The Sapiosexual Way

The Journey
2 min readDec 7, 2022

Hey Hey! Let me start by saying that doing research is one thing & writing it is another.🤯

You must be wondering how the hell we train machine-learning models!? Well honestly, even I didn’t have a clue.😅

In the past few weeks, I developed & pulled the open-source codebase for EfficientNet-B4 & B7. Using python, Tensorflow & sagemaker.

Training machine learning models wasn’t an easy task & it was time-consuming as I trained EfficientNet-B4 & B7 on deepfake datasets. Such as CelebDF (Li et al.,2019), deepfake detection challenge (DFDC; Facebook, 2020), and FaceForensics++ (Rössler et al., 2019).

I created the unseen deepfake dataset by reserving 30% of videos from CelebDF (Li et al.,2019) and the DFDC (Facebook, 2020) alongside FaceForensics++ (Rössler et al., 2019) benchmark dataset of 1000 images.

Let’s do the math of data points in the Unseen benchmark gives us 🤪

1,692 Videos from CelebDF (Li et al.,2019) + 30,000 Videos from DFDC (Facebook, 2020) and + 1,000 Images from FaceForensics++ (Rössler et al., 2019) = 32,692 Validation datapoints. 🤩

Models get trained by 100 Epochs & for more than 30,000 sample data points it took a while to get the hang of it. This means we trained the models in batches of 100 data points (100 frames of video). As I progressed in training the models, I got some of the models not picking up the learning from datasets as it is too good to be a deepfake. So I did start training the EfficientNet-B4 & B7 from scratch to make sure I was training them correctly. Phew…… 😌

But the best part: it dumped results in binary: real 0 and deepfake 1 🤭 🤩

I am trying to keep it light, but the reality was completely different. It was a huge effort to understand the EfficientNets & to figure out how I could train them, test them & validate the results from a technical standpoint. Even running the first model in Sagemaker & having a python + Tensorflow platform ready to go was a great learning curve.

For the next few weeks, I have a challenge at hand. It is to write all of this in a more intellectually & profound way so my fellow researchers can understand it (the Sapio part of the brain to be satisfied was a legit goal).

See you in a few weeks with the bang of Deepfake detection results of EfficientNet-B4 and B7 ….🫡 🤓

also, don’t forget to touch base on proud effort posts!!! by Jasmin Bharadiya

--

--

The Journey
The Journey

Written by The Journey

We welcome you to a new world of AI in the simplest way possible. Enjoy light-hearted and bite-sized AI articles.