Dynamic Neural Networks with DyNet
summary
Demonstrates building production-ready models in DyNet: a per-example dynamic computation graph API with persistent parameters, embedding lookups, trainers and forward/backward updates. Implements a Deep Averaging classifier, a biLSTM tagger with character-LSTM for OOV tokens, and tree-structured recursive nets, all written as plain Python graph-building loops. Introduces auto-batching that automatically groups independent graphs for large speedups on CPU/GPU, and reports that batching and DyNet’s low-level C++ engine yield practical, fast training without a separate graph-compile step.
AI-written summary · about
chapters (29)
- 0:00Introduction
- 2:13Neural Networks Toolkit Landscape
- 4:35Scholar Binding
- 5:18Graphs
- 6:26Dynamic vs Static
- 8:16Code
- 10:25Parameters
- 12:23Training Loop
- 13:57Example
- 14:37Coding
- 17:10What Embedding
- 17:49Similar Vectors
- 18:06Feature Vector
- 19:49Efficient Text Classification
- 19:59Deep Ranged Network
- 23:13Objects
- 24:21Summary
- 24:43Other names
- 25:05Model
- 27:20Graph
- 28:25Biased entangle
- 32:39Built tagging graph
- 34:17Recap
- 34:35Batching
- 35:49concatenate
- 36:36Auto Batching
- 37:12Single Graph
- 38:02Tree Shape Network
- 40:57Conclusions
related talks
49mDynamic Neural Networks using PyTorch
20mLearn You a PyTorch!
41mReally Deep Neural Networks with PyTorch
01h 26mDeep Neural Networks with PyTorch
37mCode First, Math Later: Learning Neural Nets Through Implementation and Examples
37mMonadic Deep Learning
37mWhat's new in Deep Learning?
30mAnyone can Build Great Deep Learning Applications - Deep Numpy