2019 Version of Fastai

For people who were not experts in 2018 or 2017

The current version of the Medium iOS app does not allow for code blocks. I will not be buying a desktop computer to write on this platform, so my code blocks will be presented as quotes.

Looks like Jeremy Howard launched the 2019 version of fastai software yesterday. The update releases the courses-v3 covering fastai version 1.x to the world. Since I was halfway though lesson 4 of the 2018 version of “deep learning for coders,” I decided to switch to the 2019 version and start over using the latest version of fastai since the 1 year old library is already behind the times (this deep learning stuff moves fast)!

For those who have a GPU machine already setup, making the switch is easy. All I did was clone the github repo for the course-v3 into a folder under the fastai repo. This will allow me to use the course-v3 repo as a sub directory in the same setup I was already using. To do this, navigate to the fastai repo (cd fastai) and type the following:

Remember to do a git pull on both the fastai repo and this course-v3 repo.

After rewatching lesson one and following along in the notebook, two things became evident: this course is now for the mainstream, and deep learning has come a long way in just a couple of years. Fastai is quickly becoming the leader in deep learning tools, training, and community. People who completed fastai have built really cool (and profitable) tools in a very short amount of time (7–14 weeks). The course now has a focus on not only ease of use for deep learning and a simple yet powerful library, but also on production. Fast.ai was updated with guidelines on production for the trained models and Jeremy now covers production during the lessons. Websites like render.com offer cloud hosting for trained models/classifiers for $5 per month. This means anyone who comes up with a good classifier can quickly upload it to Render and point people to a domain to use it. Right now Render is still invitation only, but fastai students get an invite code through the course. I will put a link to Jeremy’s teddy bear v. grizzly bear classifier on Render — very cool stuff. This new technology will open up deep learning to the masses and allow people to deploy their models publicly with little front end work. Fast.ai is continuing to prove that open source and ease of use are the future.

The second realization that is evident from lesson one is how big of a leap deep learning has taken in the past year. Instead of starting out with building an image classifier that can recognize dogs and cats, lesson one now teaches you to train a model that can recognize which of 37 breeds of dog or cat the picture in the image is from. Last year (or in my case 6 weeks ago), the accuracy on the dog or cat classifier was 98%. This iteration of fastai can recognize the exact breed with 96% accuracy with less than 2 minutes of training. Amazing.

Related posts:

Any modern website built consist of complex layouts. For better content representation one needs to use column-based layouts. This can be done in multiple ways, it can be done through “inline-block”…

Take some time off and go someplace new. Do it alone. Solitude will help bring your thoughts in order and focus on issues that really concern you and only you. Being gone from the home city will…

Pada pembuatan fungsi-fungsi ini, sudah diusahakan penamaan variable atau fungsi jelas dan tidak ada line komentar yang tidak perlu dan clean code. Jika ingin melakukan implementasi yang sudah ada…