Skip to content

Instantly share code, notes, and snippets.

@yashk2810

yashk2810/U-Net.ipynb Secret

Created May 28, 2017 15:03
Show Gist options
  • Star 1 You must be signed in to star a gist
  • Fork 1 You must be signed in to fork a gist
  • Save yashk2810/bd317f854794d2d6969240420bb1407d to your computer and use it in GitHub Desktop.
Save yashk2810/bd317f854794d2d6969240420bb1407d to your computer and use it in GitHub Desktop.
Download ZIP
U-net on Camvid
Display the source blob
Display the rendered blob
Raw
U-Net.ipynb
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
@authman
Copy link

authman commented May 28, 2017

A quick one-over and nothing specific pops out at me as wrong... looks to be a standard unet. I can only offer some ideas - maybe try adding BN into it (since you're comparing against tiramisu). Also, this paper might be of interest to you - https://arxiv.org/pdf/1511.00561.pdf it's a every so slightly modified unet such that the encoder path is VGG... which gives you the ability to use TL. They also used the same dataset so you'll be able to closer compare your results. Last thing is you're using sparse cross entropy loss, but aren't doing any class weighing. That will have some negative effect too... but from what I recall looking atyour 100-tiramisu notebook, you didn't have any in there either, so... maybe that's just a better model ?_?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment