome.

@filipre
Copy link

filipre commented Oct 19, 2018

Could you share which dimensions your tensors have? It would be interesting to know what kind of tensors are too big for Tensorboard. My current one is 2810 x 128 which already breaks the whole board.

Edit: Ah, now I understand @ashudeep's comment :) Seems like 2810 is too big.

@nfelt nfelt added theme:performance Performance, scalability, large data sizes, slowness, etc. theme:ui-polish Features or fixes that make core UI more pleasant. labels Dec 17, 2019
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
plugin:projector stat:contributions welcome theme:performance Performance, scalability, large data sizes, slowness, etc. theme:ui-polish Features or fixes that make core UI more pleasant.
Projects
None yet
Development

No branches or pull requests

4 participants