search

cc-ai / kdb

AI can help Visualizing the Impacts of Climate Change. This is an open forum to share our work
https://mila.quebec/en/ai-society/visualizing-climate-change/
Apache License 2.0
19 stars 3 forks source link

Maintain reconstruction quality with increasing decoder depth #117

Open 51N84D opened 4 years ago

51N84D commented 4 years ago

As we add downsampling/upsampling layers, the size of the latent vector is reduced. This causes a reduction in reconstruction quality.

Deeper networks result in smoother flooding (less strict adherence to the masks) so there is a tradeoff here

51N84D commented 4 years ago

A few ideas:

sashavor commented 4 years ago

I would start with the first idea (should be easier), and if it doesn't help, go with the conditioning

On Mon, Apr 6, 2020 at 2:51 PM Sunand Raghupathi notifications@github.com wrote:

A few ideas:

  • Give the latent vector more channels
  • Condition the latter feature maps of the decoder with the original image

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/cc-ai/kdb/issues/117#issuecomment-609973289, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADMMIIRWMQZ4BHLQCOZTKS3RLIQCTANCNFSM4MCQK76Q .

-- Sasha Luccioni, PhD Director of Scientific Projects (AI for Humanity, Mila), Postdoctoral Researcher (UdeM) Directrice des projets scientifiques (IA pour l'humanité, Mila), Chercheure postdoctorale (UdeM) [image: Image result for universite de montreal logo]

vict0rsch commented 4 years ago

It's easier to start with but I have less hope as IMHO the most likely bottleneck is spatial resolution but you'll see as you compare!