Prior-guided TCN fusion for finger-joint angle estimation — code & configs (review v0.1)

Code and configuration to reproduce the experiments in our Information Fusion submission:“Prior-guided TCN fusion for finger-joint angle estimation from strain sensors and dual IMUs”. What’s inside (software):• Training/evaluation code: models/, utils/, train.py, metrics.py, viz.py, losses.py, signal.py• Configs & scripts: config.yaml, env_cuda.yml (CUDA), env_cpu.yml (fallback), requirements_cuda.txt• One-command reproduction: reproduce.ps1 (Windows) / reproduce.sh (Linux/macOS)• License: Apache-2.0; Citation: CITATION.cff (DOI to be added after acceptance) How to reproduce (CUDA):1) conda env create -f env_cuda.yml conda activate anglefusion_cuda2) Place data and checkpoint from the Dataset item (see “Related materials” below): data/11..15/aligned_full_data_linear_XX.csv checkpoints/std_heads_run_20250921_1919_best.pt3) Run metrics (cross-session, causal, gate enabled, trim 1 s head/tail): python metrics.py --config config.yaml \ --ckpt_ours checkpoints/std_heads_run_20250921_1919_best.pt \ --gate --trim-head-ms 1000 --trim-tail-ms 10004) Optional single-sample overlay (sample11): python viz.py --config config.yaml \ --ckpt checkpoints/std_heads_run_20250921_1919_best.pt \ --file ./data/11/aligned_full_data_linear_11.csv --outdir viz --gate Evaluation protocol:• Strictly causal convolutions, no future leakage within windows• Cross-session split (n=5); report MAE, RMSE, R, R²; 95% CIs across sessions (Student’s t), PCC via Fisher-z• Real-time friendly heads: time-varying gate α_t, affine (a_t, b_t), residual Related materials:• Dataset + minimal checkpoints (private link / DOI reserved) — add under “Links to related material” This record is a review artifact (v0.1). We will publish and attach a DOI after acceptance.