Not possible to run full dataset on google colab

[DanielMarchand]

Hi Everyone,

There is actually nothing wrong with the code here but since I spent so much time trying to get this to work I wanted to make a brief writeup of my experience with the hope that perhaps it will be of use to someone else in the future.

I had very much wanted to run all the luna subsets on a google colab. However, the disk is just a bit too small for the GPU partitions, even with the colab pro version.

I thought I could be clever and mount my google drive with its 2TB storage capacity but that won't work. Frustratingly gdrive seems to have a lot of strange caching behavior, so even if you work entirely within your drive your /root dir will fill up and your whole system will eventually crash. This is a known issue, googlecolab/colabtools#960 but I think the way this project is set up makes it particularly vulnerable.

I then thought I could use google cloud storage instead. This will not work either because there is a big performance hit for the kind of file reading/writing that you would want to do for this project.

So, end of the day, as far as I can tell it is not currently possible to run all the subsets using google colab. I suggest anyone reading this to just use subset0 and not spend as much time as I did chasing this whale haha.

Again, loving the book, and there is nothing wrong with the code. Hopefully google increases the disk allotment to their colab servers in the future so we can enjoy running all subsets.

[Do Truong Duy]

Thanks for your shared. I'm using Colab too. But it take me more than a day for training 1 epoch on chapter 11, even after I run LunaPrepCacheApp() (also took me 1 day to complete), while the author said that it took him about 15 minutes. Can you tell me how long it takes you to train? Thank you for your interest.