Setup help

Here’s how to get set up for the course, by platform:

General setup FAQ

  • Which should I use? We recommend Paperspace, since they’ve got everything customized and set up for this course. They have a good free option, for $8/month you get plenty of extra space, and it’s a totally standard Jupyter setup so everything should “just work”. Colab is also a good free option, especially if Paperspace free instances aren’t available, but it’s a bit more fiddly.

  • What about the Local options? For users with their own equipment, the Local* options can be used in place of the cloud based options. These setups are meant for non-beginners and for those who can quickly troubleshoot installations. Remember this course is about deep learning, not system configuration setups and compatibility issues.


I know the answer is no, but I have Nvidia 1660 Ti 4GB VRam Laptop. Will I be able to get started? Is this a good idea ?

@vijaysai I’d recommend avoiding it. Nothing has been tested on that. But if you’re an expert, then you might be able to get things working with plenty of debugging and fixing.

@jeremy re AWS/Sagemaker – I saw there were Cloudformation scripts from last year to start a Notebook instanstance in Sagemaker. Will there be seperate CF Stacks to launch this years jypyter notebooks ?

No there hasn’t been any folks from AWS helping with the course this year.

So should we avoid using AWS this year – or is it not very difficult to launch the notebooks on Sagemaker without AWS help ?

I guess you could try it and see - if it’s not easy, then stick to something else that’s listed here.

I understand opinions may vary on this, but for those willing to use Docker, I’m sharing the container I will be using throughout the course:

Let me know if you need help or face issues. At the moment I would recommend the 0.0.11 version/tag.


Thanks @jeremy I can try the last year’s CF Stack script. Couple questions:

  1. They used ml.p2.xlarge as the default instance. It has an Nvidia TeslaK80 GPU with 12GB memory. Will that be good enough for this year’s course ?
  2. They used conda instead of pip in the CF script to install the fastfai library like so conda install -y fastai -c fastai. For this year’s course, can we use conda install -y fastai2 -c fastai2 ?
  3. They cloned to get all the notebook’s for last years course. Should we just clone for the notebook’s for this year’s course ?
  4. They also did: pip install jupyter_contrib_nbextensions. Other than this, do we need any extensions, other libraries / dependencies ?
1 Like

If you are interested I have set up cuda/cudnn based images of all many of the jupyterhub images, which are very convenient setups, and I have added a pytorch one as well ( docker pull if you want to use it) they can be used as easy bases for fastai docker images


It is, but it’s expensive for what it is. There are much better value options nowadays.

Thanks! Sounds nice.

I have the setup automated for a while, and I have just added support for fastai v2 installs…

In looking at the docker site for your container, you recommend installing cuda 10.2 first. Is this necessary as pytorch brings in the cuda stuff it needs? I am looking to run this container within UNRAID and would rather not install cuda on the system if I don’t have to.

My hope is that your container will recognize both of my GPUs. As of now, I can run fastaiv2 within an Ubuntu VM with GPU passthrough for one card only. I would like to have both in the oft chance that some of the libraries can parallelize.

1 Like

I just checked the Nvidia-docker documentation and

The machine running the CUDA container only requires the NVIDIA driver, the CUDA toolkit doesn't have to be installed.


  • you don’t need to install CUDA (for some reason I have been doing this for quite some time now)
  • I have updated the documentation.

Hello I’ve had some issues while doing the setup it seems there seems to be something going on with the cloning of the repository in Paperspace Gradient. I’ve got some issues when cloning the repo the fix was to instead of using the ssh command to clone use https to clone it.

perfect, I will give it a whirl a few hours before class starts. need to finish my WFH day, first. :grin:

1 Like

You know where to find me if you hit any issues. You will also need to clone the repo for the course:

git clone

All the best!


I have a Ubuntu 18.04 server with a 2080ti that i’ve used to navigate the v3 course with a v1 conda env with no problems.

Could this be as simple as cloning that environment and running pip install fastai2?

1 Like

That might just work. :slight_smile:

To complete the first two notebooks, you’ll also have to install these:

RUN pip install nbdev
RUN pip install graphviz
RUN pip install azure
RUN pip install azure-cognitiveservices-vision-computervision
RUN pip install azure-cognitiveservices-search-websearch
RUN pip install azure-cognitiveservices-search-imagesearch
RUN pip install "ipywidgets>=7.5.1"
RUN pip install sentencepiece
RUN pip install scikit_learn