Jay1208 commited on
Commit
2d5ba6a
·
verified ·
1 Parent(s): 7e1984b

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +84 -0
README.md ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: samba-eeg
3
+ pipeline_tag: feature-extraction
4
+ tags:
5
+ - eeg
6
+ - brain-computer-interface
7
+ - self-supervised-learning
8
+ - mamba
9
+ - pytorch
10
+ ---
11
+
12
+ # SAMBA
13
+
14
+ EEG representation models. [Code](https://github.com/Jiazhen-Hong/SAMBA) · [Paper](https://arxiv.org/abs/2511.18571)
15
+
16
+ SAMBA was developed during my 2025 internship at Emotiv. The original checkpoints are no longer available, so we retrained both models. SAMBA-100s follows the original training method with a rebuilt TUAB pipeline and validation split. SAMBA-2s adds HBN, Muse, and DICE to the Emotiv data for greater electrode-layout diversity.
17
+
18
+ ## Checkpoints
19
+
20
+ Both models use 128 Hz EEG. The names indicate the pretraining window; the encoders support other input lengths and channel layouts.
21
+
22
+ | | [SAMBA-2s](SAMBA-2s) | [SAMBA-100s](SAMBA-100s) |
23
+ |---|---|---|
24
+ | Window | 2 s, 256 samples | 100 s, 12,800 samples |
25
+ | Data | Emotiv¹, HBN, Muse, DICE | TUAB |
26
+ | Pretraining channels | 4, 14, 19, 129 | 16 bipolar |
27
+ | Parameters | 1.06M | 1.06M |
28
+
29
+ ¹ Alpha, Attention, Crowdsource, DriverDistraction, and STEW.
30
+
31
+ ## Usage
32
+
33
+ Requires a compatible CUDA system.
34
+
35
+ ```bash
36
+ git clone https://github.com/Jiazhen-Hong/SAMBA.git
37
+ cd SAMBA
38
+ pip install -e '.[mamba,hub]'
39
+ ```
40
+
41
+ ```python
42
+ from pathlib import Path
43
+ import torch
44
+ from huggingface_hub import snapshot_download
45
+ from samba_eeg import SambaModel
46
+
47
+ variant = "SAMBA-2s" # or "SAMBA-100s"
48
+ folder = snapshot_download(
49
+ "Jay1208/SAMBA",
50
+ allow_patterns=[f"{variant}/*"],
51
+ )
52
+ model = SambaModel.from_pretrained(Path(folder) / variant).cuda().eval()
53
+ ```
54
+
55
+ Prepare EEG at 128 Hz:
56
+ - `signals`: `[batch, channels, samples]`.
57
+ - `positions`: `[channels, 3]` or `[batch, channels, 3]`, matching the channel order and the coordinate units and frame in `config.json`.
58
+ - Normalize each channel in each window by its absolute 95th percentile. TUAB pretraining uses a 0.5–45 Hz bandpass.
59
+
60
+ ```python
61
+ signals = signals.to(device="cuda", dtype=torch.float32)
62
+ positions = positions.to(device="cuda", dtype=torch.float32)
63
+
64
+ # Skip if already normalized.
65
+ scale = torch.quantile(signals.abs(), 0.95, dim=-1, keepdim=True)
66
+ signals = signals / scale.clamp_min(1e-8)
67
+
68
+ with torch.inference_mode():
69
+ features = model.encode(signals, positions)
70
+ pooled_features = features.mean(dim=1)
71
+ ```
72
+
73
+ `features` has shape `[batch, floor(samples / 8), 128]`. Use the pooled features for linear probing, or fine-tune the encoder.
74
+
75
+ ## Citation
76
+
77
+ ```bibtex
78
+ @article{hong2026samba,
79
+ title={SAMBA: Toward an EEG Foundation Model via Differential Mamba and Coordinate-Driven Spatial Embedding},
80
+ author={Hong, Jiazhen and Mackellar, Geoffrey and Ghane, Soheila},
81
+ journal={arXiv preprint arXiv:2511.18571},
82
+ year={2026}
83
+ }
84
+ ```