Google Colab Setup for FastAI Part 2 v2

(Arnav) #22

Has anyone been able to successfully reproduce the results of Lesson 11 on Colab?
For me, the kernel restarts when I load the fasttext embeddings presumably because it ran out of memory. Any tips to fix this?

(Brian Muhia) #23

@naman-bhalla Thanks for this! I used this to run a practical session for a one-day conference I helped organize on 7th April.

(Cedric Chee) #24

My experience with Google Colab so far:

I started using Colab when it first came out during Beta stage. At that point, I have no confidence to continue using it for mildly serious work due to its stability, the show stopper issues encountered and zero ease-of-use design in mind.

  • connections: been quite good but not perfect. I think it really depends on your ISP?
  • file management:
    • Last I heard, the major issue is regarding file management as all data is lost after a session (VM instance) is destroyed.
  • GPU not available or fresh GPU with less than ~700 MB CUDA memory left.

Recently, I managed to reproduced and complete the training for these notebooks:

  • the beginning lessons (those smaller models and smaller datasets) of part1v2 within the 12 hours limit (or 1.5 idle hours).
  • artistic style transfer, super resolution, DCGAN and Wasserstein GAN with LSUN dataset (big 51 GB) of part2v1 with multiple (2 to 3) Colab sessions.

The solutions shared and discussed in this thread are good. However, these days I pretty much use the solution from Clouderizer to solve pretty much most of the problems and issues I have. Note: I am not affiliated with them. It’s the easiest & fastest way to get your notebook up & running on Colab (if you don’t mind giving your Google Drive access permission to Clouderizer). Clouderizer’s automatic 2 way data sync to Google Drive is nice despite it’s not real-time. So, to prevent training progress loss, I save the model state (checkpointing) more often. It’s a better solution but not perfect.

Go check out this thread that me and the rest have been discussing some tips & tricks & workaround when using Colab:

My early conclusion: Nevertheless, Colab is a great service for learners who can’t afford paid cloud computing services like AWS/GCP/Paperspace/etc.


What with the option: colabolatory to prototyping, preparing the model, following lesson, etc. Paperspace/AWS, whatever, to evaluate serious datasets?

(subhobrata) #26

Hi All,

I have tried most of the code in colab, it is having some or other error.
Anyone having working code in colab for part 2 full course.

Thanks in advance.

Thanks and Regards,