How to use AWS for your Deep Learning models

You just developed your first Deep Learning model and you want to train it, but you don’t have or can’t afford a GPU at the moment. You have heard of cloud computers being used to train complex models on gigantic datasets. You think, “Maybe, I can train my model on one of these cloud computers”. You search on Google and Youtube for tutorials and the sheer amount of knowledge and information regarding this topic hits you in the face like a tonne of bricks.

If this sounds familiar, well then!

I am going to try to clear the air of mystery a bit and try to make this process simpler. I started working with Deep Learning this summer and it became apparent to me that my system wouldn’t be able to do much training. This became more apparent when I had to use a GPU instance for a CNN(Convolutional Neural Network) I built to predict if a patient had pneumonia by looking at their chest X-ray scans. My Jupyter Notebook detailing my work can be found here:

When I built this model, to begin with, I used Kaggle’s own GPU to train it. The model was training on 5216 images and that can take a while. Kaggle’s GPU took a little over 30 minutes to train the model. Now if I wanted to make even the slightest change to it, I would have to spend another 30 minutes to retrain this model and then check. I thought to use one of the cloud computing providers to see if they could do it any faster. I decided to start with AWS. AWS is very expansive in the services it offers, so I had to spend some time looking for what would work for me. If you are training a simple(or even) a bit complicated Deep Learning model, you would probably want to go with their EC2 service. Amazon lists EC2 as servers in the cloud. Every server you launch for EC2 is called an instance.

What is EC2?

Related posts:

Most will recognize the quote above as the opening of President Franklin Roosevelt’s description of the attack on Pearl Harbor. In fact, this word gets misused a lot. Just this morning I heard a…

It is with great excitement that I am announcing our project, Fog, as well as our whitepaper for all to see. FOG is an ERC20-compliant token used in a system that implements zk-SNARKs encryption…

Acabo de entrar a una realidad que nadie puede ver esta en lo más oscuro de nuestro recuerdo, detrás de la mente sumergida en todo lo que hacemos diariamente por esas distracciones que llamamos vida…