Training¶

Training produces calibrated window-level posterior models for the later LBSA and TRL stages. The goal is stable class evidence across time, not a standalone window-accuracy result.

Protocol¶

Only the 80-record training split is used to fit model weights and select checkpoints during training. The 20-record development/calibration split supports temporal-policy calibration and diagnostics. The 37 external-test recordings remain untouched until final scoring.

All retained models share the same architecture and training recipe. They differ only in window length and random seed.

Setting	Value
Windows	3 s, 5 s, 8 s
Step	1 s
Seeds per scale	42, 123, 456
Batch size	256
Optimizer	AdamW
Learning rate	\(1\times10^{-3}\)
Schedule	cosine annealing
Label smoothing	0.1

Model¶

The 1.41M-parameter classifier combines three 1D convolution branches with a BiLSTM. Kernels 3, 7, and 15 capture motion at different receptive fields; the BiLSTM adds order-sensitive context across the window. The output has six classes: background plus five sports.

The backbone is deliberately compact. Its role is to provide reliable posterior trajectories so the later analysis can isolate scale arbitration and record construction.

Objective¶

The training objective combines cross-entropy, focal, and triplet terms:

\[ \mathcal{L} = \mathcal{L}_{\mathrm{CE}} + \lambda_{\mathrm{focal}}\mathcal{L}_{\mathrm{focal}} + \lambda_{\mathrm{triplet}}\mathcal{L}_{\mathrm{triplet}}, \qquad \lambda_{\mathrm{focal}}=0.2, \quad \lambda_{\mathrm{triplet}}=0.1. \tag{1} \]

cross-entropy learns the six-class decision;
focal loss emphasizes harder examples;
triplet loss improves embedding separation;
class-balanced sampling limits dominance by frequent classes.

Augmentation combines amplitude scaling, Gaussian noise, random shift, time warp, and Mixup. Label smoothing and Mixup also reduce over-confident posterior spikes near ambiguous boundaries.

Retained models¶

Nine models are trained: three seeds at each scale. The table reports training-phase internal-validation F1; it is not external-test evidence.

Scale	Seed	Best epoch	Internal-val F1
3 s	42	53	0.81
3 s	123	37	0.80
3 s	456	73	0.81
5 s	42	37	0.81
5 s	123	38	0.86
5 s	456	32	0.81
8 s	42	31	0.85
8 s	123	38	0.85
8 s	456	27	0.85

The final stack keeps one selected model per scale: seed 42 at 3 s and seed 123 at 5 s and 8 s. It does not average seeds within a scale. Cross-scale fusion is performed later by LBSA.

Interpretation¶

Longer windows show a stronger mean internal-validation trend, but they blur transitions. Shorter windows localize changes better but are noisier. Training therefore supplies complementary posterior sources; it does not decide the final record policy.

Reporting boundary

Internal-validation results select assets. They must not be mixed with the fixed external-test segment scores reported on the Results page.

Reproduce

The default training entry point is python train.py. Environment setup, authorized data placement, and saved-asset checks are documented in Quickstart and Assets.