coco_pipe.decoding.foundation_models.validation#

Opt-in real-checkpoint validation for foundation-model registry entries.

Attributes#

DEFAULT_CHANNELS

Functions#

validate_real_checkpoints([model_keys, device, token, ...])

Download each checkpoint and run one headless forward pass.

validate_real_training([model_keys, train_modes, ...])

Run tiny grouped training and checkpoint-reload smoke tests.

main()

Module Contents#

coco_pipe.decoding.foundation_models.validation.DEFAULT_CHANNELS = ['Fp1', 'Fp2', 'F7', 'F3', 'Fz', 'F4', 'F8', 'T7', 'C3', 'Cz', 'C4', 'T8', 'P7', 'P3', 'Pz',...#
coco_pipe.decoding.foundation_models.validation.validate_real_checkpoints(model_keys=('cbramod', 'labram', 'reve', 'luna'), device='cpu', token=None, ch_names=None)#

Download each checkpoint and run one headless forward pass.

Parameters:

  • ch_names (sequence of str, optional) – Montage to validate against. Defaults to DEFAULT_CHANNELS (a representative 19-channel 10-20 set). Pass the study’s actual channel list to exercise the exact channel adaptation (e.g. the real 128 x N LaBraM interpolation) the cohort run will use.

  • model_keys (collections.abc.Sequence[str])

  • device (str)

  • token (str | None)

Return type:

list[dict[str, Any]]

coco_pipe.decoding.foundation_models.validation.validate_real_training(model_keys=('cbramod', 'labram', 'reve', 'luna'), train_modes=('linear_probe', 'full', 'lora'), device='cpu', token=None, max_epochs=1, ch_names=None)#

Run tiny grouped training and checkpoint-reload smoke tests.

ch_names defaults to DEFAULT_CHANNELS; pass the study montage to train against the exact channel adaptation used in cohort runs.

Parameters:
Return type:

list[dict[str, Any]]

coco_pipe.decoding.foundation_models.validation.main()#
Return type:

None