Lesson 1 in Fast.ai Deep Learning Course
In the previous post, I tried to demonstrate “getting started” part of fast.ai. Lesson 1 is also started with explanation of installation. Getting notebook and auxiliary files are crucial to implement. The last versions of these files are in github (utils, vgg16, vgg16bn).
Structured data is needed for training. There are 12500 dog photos and 12500 cat photos in training folder. 2000 photos will be moved to valid set from total 25000 photos. Also to gain fast insights, creating sample folder which includes 20–50 photos for each class(cat,dog) will be useful.
My p2 usage request is replied by AWS, and so I am able to use it. First, I tried to use spot instance because of price. While in-demand instances cost $0.90 / hr, spot instances could be quarter of this price cost according to your bid. However there are some drawbacks of this method. When somebody offered to give more than your bid, your instance is shutting down in two minutes. And you can not save your works because of EBS. One way to solve this issue is detachment of disk before the termination, however it needs to stopping of instance. It terminates instantly, I couldn’t detach it. Some functions of AMI is not useful in spot, because it is prepared for on-demand. I didn’t want to loose more time, so I returned to on-demand although it is expensive.
Let’s turn back terminal. Aws-alias.sh makes easy to connect AWS on terminal. Just writing aws-start and then login with aws-ssh. After login, don’t be suprised when you started jupyter notebook. It is encrypted with “dl_course”. Everything is ok.
You will learn to use Kaggle CLI with these page. With Kaggle CLI, downloading data and submitting solution is possible.
After the lesson 1, there is homework section in notes. I had difficulty in getting results from model and create prediction file to submit Kaggle. After researching in forum, posts refer these file is placed in course github page.
This notebook helps me to create submission file. I compared my file structure and methods with this notebook. After I finalized my submission and started lesson 2, this file is mentioned in there. So, you should try at first your own and this is very important to learn. Otherwise you will just do replication of code.
As a result of homework, I submitted my work to Kaggle competition. It is a rough solution so I get this result:
Lesson 2 is related with more concepts and I will create article with in-depth knowledge.