Local model - Neuroencoder

The same package runs the model locally when you have the weights (shared with collaborating labs; see Access).

from neuroencoder import EPI

model = EPI.from_checkpoint()                 # ~/.cache/neuroencoder/epi300k.pt, or pass a path
emb = model.embed(eeg, sfreq=256, channel_names=ch_names, stride_seconds=30)   # numpy [N, 768]

What the model expects

Six regional averages at 250 Hz: Frontal, Central, Temporal-L, Temporal-R, Parietal, Occipital. preprocess builds them from any 10-20 montage; EOG and ECG channels are accepted in slots 7 and 8 and ignored. Each channel is z-scored inside the model; windows are 30 s (up to 32 s are used).

import neuroencoder as ne
epochs = ne.preprocess(eeg, sfreq=256, channel_names=ch_names, stride_seconds=30)   # [N, 8, 7500]
emb = model.predict(torch.from_numpy(epochs))                                          # [N, 768] on the model device

Throughput

Hardware	Windows per second	1 h of EEG, 30 s stride	1 h, 1 s stride
16 CPU cores	35	3.5 s	100 s
RTX 4090, fp32	1,900	0.06 s	1.9 s
RTX 4090, bf16 autocast	3,000	0.04 s	1.2 s

bf16 changes embeddings by at most 1e-2 per component (cosine 1.0000 to fp32); it is on by default on CUDA and can be disabled with predict(x, autocast=False). Peak GPU memory is 1.3 GB at batch 256.

Visualisation

ne.explore(emb)     # interactive embedding atlas in Jupyter
ne.plot(emb)        # static UMAP