"""Convert ``checkpoint/jung_mae.pt`` (GeoNIRS/GeoNIRS@81f8995) to braindecode's ``JungMAE``. usage: python convert.py Unwraps the ``{"model", "model_config", ...}`` payload (its configuration equals the class defaults), drops the pretraining decoder (input projection, mask token, layers, prediction layer) and writes ``config.json`` and ``model.safetensors`` with ``JungMAE.save_pretrained`` for the paper's 40 channels of 60 samples at 5 Hz. The release holds the encoder only: the head reads ``C * P * D`` features, so ``from_pretrained`` builds it for the caller's montage and ``n_outputs``. """ import os import sys import torch from safetensors.torch import save_file from braindecode.models import JungMAE DECODER = ("decoder_embed.", "mask_token", "decoder.", "prediction.") if __name__ == "__main__": src, out_dir = sys.argv[1:] payload = torch.load(src, map_location="cpu", weights_only=True) encoder = {k: v for k, v in payload["model"].items() if not k.startswith(DECODER)} torch.manual_seed(0) model = JungMAE(n_chans=40, n_outputs=2, n_times=60, sfreq=5.0) missing, unexpected = model.load_state_dict(encoder, strict=False) assert not unexpected, unexpected assert all(key.startswith("final_layer.") for key in missing), missing model.save_pretrained(out_dir) os.remove(os.path.join(out_dir, "pytorch_model.bin")) weights = {k: v for k, v in model.state_dict().items() if not k.startswith("final_layer.")} save_file(weights, os.path.join(out_dir, "model.safetensors"))