Add EEGPT backbone

CHANGELOG.md

@@ -7,6 +7,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+- Add the EEGPT backbone ([#40](https://github.com/braindecode/OpenEEGBench/pull/40)).
 
 ## [0.5.0] - 2026-05-28
 

open_eeg_bench/backbone_utils.py

@@ -0,0 +1,20 @@
+import numpy as np
+from braindecode.models import EEGPT, InterpolatedModel
+from braindecode.models.eegpt import EEGPT_CHANNELS
+from mne.channels import make_standard_montage
+
+montage = make_standard_montage("standard_1020")
+ch_pos = {
+    ch.upper(): (ch, loc) for ch, loc in montage.get_positions()["ch_pos"].items()
+}
+_EEGPT_TARGET_CHS_INFO = [
+    {
+        "ch_name": ch_pos[ch.upper()][0],
+        "kind": "eeg",
+        "loc": ch_pos[ch.upper()][1],
+    }
+    for ch in EEGPT_CHANNELS
+]
+InterpolatedEEGPT = InterpolatedModel(
+    EEGPT, _EEGPT_TARGET_CHS_INFO, name="InterpolatedEEGPT"
+)

open_eeg_bench/default_configs/backbones.py

@@ -84,11 +84,26 @@ def reve(**overrides) -> PretrainedBackbone:
     return PretrainedBackbone(**defaults)
 
 
+def eegpt(**overrides) -> PretrainedBackbone:
+    defaults = dict(
+        # model_cls="braindecode.models.EEGPT",
+        # model_cls="braindecode.models.InterpolatedEEGPT",
+        model_cls="open_eeg_bench.backbone_utils.InterpolatedEEGPT",
+        model_kwargs={"chan_proj_type": "none", "n_chans_target": 19},
+        peft_ff_modules=["qkv", "fc1", "fc2"],
+        normalization=WindowZScore(),
+        hub_repo="braindecode/eegpt-pretrained",
+    )
+    defaults.update(overrides)
+    return PretrainedBackbone(**defaults)
+
+
 ALL_BACKBONES = {
     "biot": biot,
     "labram": labram,
     "bendr": bendr,
     "cbramod": cbramod,
     "signal_jepa": signal_jepa,
     "reve": reve,
+    "eegpt": eegpt,
 }