emgbenchanon/muniverse-dynamic-v2

MUniverse Dynamic EMG Benchmark v2

416 synthetic non-stationary surface EMG recordings for benchmarking motor-unit decomposition algorithms under dynamic joint motion. Split into 336 train (subjects 0–4, 21 MU pools) and 80 held-out val (subjects 5–9, 5 MU pools) — different simulated subjects, so the val set is a clean generalisation test.

Both splits share the same 8 conditions (sinusoidal/triangular wrist Flexion–Extension at two amplitudes × two SNR levels) and ship at 70-ch and 320-ch electrode configurations.

Files

File	Size	Purpose
recordings.tar.gz	4.7 GB	Train: 336 .npz recordings, subjects 0–4, 21 MU pools (benchmark_v2/{ch070,ch320}/)
recordings_val.tar.gz	1.2 GB	Val: 80 .npz recordings, subjects 5–9, 5 MU pools (benchmark_v2_val/{ch070,ch320}/)
reproducibility_bundle.zip	36 KB	Scripts, manifests, NeuroMotion patch, PBS templates
benchmark_spec.json	7 KB	Declarative spec: subjects, muscles, conditions, combos
README.md	this page	(also inside the bundle, as the canonical reproduction guide)

SHA256:

• recordings.tar.gz: 715458efd70ca8932e1e45a642698fa11a4446e0352dbe1840e0a7e251544f56
• recordings_val.tar.gz: e4a14c291f4087a90ee10a376b57c359e3507ba4cdd68bb94b6f7b9aa6cc3756
• reproducibility_bundle.zip: f4e034726a5c2afd60f408d3e32d63be074a586235316c38926feafd3d3e0988

Splits

Split	Subjects	Pools	Recordings (per ch config)	Recordings total
Train	seeds 0–4 (5 subjects)	21 (6 small + 8 medium + 7 large)	168	336
Val	seeds 5–9 (5 subjects)	5 (1 small + 2 medium + 2 large)	40	80

Subjects 5–9 are completely disjoint from training subjects, sampled from the same NeuroMotion+BioMime generator with the same fiber-density distribution. Val combos:

Tier	Subject	Muscle	Threshold	N MUs
small	sub5	PL	0.85	29
medium	sub6	ECU	0.85	50
medium	sub7	EDI	0.85	62
large	sub8	FCU_u	0.85	79
large	sub9	ECRB	0.90	92

Quick load

import numpy as np

# Train recording
d = np.load("benchmark_v2/ch320/SA10-hi_ECU_sub03_N055.npz", allow_pickle=True)

# Val recording
d = np.load("benchmark_v2_val/ch320/SA10-hi_ECU_sub06_N050.npz", allow_pickle=True)

emg            = d["emg"]              # (T, M) float32
spike_mu_0     = d["spike_mu_0"]       # (n_spikes,) int64 — ground-truth indices
angle_profile  = d["angle_profile"]    # (T,) joint angle in degrees
effort_profile = d["effort_profile"]   # (T,) effort, fraction MVC
fs             = int(d["fs"])          # 2048

See benchmark_spec.json for the full parameter grid.

Reproduction

The dataset is fully reproducible from open code: NeuroMotion + BioMime → MUAP libraries → crosstalk filtering → NeuroMotion synthesis. Stage-by-stage scripts and PBS templates are inside reproducibility_bundle.zip. The val split was generated with the same pipeline using seeds 5–9 instead of 0–4. The original README inside the bundle walks through all five stages.

---

Original reproducibility README

Below is the canonical reproduction guide as it appears inside reproducibility_bundle.zip.

Complete configuration, scripts, and metadata to regenerate the v2 benchmark: 336 non-stationary EMG recordings across 21 motor-unit pool combinations (26–113 MUs), 5 subjects, 7 muscle types, and 8 conditions (sinusoid/triangular angle × 2 difficulties × 2 SNR).

The benchmark was generated using MUniverse as-is — no modifications to the core package except a single small patch to NeuroMotion's Triangular angle-profile generator (included in scripts/patches/). The story of this bundle is:

1. Generate 40 MUAP libraries using NeuroMotion+BioMime (5 subjects × 8 muscles at Flex-Ext DOF).
2. Analyze pairwise MUAP similarity, filter each library to a "crosstalk-clean" subset.
3. Use NeuroMotion to synthesize 336 EMG recordings from those clean subsets, each exposed to different rotation profiles and noise levels.

Contents

experiments/v2/
├── README.md                              # This file
├── benchmark_spec.json                    # Declarative specification (reviewer-readable)
├── combos.json                            # The 21 subject-muscle-threshold combos used
├── scripts/                               # Scripts used, in execution order
│   ├── 01_generate_muap_caches.py         # Stage 1: Run BioMime to make 40 caches (GPU)
│   ├── 02_analyze_crosstalk.py            # Stage 2: Pairwise similarity analysis
│   ├── 03_create_clean_caches.py          # Stage 3: Filter each cache to a clean subset
│   ├── 04_generate_recordings.py          # Stage 4: Run NeuroMotion on clean caches
│   ├── 05_fix_triangular.py               # Stage 5: Bug fix, regenerate triangular only
│   ├── build_metadata.py                  # (meta) Generates the manifest CSVs below
│   └── patches/
│       └── neuromotion_triangular_symmetric.patch    # NeuroMotion fix
└── metadata/
    ├── muap_caches_manifest.csv           # 40 caches with shapes, fiber density, sha256
    ├── combos_manifest.csv                # 21 clean combo pools
    ├── conditions_manifest.csv            # 8 condition definitions
    └── recordings_manifest.csv            # 336 recordings with full parameters

Prerequisites

1. MUniverse — tested against dynamic_zi_resampling branch (see git log). Clone the repo and install:

   git clone https://github.com/dfarinagroup/muniverse muniverse-live
   cd muniverse-live
   git checkout dynamic_zi_resampling
   pip install -e .
   

2. NeuroMotion Singularity/Docker image — pranavm19/muniverse-test:neuromotion:

   cd src/environment
   singularity pull muniverse-test_neuromotion.sif docker://pranavm19/muniverse-test:neuromotion
   

3. NeuroMotion triangular patch — apply to src/muniverse/data_generation/_run_neuromotion.py:

   cd muniverse-live
   git apply experiments/v2/scripts/patches/neuromotion_triangular_symmetric.patch
   

   (Or just copy the Triangular branch from that file into your _run_neuromotion.py line ~434.)

4. Hardware:

   • Stage 1 (MUAP generation): GPU — each of 40 caches takes ~2-3 min on A40 / RTX6000.
   • Stages 2–5: CPU only.

5. Disk:

   • MUAP caches: ~95 GB permanent
   • Clean caches: ~8 GB
   • Recordings: ~5 GB
   • Raw outputs + configs: ~10 GB ephemeral (safe to delete after gen)

Reproduction Steps

Stage 1: Generate 40 MUAP caches (GPU, ~1.5 hours)

Each cache is a (num_mus, 130, 10, 32, 96) numpy array — MUAPs over 130 angle steps for each MU on a 10×32 electrode grid.

# Generate all 40 (5 subjects × 8 muscles at Flex-Ext DOF)
python experiments/v2/scripts/01_generate_muap_caches.py

The script edits these hardcoded paths — adjust before running:

• CACHE_ROOT: where to write the caches
• RAW_ROOT: scratch dir for container outputs (ephemeral is fine)
• SIF_PATH: path to the NeuroMotion Singularity image

Output: <CACHE_ROOT>/cache_sub-simXX/subject_X_{MUSCLE}_Flexion-Extension_muaps.npy + _metadata.json + _mn_properties.csv.

Stage 2: Crosstalk analysis (CPU, ~5 min)

python experiments/v2/scripts/02_analyze_crosstalk.py

Discovers all Flex-Ext caches, computes pairwise cosine similarity at mid-pose, and writes clean-subset selections at thresholds 0.75 / 0.80 / 0.85 / 0.90 to: experiments/results/crosstalk_v3/exp_crosstalk_clean_subsets.json

Stage 3: Create 21 clean caches (CPU, ~1 min)

python experiments/v2/scripts/03_create_clean_caches.py

Reads the crosstalk JSON and writes 21 filtered caches to experiments/clean_caches_v2/ (organized as {tier}/sub{N}_{muscle}_thr{XX}/...). Also writes combos.json — the master list consumed by Stage 4.

Stage 4: Generate 336 recordings (CPU, ~55 min)

python experiments/v2/scripts/04_generate_recordings.py \
    --config experiments/clean_caches_v2/combos.json

For each combo, runs the container 8 times (one per condition), post-processes into 2 channel configs (ch070 + ch320), and writes: <BENCHMARK_ROOT>/ch{070,320}/{cond}_{muscle}_sub{N}_N{K}.npz

Stage 5 (only if you hit the bug): Fix triangular recordings

If you run Stage 4 without the NeuroMotion patch applied, the 168 triangular recordings will have flat 0° angle profiles. To detect: check any TA*.npz file — angle_profile should oscillate ±sin_amplitude.

If buggy, patch NeuroMotion and regenerate just the triangular recordings:

# Submit as array job (21 combos × ~3 min each)
qsub experiments/v2/run_regenerate_triangular.pbs

# Or run interactively, one combo at a time
for i in {0..20}; do
    python experiments/v2/scripts/05_fix_triangular.py --combo-idx $i
done

Rebuild metadata CSVs

python experiments/v2/scripts/build_metadata.py

Regenerates metadata/*.csv from the actual files on disk (useful for validation).

Data Format

Each recording is a .npz file following /BENCHMARK_DATA_FORMAT.md with:

Key	Shape	Description
emg	(T, M) float32	EMG samples-first
spike_mu_{i}	(n_spikes,) int64	GT spike sample indices per MU (empty for inactive)
angle_profile	(T,)	Joint angle over time (deg)
effort_profile	(T,)	Effort over time (fraction MVC)
fs, n_channels, n_samples, muscle, subject_seed, af, ...	scalars	Full parameters

Conditions Summary

All 8 conditions use constant effort at 50% MVC with 0.3 Hz sinusoid/triangle. Duration 10s, fs=2048.

ID	Angle Profile	af (fraction of ±65°)	±deg	SNR (dB)
SA05-hi	Sinusoid	0.5	±32.5	25
SA05-lo	Sinusoid	0.5	±32.5	20
SA10-hi	Sinusoid	1.0	±65.0	25
SA10-lo	Sinusoid	1.0	±65.0	20
TA05-hi	Triangular	0.5	±32.5	25
TA05-lo	Triangular	0.5	±32.5	20
TA10-hi	Triangular	1.0	±65.0	25
TA10-lo	Triangular	1.0	±65.0	20

Combo Summary

21 combos grouped into 3 pool-size tiers:

Tier	# Combos	Pool range	Typical recording
Small	6	26–45 MUs	PL × 3 subjects + FCU_h + ECRL + FDSI
Medium	8	47–74 MUs	ECU, FDSI, EDI, ECRB, FCU_u
Large	7	77–113 MUs	EDI, ECU, ECRB, FCU_u at threshold 0.90

See combos.json / metadata/combos_manifest.csv for the exact list.

Verification

After reproduction, verify the dataset is correct:

# Compare against recordings_manifest.csv (336 rows)
python experiments/v2/scripts/build_metadata.py
diff <(sort experiments/v2/metadata/recordings_manifest.csv) <(sort <your-regenerated>/recordings_manifest.csv)

MUAP cache SHA256 hashes (first 16 hex chars) are in metadata/muap_caches_manifest.csv for integrity checking of the Stage 1 outputs.

Citation

<pending — paper in prep>

License

Same as MUniverse core (GPL-3.0).