Get Hooked On Neural Net Inspection! That was a pun!
summary
Shows exactly what hooks reveal and how to use them with PyTorch including PyTorch forward and backward hooks, how hook callbacks receive raw tensors and how to log activations heatmaps and gradient concentration, a prime-number dimensional trick for decoding tensor shapes, a recursive children traversal to attach hooks across a large model, and a made-to-run-together parallel LSTM mirror that asserts output equality which demonstrates the approach actually works not just promises examples.
AI-written summary · about
related talks
41mReally Deep Neural Networks with PyTorch
20mLearn You a PyTorch!
01h 26mDeep Neural Networks with PyTorch
12mLightning Talk: Building Intermediate Logging for PyTorch
44mReversing Large Deep Learning AI Models
43mExploring Deep Learning Framework PyTorch
37mCode First, Math Later: Learning Neural Nets Through Implementation and Examples
27mUnderstanding Deep Neural Networks