Experimental Machine Learning with Holoviz and PyTorch in Jupeyterlab
summary
Live demo of an experimental workflow for multiclass road segmentation from SpaceNet tiles using HoloViz dynamic dashboards in JupyterLab and PyTorch FCN (VGG16) with LR range test. Notebook and Binder repo load satellite images and masks (about 1,537 samples), show interactive mid‑training predictions, use transforms/augmentations, run a learning‑rate range finder (max LR ≈ 0.02) and cyclic scheduler, and report train/validation metrics and test visualizations — practical code that runs, with some technical hiccups.
AI-written summary · about
chapters (24)
- 0:00Introduction to experimental machine learning
- 1:57Road segmentation from satellite
- 5:32Today's agenda
- 6:55Why Holoviz for ML
- 11:59Holoviz core principles
- 16:26Setting up the GitHub repository
- 19:43Import Holoviz and setup
- 21:12Points and curves with Holoviz
- 27:44Ice cream consumption data
- 30:38Holoviz images
- 33:08Combining with Holoviz containers
- 36:28Dynamic Holoviz visuals
- 38:243D Holoviz visualization
- 41:07PyTorch workflow intro
- 46:31Custom PyTorch datasets
- 52:56Data augmentation transforms
- 55:25PyTorch data loaders
- 56:10PyTorch model design
- 59:32Fully convolutional networks
- 1:02:08Learning rate tuning
- 1:04:36Learning rate range test
- 1:07:26Training with optimal LR
- 1:09:50Model performance evaluation
- 1:10:55Full road detection workflow
related talks
01h 29mComputer Vision: An (Un?)Expected Journey, with Keras and Tensorflow
28mMachine Learning Live
17mF8 Refresh | Deep Learning with PyTorch
29mMachine Learning with Deep Neural Networks in the Browser
01h 11mTutorial: Beyond Neural Search: Hands-on Building Cross-Modal/Multi-Moda...
01h 27mHands-On Computer Vision with PyTorch
49mMachine Learning 2.0 With Hugging Face
01h 10mIntroduction to Machine Learning with H2O and Python