Attempting to get a high quality image from the TensorBoard projector my last hope lies in producing a svg from a screenshot. My attempts to create an svg in inkscape ended up relatively blurry and I am hoping to get some advice on the right settings or possibly another method. Size does not matter. I just want a sharp, scalable image.

[Picture updated]

enter image description here

  • 3
    Why dont you just iterate the dataset? And draw the circles yourself.
    – joojaa
    Jul 8, 2018 at 20:23
  • 1
    A "high quality" vector version of this image would make a rectangle for each pixel. Mission accomplished; particularly since "size does not matter". You are way better off working from the original dataset.
    – Jongware
    Jul 8, 2018 at 21:40

3 Answers 3


Multilevel neural net math packages are slightly beyond the capablities of most of us - otherwise I would run it and let it write the answer.

Seriously: Obviously you can create a set of xy coordinates from your data. Check the following extension of Inkscape: cartesianPlotData2D

It's available here: https://inkscape.org/sk/~fsmMLK/resources/

There are other tools, too.


Tensorboard projector seems to use Three.js which has an SVG rendering backend. If you are experienced with javascript and web programming you might be able to get SVG output by replacing the Three.js constructor THREE.WebGLRenderer() with THREE.SVGRenderer().


Usually getting better results from Inkscape's tracing function, I rarely use Vector Magic, but it's a handy item to have, although pricey.

I popped an image copy of your dots into the program and got a surprisingly good result, selecting the "photograph" option as the source.

It's available for your perusal for a limited time: Google Drive Link

Please advise when you have retrieved it, as I intend to remove it when it serves the purpose.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.