The Stream of Consciousness
Wassup yo! ready for juicy findings on EfficientNets? (yay yay). I have to tell you guys that EfficientNets has been very efficient in the past week. I was stoked ofc, but the literature reading part keeps me sanely insane.
You know how that goes. 🤭
I dived deep & found that there are a total of OG 8 versions of it, starting from B0 to B7. So, basically, EfficientNets are a family of machine learning models based on a new scaling method that uniformly scales all dimensions of depth/width/resolution using a compound coefficient. Great Idea, right!?
As you might be wondering how that looks? Don’t worry, I got you covered. See the below EfficientNets Architecture Diagram:
So, once you have a baseline model in hand, next you will need to scale your baseline model with a compound coefficient for all three dimensions. Now you have your first EfficientNet-Call It Awesome-1.0.
lol “If you wanna be a scientist about it” then here you go: EfficientNets code
Don’t judge me, I was working with the deadline in my head so, I took pre-trained EfficientNets & formed my draft research questions straight away. It’s a profound first step to a dissertation where you can narrow down the problem into mere simple questions. This is very smart of me to create research questions, Honestly! I wasn’t expecting it until many weeks into the dissertation.
I decided to focus on two EfficientNets B4 & B7. That’s it. Why? Bcuz, they have a state-of-the-art accuracy of 84.3% on ImageNet. I want to know if the two best models compete with each other using the latest deepfake datasets, then how well they will perform. (just curious 😺)
Bottomline:
Changes in training datasets will have a high impact on the performance of models.
Phew.. so much of EfficientNets & such a tiny focus window of my brain.🙃
My plan for next week is to form research questions in terms of Quantitative research; something that I can measure & evaluate. I know i know… Uncharted waters again. 🥲
Catch up….on..what got me (Jasmin Bharadiya) here!! The Flyleaf & The Deep Dive
