Learn You a PyTorch!
summary
Demo builds a small feedforward net and trains it while showing the concrete PyTorch workflows: a 28x28 → 32 → 10 network defined as nn.Module, move-to-GPU with .cuda() and simple multi-GPU via DataParallel, and conversion to/from NumPy with torch.from_numpy and .cpu().numpy. Data loading uses custom Dataset + DataLoader with multi-threaded transforms from torchvision transforms. Runtime benefits include a dynamic computation graph that lets one debug line-level errors, clip gradients, and hook Visdom for live visualization; API was still rapidly changing in 2017.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:29What this talk covers
- 1:43What this talk isn't covering
- 2:05PyTorch - Thoughts
- 4:41API - Network Architecture 32 Hidden
- 6:48Multi GPU Support
- 7:09First Class Python Support
- 8:24Custom Dataset - Intro
- 8:58Custom Dataset - Implementation (Barebones)
- 10:04Custom Dataset Loader - Using Transforms
- 11:59Dynamic Graph Computation 6
- 13:00Dynamic Graph Computation - Why?
- 14:08Gradient Manipulation during Runtime
- 14:22How can I visualize my training output/progress?
- 15:26Summary - Non leaky API
- 16:02Questions?
related talks
01h 26mDeep Neural Networks with PyTorch
01h 05mDeep Learning with PyTorch
01h 32mIntroduction to PyTorch
01h 27mHands-On Computer Vision with PyTorch
43mExploring Deep Learning Framework PyTorch
01h 31mPyTorch Workflow Mastery: A Guide to Track and Optimize Model Performance
30mAnyone can Build Great Deep Learning Applications - Deep Numpy
41mReally Deep Neural Networks with PyTorch