A Practical Introduction to Convnets for Image Classification

on YouTube ↗51m

" summary

Practical pipeline that trains a convolutional network in Keras with a TensorFlow backend on a homemade rock paper scissors dataset of ~2,400 labeled images, resized to 32×32 resized images and split by acquisition session so training has ~600 images per class and testing ~170 per class. Model has ~56k parameters, shows severe overfitting without data augmentation, and drops test accuracy from ~75% to ~50% when augmentation is removed. Includes Jupyter demo, realtime webcam classifier and a separate drone trail follower built from 20k GoPro images.

AI-written summary · about

" chapters (6)

1. 0:00Start
2. 6:22Jupiter Notebook
3. 7:01Build the Dataset
4. 10:57Build a Dataset
5. 14:25Preparing the Training and Testing Data
6. 35:52Data Augmentation

" related talks

• 01h 33mIntroduction to Convolutional Neural Networks using TensorFlow and KerasOliver Zeigermann · PyData London 2017
• 44mIntro to ConvNetsKashif Rasul · PyData Berlin 2014
• 01h 29mComputer Vision: An (Un?)Expected Journey, with Keras and TensorflowRodolfo Bonnin · PyData London 2018
• 36mMachine Learning Zero to HeroLaurence Moroney, Karmel Allison · Google I/O 2019
• 01h 01mA practical guide to deep learningTess Ferrandez-Norlander · NDC London 2019
• 01h 01mGetting Started with Computer VisionSeth Juarez · NDC London 2021
• 01h 39mGoogle Colaboratory and Python: Deep Learning basics for CV tasksKristina Host · DSC Europe 24
• 01h 15mTutorial on Image Classification using Scikit-Image, Scikit-learn, and PyTorchMatt Litz · PyData Virginia 2025