Fastai installation on MAC OS 10.13 CPU only

(Peter) #1


I do my work mostly with Tensorflow and Python 3.6.

As Jeremy suggest that PyTorch is really good for researchers, I wanted to to try it out with the standard install as roposed on the github URL

I did the following steps

git clone
cd fastai
conda env update -f environment-cpu.yml

this last step gives me the error

Using Anaconda API:
Fetching package metadata …
Solving package specifications:
NoPackagesFoundError: Dependency missing in current osx-64 channels:

  • pytorch >=0.2.0 -> mkl >=2018

Can anybody help here. I thought that pytorch will be part of the installation
Thanks for helping


(Ramesh Sampath) #2

Yes, but because the pytorch installation for Linux and OSX are different, Jeremy might not have added it in the environment file. You could install it directly via the instructions here - models are written assuming you have a GPU and lots of RAM (32 or 16+ GB?) These are not pre-requisites to give a test run, but running DL on CPU could be very frustrating . You might want to spin up machine in Paperspace or use Google CoLab (that provides free 12 hour GPU). There are other threads that go into detail on running them in those environments.

(Peter) #3

Hi Ramesh,

thanks for your input. I will try this path of separate install of pytorch.

In the past, I was able to do pretty much any DL with my CPU only (though lots of RAM) and it was kind of slow, but still doable. I will give it a try, then maybe I go Paperspace or build my own box

thanks again


Hi Peter,

Were you able to solve the problem in MAC?

I am facing a similar problem while trying to install fastai in MAC (CPU only).

Kind regards and many thanks in advance

(Seth) #5

I was able to get the no-gpu/cpu only version working using the provided environment-cpu.yml file.

I have a 2017 Macbook Pro w/3.1 GHz i7, 16GB of RAM, and 500GB SSD. Running lesson 1 took about an hour on my machine - with no other processes running.

Not sure how other lessons will compare, but 1 hour for CPU only versus a few seconds for GPU (as shown in the lesson video) isn’t a good tradeoff.

If you value your time, I recommend using one of the recommended deep learning GPU capable providers.