Skorch - A Union of Scikit-learn and PyTorch
summary
A scikit‑learn–compatible wrapper called skorch actually hides the training loop and exposes fit / partial_fit / set_params so neural nets plug into pipelines and grid search while using PyTorch modules. Examples run MNIST and ants vs bees with a pretrained ResNet18, freezing layers, LR scheduling and checkpoints (best val accuracy saved at epoch 8). Segmentation uses a U‑Net with IOU scoring, 4 data‑loader workers and pin_memory, a cyclic one-cycle policy learning rate; code, slides and runnable examples are in the repo.
AI-written summary · about
chapters (13)
- 0:00Introduction to Scorch and its purpose
- 2:05Skorch API overview
- 3:08MNIST basics
- 6:27Hyperparameter tuning
- 8:16Custom metrics with callbacks
- 10:26Ants vs Bees demo
- 13:44Transfer learning frozen
- 14:53Checkpoints and saving
- 18:23Image segmentation
- 24:10Advanced callbacks LR
- 25:57GPU data loading
- 27:40Summary
- 29:06Q&A
related talks
01h 15mTutorial on Image Classification using Scikit-Image, Scikit-learn, and PyTorch
01h 31mPyTorch Workflow Mastery: A Guide to Track and Optimize Model Performance
28mEnhancing Machine Learning Workflows with skore
17mF8 Refresh | Deep Learning with PyTorch
01h 32mIntroduction to PyTorch
01h 13mExperimental Machine Learning with Holoviz and PyTorch in Jupeyterlab
43mStorch - GPU Accelerated Deep Learning for Scala 3
01h 26mDeep Neural Networks with PyTorch