Annotated notes to Fast.ai's lesson on Image Classification

https://www.zerotosingularity.com/blog/fast-ai-part-1-course-1-annotated-notes/

89 13 8 years ago

I'm taking this course now. Your annotations explaining each command are very useful. Thanks for sharing them

kevinb15• 8 years ago

Here (https://medium.com/@hiromi_suenaga/deep-learning-2-part-1-le...) is another set of really detailed notes. There is one for every video and they are also very good.

zerosingularity6• 8 years ago

These notes are also really good, and ahead of what I have written up so far... :)

My take on these is adding somethings I picked up along the way to add another perspective and some additional info.

zerosingularity6• 8 years ago

Hope they help in understanding... Enjoy!

sudoaza• 8 years ago

A coworker has insisted on Fast.ai do you think its preferable over Keras/TFLearn?

zerosingularity6• 8 years ago

I found/find fast.ai to be incredibly useful for its practicality, good results, and top-down approach, however, it is sometimes hard to reproduce the results as well as clearly distilling what it is you can actually learn from each lesson. Writing this post blew my socks off as to what was taught in the video, yet it took me quite some time to get it all. So I hope the posts help people with that aspect of the course.

At the moment, I'm learning fast.ai/PyTorch in parallel with Keras/Tensorflow, so at this point, I have no definitive answer to your question which one is preferable. It will probably depend and they will most likely have their own benefits (I know that the boring answer, but I need to get more experience to give you a better answer).

As an exercise I'm trying to write the fast.ai notebooks in Keras, to see how they stack up. Might need to do a post on that as well.

I hope to answer your question better in the future. Could you tell me more about what you want to achieve, I might be of more assistance?

amelius• 8 years ago

build a world-class image classifier in three lines of Python

How is that learning how to build something, as opposed to simply using a system someone else built?

sp332• 8 years ago

Fast.ai's courses start at a high level with working code right off the bat. Then they break it down and show you how the parts work. It's much more motivating when you don't have to wait a long time to see any results. [Edit: and you know you're going to get something good instead of putting a lot of work in and getting something mediocre or plain not working.]

See this paragraph in the article: https://www.zerotosingularity.com/blog/fast-ai-part-1-course...

zerosingularity6• 8 years ago

I guess titles are meant to attract and provoke a little at the same time...

Fast.ai uses a top-down approach, allowing students to get awesome results with a minimal set of instructions, and start to dig deeper from there. If you look at the takeaways from the first lesson, you can see it's more than just three lines of Python and you're done, It's a solid and fresh approach to learn some actual practical deep learning.

Additionally, it kind of depends on what you want to do, we are all using things other people built, depending on the level of abstraction we care for or are willing to deal with...

needz2• 8 years ago

I'm working on an image classifier to identify pinball machines by their backglass art. I've been using Turicreate. I can't find any comparisons of the libraries (turicreate vs fast.ai). Does anyone have any resources that compare/contrast the different solutions available for image classification?

zerosingularity6• 8 years ago

I only learned about Turicreate after watching the CreateML video from WWDC. It's on my list to play around with, but no comparisons so far.

Any feedback on using it?

needz2• 8 years ago

I'm not experienced enough with ML to provide great feedback, but I did start getting segfault errors when I trained on too many images. I have multiple pictures of nearly 1000 different pinball machines, so right now it's a lot of trial/error as I determine how many pictures of each backglass I should train on and how diverse of angles/lighting produce the most accurate model. Right now I'm at about 15 different pinball machines, each with 20-30 pictures, and I get around 85% accuracy using my test set. I suspect/worry I won't be able to make a good model that can identify 1000 different pinball machines, but my patrons have asked for it so I'm giving it my all.

Annotated notes to Fast.ai's lesson on Image Classification

Related Stories