Open Modelslightweight skeleton action recognition · runs in your browserinternal map →
…
golf swings
…
move models
…
move labels
…
stance marks
…
total labels

Every model is a lightweight skeleton classifier that runs 100% in pure TypeScript in your browser (only pose estimation touches a GPU). Input everywhere is COCO-17 keypoints in a body-frame canonical form (hip-centered, torso-scaled, nose-side mirrored). Click how to use for a copy-paste forward pass, or inspect reasoning to see the model think.

Graph strike spotterbrowser · TS
Jab, cross, jab cross, low kick and high kick, with their time intervals
archShared skeleton graph network with per-frame strike heat and duration heads; the live registry below names the selected architecture and stream
inputOne person's COCO-17 posture sequence on a fixed 30 Hz clock, retaining visibility and missing frames
weightsreplay.move_classifier_model.params (kind='spot-graph'); checksum-bound ONNX graph and fixed validation-track IDs
runs inCPU ONNX Runtime in Node; single-thread WASM in the browser. No inference GPU required.
how it reasons
Joint-removal interventions measure how each joint changes the model's score at its original peak. These are measured effects, not semantic explanations or named hidden units.
LegibleTCN (move / strike)browser · TS
Earlier named-move architecture, retained for comparison and existing non-graph models
arch3-layer causal TCN + MIL attention pool · per-joint grouped conv (17 joints × x,y,dx,dy) → pointwise mix → depthwise causal conv, GroupNorm
inputVariable-length bag of 68-dim per-frame instances (34 body-frame COCO-17 x,y + 34 Δt velocities)
weightsreplay.move_classifier_model.params (JSONB, discriminated by kind='legible-tcn')
runs inPure TypeScript, in-browser + Node — app/lib/moveClassifier/legibleTcn.ts
how it reasons
Every hidden unit maps to a named joint-channel group, so each firing is legible. The attention head shows WHICH frames drove the decision; the per-unit templates read out in English.
Attention-MIL (move)browser · TS
Arbitrary named moves — one weakly-supervised model per move (which frames ARE the move)
archAttention Multiple-Instance Learning (WTAL/STPN-style) · per-frame encoder → attention a=σ(wa·h), score s=σ(c·h), pooled Σa·s/(1+Σa) · L1 + TV on attention
input68-dim per-frame instance bags (position + velocity), order-invariant
weightsreplay.move_classifier_model.params (kind='mil')
runs inPure TypeScript — app/lib/moveClassifier/milAttention.ts
how it reasons
Attention weights localize the move inside the clip without frame labels; the score head is a per-frame 'is this the move' probability, pooled by attention.
Order-aware TCN-MIL (combos)browser · TS
Sequence / combo moves where order matters (jab-cross, feint-then-kick)
archMIL-attention + depthwise dilated 1-D temporal conv (identity-init, ~25-frame receptive field), optional AND-of-halves pool
input68-dim per-frame instance bags
weightsreplay.move_classifier_model.params (kind='mil-tc')
runs inPure TypeScript — app/lib/moveClassifier/milTemporalConv.ts (forwardTC)
how it reasons
The temporal conv gives the model a short memory so the SAME two moves in a different order score differently — explainBagTC surfaces the ordered evidence.
Golf P-system (P1–P10)browser · TS
The ten swing keyframes — P1 address · P4 top · P7 impact · P10 finish
archNearest-centroid prototype per (facing, P) + monotonic DP that enforces P1<…<P10 ordering
inputStance window features (2N-1)×34 + 2-dim facing one-hot
weightsreplay.p_position_model.params (JSONB, singleton) + localStorage pSystem:model:v2
runs inPure TypeScript — app/lib/pSystem/{train,ordering}.ts
how it reasons
Per-frame centroid scores form a matrix; assignMonotonic picks the best strictly-increasing P1..P10 path through it, so detections can never go out of order.
Stancebrowser · TS
Fighting stance — orthodox vs southpaw (live switch alerts)
arch2-layer MLP (ReLU→ReLU→sigmoid), BCE, manual SGD
inputWindowed body-frame features (2N-1)×34 (hip-centered, torso-scaled, nose-mirrored)
weightsreplay.stance_classifier_model.params (singleton) + localStorage
runs inPure TypeScript — app/lib/stanceClassifier/runPipeline.ts
how it reasons
Reads left/right shoulder + hip + lead-foot geometry over a short window; the mirror-canonical features make it view-consistent.