Fastbook chapter 5 - run performance way worse than documented

I’m currently working through Fastbook.
What struck me, is that the performance when running the cells:

is way worse than the documented performance:

Is this normal and if not - why could this be the case?

I’m getting the same thing. My first error_rate from fine_tune() or fit_one_cycle() is ~0.3 and I can’t get it below ~0.19 no matter what I try.

python --version = 3.8.8
ipython --version = 7.21
fastai --version = 2.2.7

is this related to this thread?

1 Like

Yes, thank you.
I fixed the problem by using this (yet) unmerged pull request:

git clone
git fetch origin pull/3268/head:dataloaders_fix
git checkout dataloaders_fix
pip install -e “fastai[dev]”

Just for me to clarify.
Is this a problem due to some fastbook imports or is the issue a general problem within the current fastai version (e.g.: dataloaders train & valid transformations are generally inconsistent)?

1 Like

Hello, while I am getting the same accuracy, my runtimes are much longer (~40:00) than what’s shown in the lesson or here (~00:25). I am using Google Cloud Platform (GCP) AI notebook instance with a Nvidia Tesla K80 GPU, 4 CPUs & 26GB RAM. This is making my progress very slow.

Do someone have experience with GCP that can help me out?

Something else I find curious is, the GPU runs at 100%, even though I have just started the notebook or it’s idle.