brainstorming
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).
Explorer les Skills
Développement · SKILL.md
vector-quantization
vector-quantizationBase VectorQuantize and RandomProjectionQuantizer workflows for single-stage codebook quantization, including masks and lens, image and 3D feature maps, kmeans/cosine/codebook-dim tuning, top-k/manual EMA behavior, gradient tricks, and optional FVQ bridge guidance.
Independent GitHub snapshot · Automated classification · Original documentation governs
Documentation de la Skill
GitHubname: vector-quantization description: "Base VectorQuantize and RandomProjectionQuantizer workflows for single-stage codebook quantization, including masks and lens, image and 3D feature maps, kmeans/cosine/codebook-dim tuning, top-k/manual EMA behavior, gradient tricks, and optional FVQ bridge guidance." disable-model-invocation: true metadata: disco-role: operating license: MIT
Vector Quantization
Use this sub-skill for the single-stage, codebook-based APIs in vector_quantize_pytorch.
Covers
VectorQuantizeRandomProjectionQuantizerget_output_from_indicesupdate_ema_indicestopkmaskandlenshandlingaccept_image_fmapandaccept_3d_fmapkmeans_init,use_cosine_sim,codebook_dim- stale-code expiration
- orthogonal and codebook-diversity losses
- multi-headed VQ
rotation_trickanddirectional_reparamsync_codebook- optional FVQ bridge guidance
Route elsewhere
- Residual or grouped residual codebooks -> residual-quantizers
- FSQ or FSP -> scalar-quantizers
- LFQ, ResidualLFQ, GroupedResidualLFQ, LatentQuantize, BinaryMapper, EvoLFQ -> lookup-free-and-latent
- SimVQ, ResidualSimVQ, HierarchicalVQ -> sim-and-hierarchical
Start here
Fast read
If you only need a safe check, run the smoke helper on CPU. It covers a tiny VectorQuantize forward/reconstruction check and a tiny RandomProjectionQuantizer round trip.