"""Convert the PRATIBHA-2 release (MannasAI/pratibha-2, revision 863edc9) to braindecode's ``Pratibha2``. usage: python convert.py Renames the encoder keys to braindecode's, splits each fused feed-forward input projection ``ffn.in_proj`` (value rows, then gate rows) into ``mlp.fc1`` and ``mlp.fc_gate``, and drops ``pos_enc.freq_matrix``, which the model rebuilds. ```` gets ``config.json`` and ``model.safetensors`` from ``Pratibha2.save_pretrained`` for the 10-20 montage below, 5 s windows and ``n_outputs=2``. The head is not part of the release: it is PyTorch's default init under ``torch.manual_seed(0)``; ``from_pretrained(..., n_outputs=k)`` replaces it. """ import sys import torch from safetensors.torch import load_file from braindecode.models import Pratibha2 RENAMES = [ ("patch_embed.linear.", "encoder.patch_embed."), ("pos_enc.learned_linear.0.", "pos_enc.learned_linear."), ("pos_enc.learned_linear.2.", "pos_enc.learned_norm."), ("pos_enc.", "encoder.pos_enc."), ("encoder.layers.", "encoder.blocks."), (".pre_attn_norm.", ".norm1."), (".attn.qkv_proj.", ".attn.qkv."), (".attn.out_proj.", ".attn.proj."), (".pre_ffn_norm.", ".norm2."), (".ffn.out_proj.", ".mlp.fc2."), ] MONTAGE = "Fp1 Fp2 F7 F3 Fz F4 F8 T7 C3 Cz C4 T8 P7 P3 Pz P4 P8 O1 O2".split() def convert_state_dict(state_dict): out = {} for key, value in state_dict.items(): if key == "pos_enc.freq_matrix": continue for old, new in RENAMES: key = key.replace(old, new) if ".ffn.in_proj." in key: value_rows, gate_rows = value.chunk(2) out[key.replace(".ffn.in_proj.", ".mlp.fc1.")] = value_rows.contiguous() out[key.replace(".ffn.in_proj.", ".mlp.fc_gate.")] = gate_rows.contiguous() else: out[key] = value return out if __name__ == "__main__": src, out_dir = sys.argv[1:] original = load_file(src) torch.manual_seed(0) model = Pratibha2( chs_info=[{"ch_name": name} for name in MONTAGE], n_outputs=2, n_times=1000, sfreq=200.0, ) torch.testing.assert_close( original["pos_enc.freq_matrix"], model.encoder.pos_enc.freq_matrix ) missing, unexpected = model.load_state_dict( convert_state_dict(original), strict=False ) assert not unexpected, unexpected assert all(key.startswith("final_layer.") for key in missing), missing model.save_pretrained(out_dir)