After 3 good weeks of trying and trying, i finally have a local machine running Fastai, tested for lesson 1.
Benchmark versus Paperspace is good, the first 3 epoch run from lesson 1 runs in 14 sec on my machine, 15 seconds on Paperspace:
arch=resnet34
data = ImageClassifierData.from_paths(PATH, tfms=tfms_from_model(arch, sz))
learn = ConvLearner.pretrained(arch, data, precompute=True)
learn.fit(0.01, 3)
My PC config is here: Titan XP, AMD - Threadripper 1950X 3.4GHz 16-Core Processor, M.2 Samsung Pro SSD
https://pcpartpicker.com/user/hlalibe/saved/#view=TYZsJx
I installed WIndows 10, then created a partition on my M.2 SSD drive following these settings:
https://www.google.ch/search?q=how+to+partition+drive+for+ubuntu&oq=how+to+partition+drive+for+&aqs=chrome.1.69i57j0l3.13963j0j4&client=ms-android-google&sourceid=chrome-mobile&ie=UTF-8#kpvalbx=1
Then i created a Ubuntu 16.04 installation USB stick using https://unetbootin.github.io/
setup Ubuntu with your preferences,
open Terminal in Ubuntu and follow this script to install Fastai like it is done on Paperspace:
http://files.fast.ai/setup/paperspace
You need to install Git first before running the script
https://git-scm.com/book/en/v2/Getting-Started-Installing-Git
when all is done, open terminal, cd fastai, then type jupyter notebook
Voilà, i hope these 3 weeks i’ve lost will help you get started faster !