What it does

Fast-wavenet is a reference implementation of a generation-time optimization for WaveNet and similar causal dilated convolution networks. Instead of recomputing the full binary tree of convolutions for every output sample, it caches “recurrent states” per layer and reuses them via pop/push queues. The result: inference complexity drops from exponential to linear in the number of layers.

The interesting bit

The authors confirmed with DeepMind that the original WaveNet team was already doing something similar—just not writing it down. The repo’s real value is making the trick explicit and visual: binary-tree diagrams, queue mechanics, and a worked example you can adapt to any streaming causal convnet.

Key highlights

• Complexity: O(2^L) → O(L) for generation with L layers
• Works for any causal dilated CNN, not just audio (streaming classification/regression too)
• Includes timing experiments showing crossover point where caching wins over naive GPU convolution
• Accompanied by arXiv paper with full derivation
• ~1,700 stars, TensorFlow implementation

Caveats

• For small L, naive GPU convolution can still be faster due to parallelism; the win only appears at scale
• README notes this is “mostly” the algorithm description plus example code, not a full training pipeline

Verdict

Grab this if you’re shipping WaveNet-style models to production or hacking on streaming convnets. Skip it if you just want pretrained weights and a generate.py script—this is the boring-but-critical plumbing, not the demo.