Inspired by @radek’s blog post here and his two forum posts (here, and here), I present a generic way to get up and running on AWS in less than 60 seconds (well, 60 seconds after you go through it the first time 
).
Prerequistes:
-
You must have created a persistent p2.xlarge instance in EC2 and give it a name. I simply followed the instructions from the introductory workshop material here and named my instance “fastai-part1v2”. For storage, I went with 50GB.
-
You must have the AWC CLI installed and configured with your AWS Access Key, Secret Access Key, and default region.
Steps:
-
Clone my repo here: https://github.com/ohmeow/aws_setup
-
cdinto the aws_setup directory -
In your terminal run
$ bash start-aws-instance.sh {name-of-instance | defaults to fastai-part1v2}
This will spin up your instance and ssh you into it. -
Once your ssh’d into your AWS instance, get and run my version of the install script (this step only needs to be ran once) :
$ wget https://raw.githubusercontent.com/ohmeow/aws_setup/master/fastai-install-gpu-part1-v2.sh
$ bash fastai-install-gpu-part1-v2.sh -
After that is done, start your jupyter notebook
$ ./start-jupyter-notebook
This will activate the fastai anaconda environment that is created for you in the fastai-install-gpu-part1-v2.sh script and launch jupyter notebook. -
In your browser, go to https://localhost:8888 and get to work
-
When you are done, open up another terminal on your machine and run:
$ bash stop-aws-instance.sh {name-of-instance | defaults to fastai-part1v2}
This will shut down your EC2 instance
Once you’ve gone through the steps above, you just have to …
$ bash start-aws-instance.sh {name-of-instance | defaults to fastai-part1v2}$ ./start-jupyter-notebook- Go to https://localhost:8888 and do your work
-
$ bash stop-aws-instance.sh {name-of-instance | defaults to fastai-part1v2}when done.
Notes:
-
You’ll notice that while relying heavily on @radek’s version of the install script in step 4, I did make a few changes. The biggest one is simply using the environment.yml file in the fastai repo to create a “fastai” environment and install all the required packages for the course from it.
-
This is for setting up a persistent EC2 instance. If you are into saving $$$ and want to see how to get things working with spot instances, see @radek’s posts mentioned above. I purposely decided to go with using persistent images (for the time-being at least) because it was much more straight-forward to get operational quickly. The scripts are more generic and will allow you to completely avoid having to log into AWS to be operational.
-
This is a work in progress so please feel free to post recommendations and submit pull-requests if you got ideas on how to improve.
-
I haven’t tested on anything outside of a p2.xlarge instance so if you are using other GPU friendly builds, let me know if the above still works or not.
