amirsh1376/hdemg-encoder-report

Overnight HD-EMG Encoder Campaign — Report (2026-11-07, 23:50–06:00 EST)

Goal: find the best architecture + self-supervised objective for encoding the new 128-channel, 2 kHz Intan HD-EMG, fine-tuned to predict the top-quality MANO hand fits, on 4× H200 for one night.

Headline: EMG→pose works. The best recipe — spectrogram encoder (per-channel log-STFT ViT), MAE pretraining, fine-tuned on the top-15% quality fits — reaches 18.9 ± 0.2 mm in-session val MPJPE across 4 seeds (vs 34.6 mm mean-pose baseline, 21.7–25.6 mm shuffled-EMG control). The gains survive strict subject holdout for in-session decoding. The open problem is cross-session transfer: no variant beat the shuffled-EMG control on a held-out session (electrode-shift domain gap), and neither finetune-time nor SSL-time electrode-shift augmentation closed it.

Final results (best epoch by in-session MPJPE; mm)

run	best ep	in-sess	in-sess mean-pose	in-sess shuffled	x-sess	x-sess mean-pose	x-sess shuffled
ft_spec_top15 (winner, 4 seeds)	~22	18.7–19.1	34.6	21.7–25.6	29.7–32.2	30.6	27.3–28.5
ft_subject_holdout (no subj07 in train)	27	18.9	34.6	23.6	29.7	30.6	28.0
ft_spec (top-30)	31	20.0	36.6	26.5	34.4	33.7	29.4
ft_best_shot (spec-long enc, 80 ep)	39	21.1	34.6	23.8	34.3	30.6	27.7
ft_spec_long (120-ep SSL)	33	21.9	36.6	23.2	36.6	33.7	31.0
top30_insession (raw transformer)	21	22.0	36.6	28.2	33.2	33.7	31.4
ft_spec_win4s (4 s windows)	20	22.0	36.6	32.5	39.4	33.7	31.8
ft_spec_top15_scratch (no SSL)	35	22.4	34.6	25.5	33.6	30.6	28.3
ft_large_stable (59M raw, LLRD)	23	22.5	36.6	24.0	32.6	33.7	31.9
ms_insession (MotorSense TDS+biLSTM)	27	22.5	34.6	42.0	30.5	30.6	30.3
top30_aug_strong (ft-time augs)	25	22.7	36.6	24.9	33.5	33.7	30.8
ft_augmae (SSL-time augs)	27	24.2	36.6	30.2	35.2	33.7	31.0
ft_clip (CLIP-multiview enc)	19	24.3	34.6	26.2	31.5	30.6	28.4

Wave-1 runs (cross-session val only, before the dual-val feature): ft_mae_raw_base 27.2, scratch raw 27.8, ft_mae_raw_large 27.8, MotorSense baseline 28.6 — all ≈ their shuffled controls, which is what forced the dual-val redesign mid-night.

Linear probes (frozen features → ridge): MAE-raw in-session 25.3 vs 34.6 mean-pose (features carry pose); CLIP-multiview 26.9; cross-session all ≈ baseline. CLIP-multiview retrieval: frame↔EMG top-1 54–68% vs 0.2% chance on training sessions.

Conclusions

1. The HD-EMG encoder extracts real pose signal. Real-EMG beats shuffled-EMG in-session by 3–7 mm on the winner (consistent across 4 seeds), and the mean-pose baseline by ~16 mm.
2. Spectrogram input beats raw-waveform at this data scale (18.9 vs 22.0), echoing emg2qwerty-style findings. Model capacity beyond ~15M params doesn't help (large = 22.5).
3. MAE pretraining is worth ~3.5 mm on the winning config (18.9 pretrained vs 22.4 scratch); it also converges ~2× faster (best epoch 22 vs 35). Longer SSL pretraining (120 ep) did NOT help (21.9) — 60 epochs suffices at this dataset size.
4. Top-15% quality filtering beats top-30% (18.9 vs 20.0) — cleaner supervision beats 2× data.
5. Cross-session transfer is the open problem. Every variant ≈ or worse than shuffled control cross-session. Electrode-shift-style augmentation (gain/channel-mix/noise) failed at both finetune time (22.7, diluted the EMG gap) and SSL time (24.2). Recommended next directions: per-session input adapters/alignment, test-time normalization over longer context, more sessions per subject in future recordings.
6. CLIP-multiview (EMG ↔ frozen DINOv2 frame embeddings across all 6 calibrated views, multi-positive InfoNCE, same-session batches per MotorSense) learns strong view-invariant alignment (54–68% retrieval top-1) but is worse than MAE for pose decoding (24.3 vs 18.9). Worth keeping for semantic/action tasks rather than pose.
7. MotorSense's TDS-conv+biLSTM (the paper's encoder, input stem adapted to 128 ch, EMG-only — sessions have no IMU) is the most EMG-dependent model (shuffled control degrades to 42 mm) but absolute accuracy trails the spectrogram ViT (22.5 vs 18.9).

Optimizations (measured)

change	before	after
GPU-resident data loader (EMG lives on GPU, zero workers)	~900 win/s (CPU loader, 3 workers)	2,900–3,540 win/s; pretrain epoch 58 s → 11–13 s
Finetune epoch time (batch 512 vs 96, GPU loader)	81–95 s	38–39 s
HF download auth (fixed 429 throttling)	~15 MB/s	~700 MB/s
Frame-embedding extraction, 4-way process parallelism	22 img/s (1 proc)	~36 img/s aggregate (CPU-decode bound)
Caches moved off over-quota surgtac share (HF_HOME, TORCH_HOME)	hard failures	working

Known env quirks: autofs idle-timeout on /beacon-projects/srgfoundation intermittently mis-resolves paths (a keep-alive loop now pins it); CLI float overrides like 1e-4 parsed as strings (fixed in repo); stray envs/ + third_party_motorsense/ dirs were accidentally created at the share root during a mount flake and can be deleted.

Artifacts

• Winning encoder: artifacts/pretrain/mae_spec_base/checkpoint_last.pt
• Winning finetuned model: artifacts/finetune/ft_spec_top15/checkpoint_best.pt (+ seeds 1–3)
• Full per-epoch metrics: artifacts/*/*/results.jsonl; table generator scripts/compile_report.py
• Profiling details: artifacts/PROFILE.md

---

Daytime campaign (2026-11-07, 07:39–~12:00): in-session MPJPE below 18.9

Headline: 18.9 ± 0.2 → 18.25 ± 0.18 mm (5 seeds), entirely from how the fit labels are SAMPLED. Winner: train over ALL validity-gated fits (343k frames, not the top-15% cut) with sampling probability exp((quality_score − 1)/τ), τ = 0.2. Validation frames are bit-identical across every run below (same 9,280 in-session frames + same held-out session); only training changed.

τ sweep (soft quality-weighted sampling; hard top-15 cut = 18.9)

τ	0.05	0.10	0.20	0.30	0.40
in-sess mm	19.5	18.5	18.11 (seeds 18.25±0.18)	18.34	18.25

Everything tried (in-session, fixed val; full table: DELTA_TABLE.md)

Positive:

• Soft quality-weighted sampling, τ=0.2: 18.25 ± 0.18 (5 seeds).

Ties with the winner (do not add on top of it):

• 100 ms label smoothing (18.14), motion-weighted sampling (18.30), pose-diversity sampling k-means 1/count^0.5 (18.37, shuffled 21.8) — quality weighting evidently does not starve pose diversity.

Negatives (all on the same fixed val):

• Mask ratio 0.85 / 0.90: 20.0 / 20.4 (0.75 confirmed; 2 s windows leave too few visible time slices at high ratios).
• Temporal decoding over FROZEN features, horizons 0/4/8/16 s: 20.5–21.8, monotonically worse with context.
• Dense along-window supervision (17 time-queries, emg2pose-style): 19.3.
• Rotation-decomposed targets (6D global rot + canonical articulation, recomposed): 20.6 — orientation error compounds through recomposition, although orientation is >half the target variance (canonical std 8.7 mm vs 18.9 mm world).
• Channel dropout 10%: 20.6 (with every other augmentation negative too: in-session training is not overfitting-limited).
• DINOv2 ViT-S trunk transplant, direct finetune: 22.4 = scratch (image pretraining alone transfers nothing; the 3.5 mm pretraining gain is EMG-domain MAE specifically). DINOv2→MAE warm-start chain: in flight.
• Scaled SSL corpus (13 sessions, 2.2×): 19.5 — other subjects' unlabeled EMG DILUTES; pretrain corpus now pinned to the 5 train sessions in config.
• STFT frontend n_fft=128: 19.5. Fingertip-weighted loss: 19.0. Multitask joints+MANO-PCA: 19.0. DINOv2 distillation aux (w=0.25): 19.0.

Real-vs-shuffled delta (user's target metric)

Per-run real/shuffled/Δ columns: artifacts/DELTA_TABLE.md. Key insight: at equal top-tier accuracy (≤18.5 mm), Δ varies 3–4× — smooth100 (18.1, Δ+7.0) and τ=0.4 (18.3, Δ+8.5) rely far more on EMG than plain τ=0.2 (18.1, Δ+2.9). Caveat: Δ also rewards brittle priors (MotorSense: Δ+19.5 at 22.5 mm absolute), so report both columns, never Δ alone. Combo run (τ0.4+smooth+motion, ft_delta_combo): in flight.

Afternoon additions (user-directed, 11:00–14:00)

• Audio-pretraining playbook, all four mechanisms tested — in-domain MAE remains unbeaten: DINOv2 trunk transplant alone 22.4 (= scratch); DINOv2 → MAE warm start 19.3 (worse than scratch-MAE 18.9); HuBERT-style discrete k-means targets 20.3; (DINOv2-as-distillation-target was 19.0 earlier).
• Dataset-size for fine-tuning (user question): 13-session fits, same val: 18.88 step-matched, 18.81 at 2.7× compute — both WORSE than 5-session 18.25 ± 0.18. More subjects' labels dilute; matched compute is not the cause.
• Enlarged-validation benchmark (ft_bigval, new numbers, own scale): in-session over 10 sessions' tails (13,939 frames): 20.0 vs 38.3 mean-pose / 22.9 shuffled — decoding replicates across the new subjects. Cross-session over 4 held-out re-donnings (31,421 frames, every subject has a train session): 32.3 vs shuffled 29.7 — the cross-session gap is electrode-state, NOT subject identity (4× the earlier evidence).
• Masking × window joint cell (user hypothesis): 4 s + mask 0.85 = 21.5 — higher masking helps at longer windows (vs 22.0 at 4 s + 0.75) but the long window itself stays the dominant penalty; 2 s + 0.75 remains best.
• Sleeve CNN-attention architecture (after github.com/lokichubs/sEMG-model, circular electrode conv + multi-scale convs + post-norm attention, ELU): from scratch 23.8 (shuffled 26.3); with MAE pretraining 22.7 (shuffled 24.6) — pretraining transfers to this architecture too (+1.1 mm) but it trails every ViT variant. NOTE: that repo's sleeve_geometry_model.py encodes an inferred physical layout for what looks like the same sleeve: 26 rings × 5 slots built from 10 vertical strips — worth confirming against the hardware; it would unlock 2D spatial modeling.
• Δ-combo (τ0.4+smooth+motion): 18.69/Δ5.0 — the high-Δ ingredients do not stack; τ0.4 alone (18.3/Δ8.5) or τ0.2+smooth (18.1/Δ7.0) are the Pareto picks when the real-vs-shuffled delta is the headline metric.

In flight at compile time

ft_scale13 (fits for 8 new sessions downloaded; 13-session finetune, step-matched, same val — the dataset-size question), ft_delta_combo, DINOv2→MAE→ft chain, 4 s × mask-0.85 joint cell, HuBERT-style discrete-target SSL chain queued next, ft_bigval (enlarged validation: in-session tails of 10 train sessions + 4 held-out cross-sessions). Results land in results.jsonl per run; regenerate tables with scripts/compile_report.py + the DELTA_TABLE snippet.

Facts recorded today

• Electrode sleeve placement is identical across recordings; channel index i is the same physical site in every session (user). Grid adjacency map still unknown — needed for 2D spatial modeling.
• Fit labels now exist for ALL 14 sessions (downloaded); EMG for all 14 also local. Auto-discovery in pretrain configs was a contamination hazard and is now pinned.

---

Overnight fusion campaign (2026-11-08, 00:15–03:30): EMG + WiLoR predictions

Directive: build on the Runpod "fusion server" agent's work (reports/hdemg_emg_wilor_fusion.md in the dataset repo) but do DIFFERENT things, and use WiLoR only as a predictor at training time — no precomputed/cached features (we don't have that compute or storage).

Design (deliberately complementary to the fusion server)

	Fusion server (Runpod, 80 pods)	This campaign (4 GPUs)
WiLoR input	ViT-H backbone tokens, PCA-256, 2.5 TB cached offline	uploaded predictions only (MANO params → joints computed at startup, in RAM, nothing on disk)
Views	one random camera per sample	all 6 cameras fused at once
Output frame	camera frame (deployable track)	rig world frame, our fixed in-session protocol (comparable to every EMG-only row)
Temporal	center frame only	optional WiLoR-trajectory tokens over ±1 s

New modules: hdemg/finetune/manomin.py (pure-torch MANO joints forward from MANO_RIGHT.pkl, no smplx/chumpy; verified against the fits: PA-MPJPE 4.6–7.9 mm, per-camera 10–38 mm — same as the server's smplx numbers), wilor_store.py (per-session world-frame per-view joints + reliability features: orientation 6d, detector score, self-occlusion depth, crop size; built on the GPU in ~90 s at startup), mvfusion.py (token fusion: 16 pooled EMG tokens + 6 per-camera WiLoR tokens [+ trajectory tokens]; concat or WiLoR-anchored head; modality dropout gives fused / emg_only / wilor_only readouts from one checkpoint; independent per-view dropout; k-view eval sweep; WiLoR-permuted control).

Also adopted from the server's code (reviewed hunk by hunk): EMA of weights, micro-batch accumulation, controls_v2 (the legacy "shuffled-EMG" control was a roll-by-1 inside time-sorted batches — windows a median 42 ms apart, ~98% overlap, nearly a no-op; the proper control is a within-session permutation), checkpoint_best_insession, and their channel-correlation sleeve geometry (two 13×5 grids, hdemg/data/spatial.py, inactive).

EMG-only: new record

run	in-session MPJPE	legacy roll1 control	PROPER permutation control	mean-pose
ft_ema_lr5e5 (τ0.2 + EMA 0.999 + lr 5e-5), 5 seeds	17.94 ± 0.10 mm (17.86–18.06)	~18.9–19.1	33.8–34.1	37.2

EMA + lr 5e-5 replicates the server's finding on our protocol (−0.3 mm vs 18.25±0.18, and tighter across seeds: ±0.10 vs ±0.18). The proper permutation control rewrites the attribution story: EMG accounts for ~16 mm, not the ~5 mm the legacy roll control suggested. τ×EMA interaction: τ 0.1/0.2/0.3/0.4 under EMA → 18.23 / 17.94 / 18.04 / 18.00 — τ0.2 stays the peak and EMA flattens the curve (label sampling saturated, as the server found). 60 epochs instead of 40: 17.97 — no gain.

Multi-view WiLoR + EMG (fixed in-session val, mm)

Raw multi-view score-weighted WiLoR mean (no training): 9.6 on val frames (10.9–15.5 over whole sessions). Trained fusion:

run	fused (6 views)	vision-only	1 view fused	1 view vision-only
ft_mv_concat (40 ep)	3.26	3.44	12.88	15.40
ft_mv_anchor	3.72	4.24	14.63	17.73
ft_mv_temporal (±1 s traj)	4.01	4.54	14.29	16.61
ft_mv_noinit (no EMG warm start)	3.27	3.84	13.02	16.08
ft_mv_views05 (concat + per-view drop 0.5; 3 seeds)	3.96 ± 0.02	4.89	7.56–7.70	8.80–8.98
ft_mv_views05_ep80	3.67	4.00	7.61	8.39
ft_mv_anchor_views05 (2 seeds)	4.73–4.82	5.76–5.80	7.68–7.70	9.53–9.61
ft_mv_views08 (drop 0.8)	5.78	6.52	7.98	8.90
ft_mv_views05_pw50 (+whole-vision drop 0.5)	4.49	5.06	8.29	9.32
ft_mv_views05_pe0 (EMG never dropped)	4.05	(path untrained)	7.67	—
ft_mv_visonly (DEDICATED vision-only, EMG never present)	—	4.47	—	8.42

k-view sweep of the winner (ft_mv_views05, fused / dedicated-vision-only): 1 view 7.6/8.4 · 2 views 5.4/6.0 · 3 views 4.6/5.1 · 4 views 4.2/4.7 · 6 views 3.96/4.47.

Controls: WiLoR-permuted (store rows shuffled within session) 24–29 mm on every fusion model; EMG-permuted (controls_v2) on ft_mv_concat: 3.26 → 4.13.

Conclusions

1. EMG improves on vision at every view count, most where vision is weakest: +0.5 mm over the dedicated vision-only model at 6 views, +0.7–0.9 mm at 1 view (up to +1.9 mm against the same-model vision-only readout, and +2.5–3.1 mm in models not trained for missing views).
2. Per-view dropout training is what makes degraded vision work: without it, single-view error collapses from ~7.6 to 12.9–14.6 mm.
3. Not needed: WiLoR-trajectory tokens, EMG encoder warm start, always-on EMG, heavier whole-vision dropout. Anchoring (tanh-gated residual on the WiLoR mean) is safe but concat is better here.
4. Circularity caveat for all vision rows: the val ground truth is MANO fits optimized from these same WiLoR detections, and val frames are the top-15% by fit quality. The defensible claims are the EMG deltas (fused vs vision-only, EMG-permuted controls), not the absolute 3–4 mm numbers.
5. The EMG-only line to carry forward: 17.86 mm with EMA+lr5e-5, and controls_v2 as the standard attribution metric (real | roll1 | permuted | mean-pose).

---

Appendix: every run, real vs shuffled-EMG (in-session)

run	in-sess MPJPE (real EMG)	shuffled EMG	delta
ft_ema_lr5e5	17.9	25.3*	+7.4
ft_weighted_tau20_seed2	18.0	23.1	+5.0
ft_weighted_tau20	18.1	21.0	+2.9
ft_smooth100	18.1	25.2	+7.0
ft_weighted_tau20_seed4	18.2	23.1	+4.9
ft_weighted_tau40	18.3	26.8	+8.5
ft_motion_g1	18.3	24.9	+6.6
ft_weighted_tau30	18.3	20.6	+2.2
ft_diversity_b50	18.4	21.8	+3.5
ft_weighted_tau20_seed1	18.4	21.2	+2.8
ft_weighted_tau20_seed3	18.4	20.6	+2.1
ft_weighted_tau10	18.5	21.7	+3.1
ft_delta_combo	18.7	23.6	+5.0
ft_distill_w25	18.7	22.2	+3.5
ft_spec_top15_seed1	18.7	21.7	+3.0
ft_scale13_long	18.8	21.7	+2.9
ft_spec_top15	18.9	23.6	+4.7
ft_scale13	18.9	23.0	+4.1
ft_spec_top15_seed3	18.9	22.4	+3.5
ft_subject_holdout	18.9	23.6	+4.6
ft_tips2	19.0	22.1	+3.1
ft_multitask	19.0	21.3	+2.3
ft_spec_top15_seed2	19.1	25.6	+6.4
ft_dense_w	19.3	22.3	+3.0
ft_dinoinit_mae	19.3	24.8	+5.4
ft_spec_f128	19.5	24.8	+5.3
ft_spec_alldata	19.5	24.3	+4.8
ft_weighted_tau05	19.5	25.2	+5.7
ft_spec	20.0	26.5	+6.5
ft_spec_m85	20.0	25.8	+5.8
ft_bigval	20.0	22.9	+2.9
ft_discrete	20.3	25.7	+5.4
ft_spec_m90	20.4	24.3	+3.9
td_h0	20.5	21.9	+1.3
ft_canon	20.6	33.6	+13.1
ft_chandrop10	20.6	24.0	+3.3
td_h4	20.7	21.7	+0.9
ft_best_shot	21.1	23.8	+2.8
ft_spec_w4_m85	21.5	26.1	+4.7
td_h8	21.5	22.7	+1.2
td_h16	21.8	22.7	+1.0
ft_spec_long	21.9	23.2	+1.2
top30_insession	22.0	28.2	+6.3
ft_spec_win4s	22.0	32.5	+10.5
ft_spec_top15_scratch	22.4	25.5	+3.1
ft_dinoinit	22.4	25.0	+2.6
ft_large_stable	22.5	24.0	+1.4
ms_insession	22.5	42.0	+19.5
ft_sleeve_mae	22.7	24.6	+1.9
top30_aug_strong	22.7	24.9	+2.3
ft_sleeve_scratch	23.8	26.3	+2.5
ft_canon_w20	23.9	—	—
ft_augmae	24.2	30.2	+6.0
ft_clip	24.3	26.2	+1.9

*ft_ema_lr5e5's control column is the legacy roll-by-1 "shuffle" (25.3) for comparability with the rows above. The fusion server showed that control is nearly a no-op (the rolled window is ~42 ms away, 98% overlap); a PROPER within-session permutation on the same checkpoint gives 33.96 mm vs mean-pose 37.2 — EMG accounts for ~16 mm, not ~5. All runs from 2026-11-08 on also log controls_v2 (real | roll1 | permuted) in results.jsonl.