FatimahEmadEldin/PG-XFormer-artifacts

PG-XFormer — reproducibility artifacts

Code, split definitions, per-clip predictions, and feedback-audit outputs for the paper "PG-XFormer: Pose-Guided Cross-Modal Exercise Monitoring with Agentic Feedback". Every number in the paper can be recomputed from the files here with code/r7_aggregate.py.

Related repositories (see the PG-XFormer collection):

• Model weights and model card: FatimahEmadEldin/PG-XFormer
• Synthetic training data (InfiniteRep archive): FatimahEmadEldin/infiniterep-physiotherapy

All PG-XFormer resources are grouped in the PG-XFormer collection.

Contents

Path	What it is
splits/infiniterep_{train,dev,test}.csv	The instance-disjoint, class-stratified split used in the paper: 560 / 70 / 70 clips (56 / 7 / 7 per exercise, ten exercises). Keys exercise, instance_id match the InfiniteRep archive.
predictions/synthetic/<run>.npz	Held-out test predictions for every synthetic run: labels, preds, probs (70 × 10 softmax), macro_f1, accuracy, kappa, ece, train_seconds.
predictions/synthetic/<run>_history.csv	Per-epoch training loss / accuracy and development-set accuracy and macro-F1.
predictions/kimore/kimore_<setting>.json	KIMORE five-fold results per setting: per-fold metrics, test labels, predictions, and probabilities. Recording indices follow the order of the KIMORE exercise tables (ex1–ex5).
agent/reports_natural.jsonl	Kinematic error reports of the 70 held-out synthetic clips (64 clean, 4 deviating, 2 critical).
agent/reports_stress.jsonl	Controlled stress set: one injected deviation of known identity per report (60 deviating, 10 critical).
agent/<model>_{cpu,gpu,openrouter}.json, agent/qwen235b.json	Every generated cue with its score, for raw and safety-gated pipelines (Qwen3-235B, Qwen3-1.7B, Qwen3-0.6B).
agent/qwen235b_original_audit_cues.jsonl	Cues of the original 235B audit (natural reports).
results/revision_numbers.json	All aggregated numbers reported in the paper (means, standard deviations, bootstrap CIs, McNemar / t / Wilcoxon tests).
results/cost.json	Parameters, GFLOPs, and batch-1 latency per configuration.
study2/PREREGISTRATION_study2.md	Pre-registration of the label-efficiency and movement-quality study, written before any run (an earlier timestamped copy is in the authors' working repository).
study2/le_<arm>_k<budget>.json	Label-efficiency study: per fold and draw, training-set size, metrics, test labels and predictions (k0 = all labels). Arms: syn_ft, scratch, skel_syn, skel_scratch, probe.
study2/quality_cts.json	Movement-quality regression (clinician total score) from frozen features, with out-of-fold predictions and bootstrap comparisons.
study2/study2_results.json	Output of code/r11_study2_analysis.py (the pre-registered tests).
code/	Training, evaluation, and analysis scripts (see below).

Run names

<fusion>_<objective>_s<seed> with fusion ∈ {xattn (symmetric cross-attention), concat, bottleneck, video (video only), skeleton (skeleton only)}, objective ∈ {ce (cross-entropy), align (cross-entropy + InfoNCE, λ = 0.1)}, and seed ∈ {0, 1, 2}.

KIMORE settings

head_only_ref (untrained head, chance reference), linear_probe (frozen synthetic network), syn_adapt (synthetic initialisation + fine-tuning), kim_matched (same recipe without synthetic pretraining), kim_scratch (KIMORE-only, from-scratch recipe), skel_scratch (skeleton stream only).

Code

Script	Purpose
pgxformer_colab_standalone.py	Model definitions (CTR-GCN, VideoMAE wrapper, fusion variants) and the training/evaluation engine.
r2_train_synth.py	One synthetic training run (configuration, seed, optional InfoNCE).
r3_kimore.py	KIMORE five-fold study with inner validation, including the Kinect skeleton parser.
r4_build_reports.py	Builds the natural and stress-set kinematic error reports.
r5_agent_eval.py	Raw vs. safety-gated feedback evaluation (rejection gate + output verifier).
r6_cost.py	Parameters, GFLOPs, latency.
r1_stats_existing_runs.py, r7_aggregate.py	Statistics and aggregation into revision_numbers.json.
r9_label_efficiency.py, r10_quality.py, r11_study2_analysis.py	Study 2 (label efficiency, movement quality, pre-registered analysis).
modal_app.py, r_hf.py, r8_status.py	Parallel, resume-safe execution of the full grid on cloud GPUs.

Quick recomputation of the paper's numbers:

pip install numpy scipy scikit-learn
python code/r7_aggregate.py   # expects predictions/ and agent/ under out/; see the script header

Retraining requires the InfiniteRep archive (linked above) and, for the clinical study, the KIMORE dataset, which must be obtained from its authors (it is not redistributed here).

Data and licences

• InfiniteRep is a synthetic dataset by Infinity AI; the archive linked above is a research-use mirror with attribution.
• KIMORE (Capecci et al., IEEE TNSRE 2019) is not included; only model predictions on it are.
• The files in this repository (code, splits, predictions, audit outputs) are released under CC BY 4.0.

Citation

Please cite the PG-XFormer paper (citation to be added on publication) and the InfiniteRep and KIMORE datasets.