search

NeuromatchAcademy / course-content-dl

NMA deep learning course
https://deeplearning.neuromatch.io/
Creative Commons Attribution 4.0 International
731 stars 265 forks source link

Torch downgrade removed from W3D3 (NeuromatchAcademy#905) #934

Closed colleenjg closed 8 months ago

colleenjg commented 8 months ago

This commit fixes issue #905. torch 2.0 introduced an out of memory bug in the tutorial, which was addressed with commit ed8f9f2 by downgrading to torch 1.11. The pytorch bug was fixed in torch 2.1 through commit fff4a9. Since Google Colab now comes with torch 2.1 installed by default, the downgrade is no longer needed.

I also changed the supervised learning performance reported in table from 98.67% to 98.70%. Unlike the other methods (VAE, SimCLR, etc.), the supervised result has never been perfectly identical on each run. It varies a little bit around 98.70% (from 98.67% to 98.73%, when I've run it). To avoid confusion for users, the first time this result is reported, it is now reported with a ~ (i.e., as ~98.70%) to signal to users that their result may differ slightly from it.

colleenjg commented 8 months ago

Tagging @spirosChv to let you know about this PR reversing the torch downgrade you implemented a few months ago to fix a new memory bug that showed up in W3D3. :)

TL;DR: torch on Google Colab has now been updated to 2.1. From my tests, this fixes the memory bug you found (as planned!), and so torch no longer needs to be downgraded.

spirosChv commented 7 months ago

Fixes this https://github.com/NeuromatchAcademy/course-content-dl/issues/905