Hello everybody! For individuals who have no idea me but, my title is Francois, I’m a Analysis Scientist at Meta. I’ve a ardour for explaining superior AI ideas and making them extra accessible.
Right this moment, I’m excited to delve into one of the crucial important breakthroughs in Pc Imaginative and prescient post-Imaginative and prescient Transformers: Masked Autoencoders (MAE). This text serves as the sensible implementation companion to my earlier put up: The Ultimate Guide to Masked Autoencoders (MAE)
For the next tutorial, we’ll use the code accessible on this github repository:
Here’s a transient reminder of the way it works:
Right here’s how the methodology works:
1. The picture is cut up into patches.
2. A subset of those patches is randomly masked.
3. Solely the seen patches are fed into the encoder (that is essential).