PhysioWave-T-sEMG: multi-route pretrained sEMG encoder

One of the three PhysioWave-T encoders (EEG, ECG, sEMG); the RALF fusion model combines them. Code: https://github.com/ForeverBlue816/PhysioWave.

Masked dual-reconstruction pretraining on surface EMG from wristbands, forearm arrays and high-density grids. One shared transformer serves three electrode layouts through route-specific wavelet frontends, and electrode identity enters as a learned embedding of the electrode NAME rather than of its index.

Trained for 45 epochs of 60 configured; validation was logged for the first 39, which is what metrics_epoch.jsonl holds (the last logged validation at step 39000) on a mixture of public sEMG corpora, 12.1 M parameters. The files here are built around pretrain_emg_c1_moe_epoch38.pth, epoch 39.

Objective

0.5*L_spec + 0.5*L_raw + 0.001*L_foldKL

Two heads at equal weight. One predicts folded wavelet patches, the other predicts the preprocessed sEMG waveform of the same masked patches. Masking is applied BEFORE the wavelet frontend (True), at ratio 0.7, and the spec target is per-patch normalised (True).

Training curves

Pretraining ran for 45 epochs on 16 GPUs over the six corpora, with validation logged for the first 39. The two reconstruction terms behave differently: the folded-wavelet loss is lowest early (epoch 7, which is what best.pth selects on the total), while the raw-waveform loss keeps falling to the end. The released main file is the epoch-38 checkpoint, the one with the lowest raw-waveform loss, and the one every downstream sEMG result of the paper was obtained with.

Train and validation per epoch: the total loss, the two reconstruction terms (folded wavelet patch, raw waveform) and the masked spectral correlation.

Validation, at the released checkpoint (epoch 39)

metric	value	what 1.0 would mean
total loss	0.2580	—
spec MSE	0.3207	—
spec NMSE	0.321	predicting zero
spec Pearson r	0.822	perfect
raw SmoothL1	0.1953	—
raw NMSE	0.381	predicting zero
raw Pearson r	0.782	perfect
route-macro total	0.2315	—

NMSE is dimensionless: MSE(prediction, target) / MSE(0, target), so 1.0 is what predicting zero scores on that head's own target. The two losses are NOT comparable to each other -- one is an MSE on normalised wavelet coefficients, the other a SmoothL1 on the window-normalised waveform. The columns that compare the heads are the correlations and the NMSEs.

The route-macro row weights the three routes equally. The plain rows weight by sample count, and validation is dominated by the largest corpora, so a model specialising toward the smaller ones moves the two rows in opposite directions.

Files

Which file do you want

your situation	download
fine-tune on any montage	emg_c1_encoder.pth
your montage IS one of the three routes below	pretrain_emg_c1_moe_epoch38.pth, then cut the route out (one command, below)
continue pretraining, or take the model apart yourself	pretrain_emg_c1_moe_epoch38.pth

pretrain_emg_c1_moe_epoch38.pth is the complete model -- all three wavelet frontends, the patchers, both reconstruction decoders. Nothing is held back; only the optimiser state was stripped, which no downstream use needs. Take it if you want to keep or strip pieces yourself.

emg_c1_encoder.pth is the same weights minus everything route-bound and pretraining-only. It exists because it cannot be loaded wrong.

file	what it is
emg_c1_encoder.pth	encoder alone: channel embedding, gate, transformer
pretrain_emg_c1_moe_epoch38.pth	complete model, the reported checkpoint (epoch 39)
pretrain_emg_c1_moe_final.pth	complete model, last epoch (epoch 45)
pretrain_emg_c1_moe_best.pth	complete model, lowest validation total loss (epoch 7)
pretrain_emg_c1_moe_best_macro.pth	complete model, lowest route-macro total (epoch 25)
metrics_epoch.jsonl	the full training curve
config_resolved.yaml	every resolved hyperparameter

Why several checkpoints

Each is the minimum of a different validation metric, and they are published together so a downstream comparison can ask which criterion transfers instead of assuming the one that names best.pth. Reconstruction loss cannot answer that question; a downstream task can.

best.pth is selected on the TOTAL loss -- the loss that was trained (epoch 6 counted from 0). It is not the best spec loss, and not the best raw loss.

Why pretrain_emg_c1_moe_epoch38.pth is the main file. It has the lowest validation raw loss; best.pth wins the total loss at epoch 7, before the raw head had converged (raw r 0.740 against 0.782). The paper's sEMG downstream and fusion results use this file.

pretrain_emg_c1_moe_best_spec.pth would have been byte-identical to pretrain_emg_c1_moe_best.pth -- the same epoch won both bars -- so it is not published separately. pretrain_emg_c1_moe_best_raw.pth would have been byte-identical to pretrain_emg_c1_moe_epoch38.pth -- the same epoch won both bars -- so it is not published separately.

What is in the complete model

component	parameters	share
shared transformer	10.63 M	88.2%	encoder, always transfers
channel-name embedding	0.02 M	0.1%	always transfers
channel projection	0.02 M	0.2%	always transfers
wavelet frontends (all three routes)	0.79 M	6.6%	route-bound; downstream usually builds its own
patchers (one rate)	0.10 M	0.8%	route-bound
spec decoder	0.24 M	2.0%	pretraining only
raw decoder	0.24 M	2.0%	pretraining only
total	12.05 M	100%

The shared transformer is 88% of the model and the three frontends together are 7%, which is why there is no per-route encoder file here: three of them would duplicate that transformer three times to deliver frontends that are a small fraction of it. Cut the one you want out of the complete checkpoint instead.

What transfers, and what does not

downstream.py splits a checkpoint into two sets. The channel embedding, its gate and the shared transformer always load. The wavelet frontend and the patcher load only when your montage is the route they were trained for -- same channel count, same patch length -- and are skipped otherwise rather than reshaped.

The usual sEMG benchmark montages -- armbands, forearm rings, NinaPro's electrodes -- are none of these routes, so a fresh frontend is the normal case. Their electrode names are outside the vocabulary too. Each unknown name gets an embedding row of its own, initialised from the pretrained <unk> row and trained as a montage adapter; new_channel_rows=False keeps the one shared <unk> vector instead, and channels are then told apart by the position encoding alone.

The three routes the shared transformer was pretrained across:

route	channels	rate
A24_2000	24	2000 Hz
G64_2000	64	2000 Hz
W16_2000	16	2000 Hz

To fine-tune ON one of them exactly, take that route's frontend and patcher out of the complete checkpoint:

python scripts/export_eeg_pretrained_encoder.py \
    --checkpoint pretrain_emg_c1_moe_epoch38.pth --route G64_2000 \
    --output encoder_G64_2000.pth

The exporter reads the modality from the checkpoint; its name is historical. Omit --route and you get emg_c1_encoder.pth again. Both pretraining decoders are dropped either way: fine-tuning against the pretext objective by accident is the failure that prevents.

The last checkpoint (epoch 45) has no logged validation row, so it cannot be compared with the released one here. Which transfers better is a downstream question -- measure both.

The channel vocabulary travels with the weights

An embedding row means whatever channel held that id when the row was learned. Every checkpoint here carries the vocabulary and its hash. Resolving electrode names against a different vocabulary silently trains on relabelled channels, which is why the hash is checked rather than assumed.

Architecture

embed_dim   384      depth 6      heads 6
norm        rmsnorm     ffn swiglu     qk_norm True
wavelet     max_level 3, kernel 16, bank db2, db4, db7, sym4, sym5
channel     id embedding, injected at the token site
mask        frequency_guided, ratio 0.7

Loading

EMGC1Downstream builds the frontend and patcher for your montage and loads this file into everything else. Nothing here is route-bound, so the montage, the window length and the patch length are yours to choose.

import torch
from physiowave.emg_c1.downstream import EMGC1Downstream

ENC = "emg_c1_encoder.pth"
ck = torch.load(ENC, map_location="cpu", weights_only=False)
mc = ck["model_config"]                  # the width the weights were trained at
CH = [f"ch{i}" for i in range(1, 9)]   # an 8-electrode armband

model = EMGC1Downstream(
    in_channels=len(CH), window_samples=2000, sampling_rate=2000.0,
    patch_samples=250, num_classes=2, channel_names=CH,
    embed_dim=mc["embed_dim"], depth=mc["depth"], num_heads=mc["num_heads"],
    channel_embed_dim=mc.get("channel_embed_dim", 64),
    channel_vocab_size=ck["channel_vocab_size"])

report = model.load_pretrained(ENC)
print(model.describe_transfer(report))
# loaded: transformer N, channel encoder N, gate yes, frontend 0, patcher 0

class Meta:                               # or your dataset's ChannelMeta
    channel_names = CH
    channel_mask = None

out = model(torch.randn(2, len(CH), 2000), Meta())   # {"logits": [2, 2]}

frontend 0, patcher 0 is the expected line, not a warning: those are the two things your montage supplies.

The frontend you build is not randomly initialised. Its filters are seeded from real wavelet families -- db2, db4, db7, sym4, sym5 -- and learned from there during fine-tuning. What a fresh frontend costs you is the adaptation those filters underwent during pretraining, not the wavelet prior itself.

Names, not indices, are what this encoder and your data agree on. Pass real electrode names and check channel_vocab_sha256 -- load_pretrained refuses a mismatch rather than silently training on relabelled channels.

Code: https://github.com/ForeverBlue816/PhysioWave

---

Prepared by scripts/prepare_hf_release.py from pretrain_emg_c1_moe on 2026-10-29.