Replace the stale "centred-skeleton stub returning confidence=0, wiring is a separate PR" sentence with a description of what now ships: - `InferenceEngine::with_weights` loads `pose_v1.safetensors` through Candle and runs the documented Conv1d + Linear forward pass - output shape is 17 (x, y) keypoint pairs with the published confidence = 0.185 (validation PCK@50) - when the weights file is missing on disk, the engine logs a `tracing::warn!` and falls back to the centred-skeleton stub so the runtime contract is preserved The 3% PCK@20 / 18.5% PCK@50 accuracy table and the "Honest reading" section are unchanged — wiring the weights does not change the trained model, only stops the cog from emitting placeholder values. |
||
|---|---|---|
| .. | ||
| cog | ||
| src | ||
| tests | ||
| Cargo.toml | ||