Machine Learning Project: Friendnet | Custom CNN, ResNet50, CLIP & Ensemble Demo.
In this video, I walk through FriendNet, a deep-learning facial recognition project I built during school to identify and classify my friends from images.
I cover the complete machine-learning pipeline—from collecting videos and extracting facial frames to preprocessing the dataset, training multiple models, evaluating their performance, and deploying the final system through a FastAPI-powered application.
What you’ll learn
How convolutional neural networks extract features from images
How I built a custom 2D CNN using PyTorch
How transfer learning works with ResNet50 and OpenAI’s CLIP
Why pretrained layers can be frozen during training
Dataset splitting, normalization, and image augmentation
Cross-entropy loss, stochastic gradient descent, and early stopping
How Class Activation Maps help explain model predictions
How ensemble predictions combine several models for more reliable results
A live demonstration using previously unseen images
Model performance
Custom FriendNet CNN: approximately 95% test accuracy
ResNet50: approximately 100% test accuracy
CLIP: approximately 98% test accuracy
Chapters
00:00 Introduction
02:02 Building FriendNet from scratch
02:31 Understanding convolutional neural networks
07:13 Transfer learning with ResNet50 and CLIP
09:13 Freezing layers and replacing classification heads
10:50 Potential applications of facial recognition
11:56 Dataset preparation and splitting
13:01 Image normalization and augmentation
14:16 Custom CNN architecture
16:19 Training process and hyperparameters
19:09 Early stopping and overfitting
20:28 Model performance results
21:44 Model explainability with Class Activation Maps
24:02 Application deployment
25:01 Live model and ensemble demonstration
27:05 Conclusion
This project was built with Python, PyTorch, ResNet50, OpenAI CLIP, FastAPI, computer vision, transfer learning, and ensemble learning.
Presentation link: https://friendly.chrismba.com/present