Hello Reader,

Welcome to another edition of the AIFEE newsletter!

AIFEE stands for Artificial Intelligence For Everyone and Everything!

Another week, another Machine Learning topic!

How to personalize image generation models?

We’ve all seen the incredible results lately for image generation using deep learning.

Models such as DALL-E, DALL-E 2, Imagen and Stable Diffusion all show tremendous potential of understanding users text input.

I’ve also seen these models being used to modify images but there was something lacking. Personalization.

Imagine that you have some specific subject in an image and you want to generate new images using contexts around that subject.

For example, imagine that you have a specific t-shirt image and you want to see how it would look if worn by somebody else. These scenarios are hard to achieve using pure image generation models mentioned above.

Introducing DreamBooth.

This is an approach that can help with personalization. It takes a few images (3 to 5) of your personal subject and it also takes an image-generation model as input.

The output is :

• A unique identifier for your subject, let’s call it V.
• A fine-tuned/“personalized” text-to-image model that encodes V.

This is shown in the figure below ↓

Pytorch tip!

Sometimes you might need to extract only a part of a deep learning model.
​
For example, you might want to extract only convolutional layers of a pretrained model such as InceptionV3 or ResNet50.
​
How would you achieve that using Pytorch?
​
Below is a code snippet showcasing exactly how ↓

​

So you've trained, evaluated and deployed your ML model. Are you done?

The answer is no! Training, evaluation and deployment of your ML models is not enough.

There are actually more tests that need to continuously be run to ensure the validity of your AI system.

These tests need to be done for both your models as well as your data.

Some of these tests are :

• Continuously testing your model performance.
• Continuously testing data integrity.
• Continuously testing data distribution mismatches.

Implementing all of these tests by yourself could be time consuming.

Introducing Deepchecks.

It’s a Python open source library that you can use to run all of the tests mentioned above as well as other types of tests.

Currently you can use it for both computer vision tasks as well as tabular data tasks.

That's it for this week's newsletter! See you next time!

​

​