Is it a quokka or a capybara?
On the recommendation of some old colleagues, I’ve decided to work my way through fastai’s course to learn some deep learning basics. I like it! Very programmer-oriented, which is exactly what I need, given my personal technical background. Fastai seems like a very cool library — a wrapper that makes training a model almost as easy as just learning a new API.
I didn’t do anything super crazy with the first “homework assignment” for the course — just modified the classifier to differentiate between quokkas and capybaras, instead of between birds and forests.
Really like the image downloader capabilities. This is a very satisfying array of training examples to look at:
Some thoughts on everything:
- Made a tiny modification to the example code to ensure that the 0th image wasn’t passed into the model to ensure that there was no way the image would end up in train and that I could always use the 0th result to validate. Probably there’s a cleaner way to do this, but I’m lazy!
- Apparently the probability that a (single) bird is a quokka is 0.9661. Hah! :)
- I am new to the concept of fine tuning, aka tweaking a model that’s already been pre-trained to serve a specific purpose, such as classifying quokkas vs capybaras. Am blown away by how few examples it takes. This is so wild.
- That see-through plant detector tablet thing that Sigourney Weaver held in Avatar made a very strong, futuristic impression on me over a decade ago. The see-through aspect might not be here quite yet, but it’s amazing that the plant detector is more than possible today.
- Same goes for the shoe brand detector that Leonard/Sheldon/Penny were building in The Big Bang Theory. What kind of crazy “futuristic” tech are TV shows and movies going to turn to, now that what was once the future is already here??
- Kaggle has come a long way since the last time I’ve checked it out. Am super impressed.
Speaking of Kaggle, here’s a link to my very first public notebook.
That’s it! On to lesson 2!
(this post is also on substack)