braindecode 1.8.1.dev180679353__tar.gz → 1.8.1.dev180761562__tar.gz

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/NOTICE.txt

@@ -40,6 +40,7 @@ The following components are licensed under the MIT License:
 
 - `braindecode/models/ctnet.py`
 - `braindecode/models/dance.py`
+- `braindecode/models/duin.py`
 - `braindecode/models/medformer.py`
 - `braindecode/models/tcformer.py`
 - `braindecode/models/ifnet.py`

{braindecode-1.8.1.dev180679353/braindecode.egg-info → braindecode-1.8.1.dev180761562}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev180679353
+Version: 1.8.1.dev180761562
 Summary: Deep learning software to decode EEG, ECG or MEG signals
 Author-email: Robin Tibor Schirrmeister <robintibor@gmail.com>, Bruno Aristimunha Pinto <b.aristimunha@gmail.com>, Alexandre Gramfort <agramfort@meta.com>
 Maintainer-email: Alexandre Gramfort <agramfort@meta.com>, Bruno Aristimunha Pinto <b.aristimunha@gmail.com>, Robin Tibor Schirrmeister <robintibor@gmail.com>

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/augmentation/__init__.py

@@ -10,6 +10,7 @@ from .transforms import (
     ChannelsReref,
     ChannelsShuffle,
     ChannelsSymmetry,
+    CutMix,
     FrequencyShift,
     FTSurrogate,
     GaussianNoise,
@@ -36,6 +37,7 @@ __all__ = [
     "SignFlip",
     "FTSurrogate",
     "ChannelsShuffle",
+    "CutMix",
     "ChannelsDropout",
     "GaussianNoise",
     "ChannelsSymmetry",

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/augmentation/base.py

@@ -171,7 +171,13 @@ class Compose(Transform):
         self.transforms = transforms
         super().__init__()
 
-    def forward(self, X, y):
+    def forward(self, X, y=None):
+        # like Transform.forward, return only X when no target is given (each
+        # transform then returns a single tensor, not an (X, y) pair)
+        if y is None:
+            for transform in self.transforms:
+                X = transform(X)
+            return X
         for transform in self.transforms:
             X, y = transform(X, y)
         return X, y

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/augmentation/functional.py

@@ -22,6 +22,21 @@ from torch.fft import fft, ifft
 from torch.nn.functional import pad
 
 
+def _max_float_dtype(device: torch.device | str) -> torch.dtype:
+    """float64, or float32 on Apple MPS, which has no float64."""
+    return torch.float32 if torch.device(device).type == "mps" else torch.float64
+
+
+def _to_device(t: torch.Tensor, device: torch.device | str) -> torch.Tensor:
+    """Move ``t`` to ``device``, in single precision on MPS (no float64 there)."""
+    if torch.device(device).type == "mps":
+        if t.dtype == torch.float64:
+            return t.float().to(device)
+        if t.dtype == torch.complex128:
+            return t.to(torch.complex64).to(device)
+    return t.to(device)
+
+
 def identity(X: torch.Tensor, y: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
     """Identity operation.
 
@@ -92,10 +107,11 @@ def _new_random_fft_phase_odd(
     rng = check_random_state(random_state)
     random_phase = torch.from_numpy(
         2j * np.pi * rng.random((batch_size, c, (n - 1) // 2))
-    ).to(device)
+    )
+    random_phase = _to_device(random_phase, device)
     return torch.cat(
         [
-            torch.zeros((batch_size, c, 1), device=device),
+            torch.zeros((batch_size, c, 1), device=device, dtype=random_phase.dtype),
             random_phase,
             -torch.flip(random_phase, [-1]),
         ],
@@ -113,12 +129,13 @@ def _new_random_fft_phase_even(
     rng = check_random_state(random_state)
     random_phase = torch.from_numpy(
         2j * np.pi * rng.random((batch_size, c, n // 2 - 1))
-    ).to(device)
+    )
+    random_phase = _to_device(random_phase, device)
     return torch.cat(
         [
-            torch.zeros((batch_size, c, 1), device=device),
+            torch.zeros((batch_size, c, 1), device=device, dtype=random_phase.dtype),
             random_phase,
-            torch.zeros((batch_size, c, 1), device=device),
+            torch.zeros((batch_size, c, 1), device=device, dtype=random_phase.dtype),
             -torch.flip(random_phase, [-1]),
         ],
         dim=-1,
@@ -178,7 +195,7 @@ def ft_surrogate(
         and 0 <= phase_noise_magnitude <= 1
     ), f"eps must be a float between 0 and 1. Got {phase_noise_magnitude}."
 
-    f = fft(X.double(), dim=-1)
+    f = fft(X.to(_max_float_dtype(X.device)), dim=-1)
     device = X.device
 
     n = f.shape[-1]
@@ -580,6 +597,7 @@ def _frequency_shift(X: torch.Tensor, fs: float, f_shift: float) -> torch.Tensor
         _f_shift = torch.as_tensor(f_shift).float()
     else:
         raise ValueError(f"Invalid f_shift type: {type(f_shift)}")
+    _f_shift = _f_shift.to(X.device)
     f_shift_stack = _f_shift.repeat(N_padded, n_channels, 1)
     reshaped_f_shift = f_shift_stack.permute(
         *torch.arange(f_shift_stack.ndim - 1, -1, -1)
@@ -911,7 +929,15 @@ def _rotate_signals(
     sensors_positions_matrix: torch.Tensor,
     spherical: bool = True,
 ) -> torch.Tensor:
-    sensors_positions_matrix = sensors_positions_matrix.to(X.device)
+    if X.device.type == "mps":
+        # MPS has no float64: compute the small (n_chans x n_chans) spherical
+        # spline matrices on CPU in double precision, only move the result
+        # (two separate steps: a combined device + dtype .to() from MPS
+        # returns zeros with some torch versions)
+        sensors_positions_matrix = sensors_positions_matrix.cpu().double()
+        rotations = [rotation.cpu().double() for rotation in rotations]
+    else:
+        sensors_positions_matrix = sensors_positions_matrix.to(X.device)
     rot_sensors_matrices = [
         rotation.matmul(sensors_positions_matrix) for rotation in rotations
     ]
@@ -921,9 +947,10 @@ def _rotate_signals(
                 torch.as_tensor(
                     _torch_make_interpolation_matrix(
                         sensors_positions_matrix.T, rot_sensors_matrix.T
-                    ),
-                    device=X.device,
-                ).float()
+                    )
+                )
+                .float()
+                .to(X.device)
                 for rot_sensors_matrix in rot_sensors_matrices
             ]
         )
@@ -1075,6 +1102,54 @@ def mixup(
     return X_mix, (y_a, y_b, lam)
 
 
+def cutmix(
+    X: torch.Tensor,
+    y: torch.Tensor,
+    idx_perm: torch.Tensor,
+    cut_t_start: torch.Tensor,
+    cut_t_len: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Paste a time region of shuffled examples into the original examples.
+
+    See [1]_ for details.
+
+    Parameters
+    ----------
+    X : torch.Tensor
+        EEG data in form ``batch_size, n_channels, n_times``
+    y : torch.Tensor
+        Target of length ``batch_size``
+    idx_perm : torch.Tensor
+        Permuted indices of example that are mixed into original examples.
+    cut_t_start : torch.Tensor
+        Start index of the pasted time region for each example.
+    cut_t_len : torch.Tensor
+        Length of the pasted time region for each example.
+
+    Returns
+    -------
+    tuple
+        ``X``, ``y``. Where ``X`` is augmented and ``y`` is a tuple of length
+        3 containing the labels of the two mixed examples and the fraction of
+        each example kept from the original, ``1 - cut_t_len / n_times``.
+
+    References
+    ----------
+    .. [1] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun,
+        Junsuk Choe, Youngjoon Yoo (2019). CutMix: Regularization Strategy
+        to Train Strong Classifiers with Localizable Features. In 2019
+        IEEE/CVF International Conference on Computer Vision (ICCV)
+        Online: https://arxiv.org/abs/1905.04899
+    """
+    n_times = X.shape[-1]
+    t = torch.arange(n_times, device=X.device)
+    start = cut_t_start[:, None]
+    in_cut = (t >= start) & (t < start + cut_t_len[:, None])
+    X_mix = torch.where(in_cut[:, None], X[idx_perm], X)
+    lam = (n_times - cut_t_len).to(X.dtype) / n_times
+    return X_mix, (y.clone(), y[idx_perm].clone(), lam)
+
+
 def segmentation_reconstruction(
     X: torch.Tensor,
     y: torch.Tensor,

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/augmentation/transforms.py

@@ -3,6 +3,7 @@
 #          Gustavo Rodrigues <gustavenrique01@gmail.com>
 #          Bruna Lopes <brunajaflopes@gmail.com>
 #          Sarthak Tayal <sarthaktayal2@gmail.com>
+#          Anton Soloviev <anton@praviel.com>
 #
 # License: BSD (3-clause)
 
@@ -17,6 +18,7 @@ from braindecode.util import resolve_montage_name
 
 from .base import Transform
 from .functional import (
+    _max_float_dtype,
     amplitude_scale,
     band_rotation,
     bandstop_filter,
@@ -24,6 +26,7 @@ from .functional import (
     channels_permute,
     channels_rereference,
     channels_shuffle,
+    cutmix,
     frequency_shift,
     ft_surrogate,
     gaussian_noise,
@@ -492,6 +495,7 @@ class SmoothTimeMask(Transform):
                 size=X.shape[0],
             ),
             device=X.device,
+            dtype=_max_float_dtype(X.device),
         ) * (seq_length - mask_len_samples)
         return {
             "mask_start_per_sample": mask_start,
@@ -673,7 +677,11 @@ class FrequencyShift(Transform):
             return super().get_augmentation_params(*batch)
         X = batch[0]
 
-        u = torch.as_tensor(self.rng.uniform(size=X.shape[0]), device=X.device)
+        u = torch.as_tensor(
+            self.rng.uniform(size=X.shape[0]),
+            device=X.device,
+            dtype=_max_float_dtype(X.device),
+        )
         max_delta_freq = self.max_delta_freq
         if isinstance(max_delta_freq, torch.Tensor):
             max_delta_freq = max_delta_freq.to(X.device)
@@ -831,7 +839,10 @@ class SensorsRotation(Transform):
         if isinstance(max_degrees, torch.Tensor):
             max_degrees = max_degrees.to(X.device)
         random_angles = (
-            torch.as_tensor(u, device=X.device) * 2 * max_degrees - max_degrees
+            torch.as_tensor(u, device=X.device, dtype=_max_float_dtype(X.device))
+            * 2
+            * max_degrees
+            - max_degrees
         )
         return {
             "sensors_positions_matrix": self.sensors_positions_matrix,
@@ -1095,6 +1106,55 @@ class Mixup(Transform):
         }
 
 
+class CutMix(Mixup):
+    """Implements CutMix for EEG data.
+
+    A random time region of each example is replaced by the same region of
+    a shuffled example of the batch, see [1]_. As in the reference code [2]_,
+    a coefficient ``lam`` is drawn from a Beta distribution, the region is
+    ``(1 - lam) * n_times`` long (the 1-D analogue of the cut area ratio
+    ``1 - lam``) around a uniformly drawn center, it is clipped at the trial
+    boundaries, and ``lam`` is then recomputed from the pasted length.
+
+    Like :class:`Mixup`, it applies to whole batches and returns targets
+    ``(y_a, y_b, lam)`` for :func:`braindecode.training.mixup_criterion`.
+
+    Parameters
+    ----------
+    alpha : float
+        Parameter of the Beta distribution of ``lam``. ``alpha <= 0`` cuts
+        nothing.
+    beta_per_sample : bool (default=False)
+        By default, one ``lam`` per batch is drawn. If True, one per sample.
+    random_state : int | numpy.random.Generator, optional
+        Seed to be used to instantiate numpy random number generator instance.
+        Defaults to None.
+
+    References
+    ----------
+    .. [1] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun,
+       Junsuk Choe, Youngjoon Yoo (2019). CutMix: Regularization Strategy
+       to Train Strong Classifiers with Localizable Features. In 2019
+       IEEE/CVF International Conference on Computer Vision (ICCV)
+       Online: https://arxiv.org/abs/1905.04899
+    .. [2] https://github.com/clovaai/CutMix-PyTorch/blob/master/train.py
+    """
+
+    operation = staticmethod(cutmix)  # type: ignore[assignment]
+
+    def get_augmentation_params(self, *batch):
+        """Return idx_perm and the cut_t_start, cut_t_len of each region."""
+        params = super().get_augmentation_params(*batch)
+        lam = params.pop("lam")
+        n_times = batch[0].shape[-1]
+        half = ((1 - lam) * n_times).long() // 2
+        center = torch.as_tensor(self.rng.randint(n_times, size=len(lam)))
+        center = center.to(lam.device)
+        start = (center - half).clamp(min=0)
+        end = (center + half).clamp(max=n_times)
+        return dict(params, cut_t_start=start, cut_t_len=end - start)
+
+
 class SegmentationReconstruction(Transform):
     """Segmentation Reconstruction from Lotte (2015) [Lotte2015]_.
 
@@ -1283,7 +1343,7 @@ class MaskEncoding(Transform):
         )
 
         time_start = self.rng.randint(
-            0, n_times - segment_length, (batch_size, self.n_segments)
+            0, n_times - segment_length + 1, (batch_size, self.n_segments)
         )
         time_start = torch.from_numpy(time_start)
 

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/datasets/base.py

@@ -634,7 +634,9 @@ class RecordDataset(Dataset[tuple[np.ndarray, int | str, tuple[int, int, int]]])
                         f"'{key}' already in description. Please "
                         f"rename or set overwrite to True."
                     )
-                    self._description.pop(key)
+                    # drop on a copy: windows datasets share the description
+                    # Series of the dataset they were created from
+                    self._description = self._description.drop(key)
             self._description = pd.concat([self.description, description])
 
     @property
@@ -1900,6 +1902,11 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
             dataset description.
         """
         description = pd.DataFrame(description)
+        if len(description.columns) and len(description) != len(self.datasets):
+            raise ValueError(
+                f"Length of values ({len(description)}) does not match "
+                f"number of datasets ({len(self.datasets)})."
+            )
         for key, value in description.items():
             for ds, value_ in zip(self.datasets, value):
                 ds.set_description({key: value_}, overwrite=overwrite)

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/datasets/mne.py

@@ -159,6 +159,11 @@ def create_from_mne_epochs(
 
     list_of_windows_ds = []
     for i, epochs in enumerate(list_of_epochs):
+        # drop rejected epochs first (on a copy) so that iterating over the
+        # epochs stays aligned with ``epochs.events``; preloaded epochs have
+        # already dropped them, so skip copying their data
+        if not epochs.preload:
+            epochs = epochs.copy().drop_bad()
         event_descriptions = epochs.events[:, 2]
         original_trial_starts = epochs.events[:, 0]
         stop = len(epochs.times) - window_size_samples

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/datasets/tuh.py

@@ -122,7 +122,8 @@ class TUH(BaseConcatDataset):
                 # Assume it is an integer specifying number
                 # of recordings to load
                 recording_ids = range(recording_ids)
-            descriptions = descriptions[recording_ids]
+            # select by chronological position, not by (file listing) label
+            descriptions = descriptions.iloc[:, list(recording_ids)]
 
         # workaround to ensure warnings are suppressed when running in parallel
         def create_dataset(*args, **kwargs):

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/datautil/serialization.py

@@ -140,15 +140,19 @@ def _load_signals(fif_file, preload, is_raw):
         with open(pkl_file, "rb") as f:
             signals = pickle.load(f)
 
-        if all(Path(f).exists() for f in signals.filenames):
+        # The pickle stores absolute paths: check that they exist and point to
+        # this fif file, not to the folder the dataset was copied from.
+        if all(Path(f).exists() for f in signals.filenames) and (
+            Path(signals.filenames[0]).resolve() == fif_file.resolve()
+        ):
             if preload:
                 signals.load_data()
             return _compact_fif_tags(signals)
         else:  # This may happen if the file has been moved together with the pickle file.
             warnings.warn(
                 f"Pickle file {pkl_file} exists, but the referenced fif "
-                "file(s) do not exist. Will read the fif file(s) directly "
-                "and re-create the pickle file.",
+                "file(s) do not exist or are not next to it. Will read the "
+                "fif file(s) directly and re-create the pickle file.",
                 UserWarning,
             )
 
@@ -220,7 +224,10 @@ def load_concat_dataset(path, preload, ids_to_load=None, target_name=None, n_job
     # target_name.json, raw_preproc_kwargs.json, window_kwargs.json,
     # window_preproc_kwargs.json
     if ids_to_load is None:
-        ids_to_load = [p.name for p in path.iterdir()]
+        # skip stray entries such as .DS_Store or a README next to the datasets
+        ids_to_load = [
+            p.name for p in path.iterdir() if p.is_dir() and p.name.isdigit()
+        ]
         ids_to_load = sorted(ids_to_load, key=lambda i: int(i))
     ids_to_load = [str(i) for i in ids_to_load]
     first_raw_fif_path = path / ids_to_load[0] / f"{ids_to_load[0]}-raw.fif"

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/models/__init__.py

@@ -22,6 +22,7 @@ from .deep4 import Deep4Net
 from .deepsleepnet import DeepSleepNet
 from .dgcnn import DGCNN
 from .diver1 import DIVER1
+from .duin import DuIN
 from .eeg_clip import EEGCLIP
 from .eegconformer import EEGConformer
 from .eegdino import EEGDINO
@@ -119,6 +120,7 @@ __all__ = [
     "Deep4Net",
     "DeepSleepNet",
     "DIVER1",
+    "DuIN",
     "BrainModule",
     "Brant",
     "EEGCLIP",

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/models/barista.py

@@ -14,6 +14,7 @@ from braindecode.models.base import EEGModuleMixin
 from braindecode.models.util import extract_channel_locations_from_chs_info
 from braindecode.modules import GatedLinearUnit, PatchTokenizer
 from braindecode.modules.attention import _scaled_dot_product_attention
+from braindecode.modules.blocks import _check_on_non_divisible, _patch_count
 
 
 class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
@@ -77,7 +78,8 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
     .. rubric:: Temporal, Spatial, and Spectral Encoding
 
     - Temporal encoding uses non-overlapping patches and patch-index rotary
-      embeddings. Samples beyond the last complete patch are dropped.
+      embeddings. By default a trailing partial patch is right zero-padded;
+      ``on_non_divisible="crop"`` drops it, as the reference does.
     - Spatial encoding uses dataset-provided coordinate or region indices.
       ``spatial_scale="none"`` disables it.
     - The CNN operates on waveforms. There is no explicit spectral transform.
@@ -94,7 +96,8 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
     Treebank's coordinate convention.
 
     With ``pooling="mean"``, channel counts and window lengths can vary between
-    batches, provided each window contains a full patch. Pass each recording's
+    batches (with ``on_non_divisible="crop"``, provided each window contains a
+    full patch). Pass each recording's
     indices to :meth:`forward`; otherwise it uses the constructor's indices
     or MNE positions. All samples in one batch
     share a montage. The encoder uses PyTorch attention on separate batch
@@ -141,9 +144,13 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
         in millimetres when deriving indices from ``chs_info``.
     patch_size : int
         Number of samples per temporal patch, default 512 (250 ms at the
-        paper's 2048 Hz). Windows are tokenized into whole patches, so a window
-        that is not a multiple of ``patch_size`` loses its trailing samples, as
-        in the reference.
+        paper's 2048 Hz).
+    on_non_divisible : {"pad", "crop", "error"}, default="pad"
+        How :class:`~braindecode.modules.PatchTokenizer` treats a time axis that
+        is not a multiple of ``patch_size`` (right zero-pad, crop, or raise).
+        The reference implementation crops (``"crop"``: the trailing samples
+        are lost); both give the same model and output when ``n_times`` is a
+        multiple of ``patch_size``.
     d_model : int
         Token embedding dimension.
     n_layers : int
@@ -180,6 +187,9 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
        Systems 38. https://arxiv.org/abs/2512.12135
     """
 
+    # Configs and checkpoints saved before ``on_non_divisible`` existed cropped.
+    _legacy_on_non_divisible = "crop"
+
     def __init__(
         self,
         # --- signal-related (handled by EEGModuleMixin) ---
@@ -195,6 +205,7 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
         spatial_indices: list[int] | list[list[int]] | None = None,
         coord_bins: int = 200,
         patch_size: int = 512,
+        on_non_divisible: str = "pad",
         d_model: int = 64,
         n_layers: int = 12,
         num_heads: int = 4,
@@ -243,8 +254,7 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
                 raise ValueError(f"{name} must be positive, got {value}.")
         if cnn_depth < 0:
             raise ValueError(f"cnn_depth must be non-negative, got {cnn_depth}.")
-        if self.n_times < patch_size:
-            raise ValueError("n_times must contain at least one full patch.")
+        _check_on_non_divisible(on_non_divisible)
         if d_model % num_heads:
             raise ValueError(
                 f"d_model ({d_model}) must be divisible by num_heads ({num_heads})."
@@ -258,6 +268,7 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
         self.spatial_scale = spatial_scale
         self.coord_bins = coord_bins
         self.patch_size = patch_size
+        self.on_non_divisible = on_non_divisible
         self.d_model = d_model
         self.n_layers = n_layers
         self.num_heads = num_heads
@@ -266,14 +277,15 @@ class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
         self.drop_prob = drop_prob
         self.activation = activation
 
-        # The trailing samples are dropped when n_times is not a multiple of
-        # patch_size, as in the reference.
-        self.n_patches = self.n_times // patch_size
+        # ceil(n_times / patch_size) with "pad"; floor with "crop" (the
+        # reference: trailing samples dropped, raises below one patch); exact
+        # with "error" (raises otherwise). The learned pooling is sized to it.
+        self.n_patches = _patch_count(self.n_times, patch_size, on_non_divisible)
 
         self.patch_tokenizer = PatchTokenizer(
             patch_size=patch_size,
             n_times=self.n_times,
-            on_non_divisible="crop",
+            on_non_divisible=on_non_divisible,
             output_order="patch_channel",
         )
         # Every (patch, channel) pair is encoded independently, so the grid is

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/models/base.py

@@ -521,6 +521,11 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
         output_shape : tuple[int, ...]
             shape of the network output for `batch_size==1` (1, ...)
         """
+        # Run in eval mode, so that the probe input does not update BatchNorm
+        # running stats or draw dropout masks; restore each module's own mode
+        # afterwards (some may have been put in eval mode on purpose).
+        modes = {module: module.training for module in self.modules()}
+        self.eval()
         with torch.inference_mode():
             try:
                 return tuple(
@@ -540,6 +545,9 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
                 if error is None:
                     raise
                 raise error from exc
+            finally:
+                for module, training in modes.items():
+                    module.training = training
 
     def get_config(self) -> dict:
         """Return a JSON-serializable dict of all ``__init__`` parameters.
@@ -608,9 +616,28 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
         """
         config = dict(config)  # shallow copy
         config.pop("braindecode_version", None)
+        cls._fill_legacy_defaults(config)
         resolve_type_kwargs(cls, config)
         return cls(**config)
 
+    @classmethod
+    def _fill_legacy_defaults(cls, kwargs: dict) -> None:
+        """Rebuild configs saved before ``on_non_divisible`` existed as they were.
+
+        Models that used to crop a time axis not divisible by their patch size
+        set ``_legacy_on_non_divisible = "crop"``. A saved config or checkpoint
+        without the ``on_non_divisible`` key then loads with ``"crop"``, so its
+        head and position tables keep their shapes; a new model defaults to
+        ``"pad"``.
+        """
+        legacy = getattr(cls, "_legacy_on_non_divisible", None)
+        if legacy is None or "on_non_divisible" in kwargs:
+            return
+        # A subclass with its own signature may not take the argument.
+        params = inspect.signature(cls.__init__).parameters.values()
+        if any(p.name == "on_non_divisible" or p.kind is p.VAR_KEYWORD for p in params):
+            kwargs["on_non_divisible"] = legacy
+
     def _set_n_outputs(self, n_outputs: int) -> None:
         """Record a new ``n_outputs`` on the model and its saved configs.
 
@@ -911,6 +938,9 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
         ):
             model_kwargs.pop("braindecode_version", None)
             filename = model_kwargs.pop("filename", None)
+            # The Hub mixin passes only the keys of config.json (plus the
+            # caller's), so a missing key means a pre-on_non_divisible save.
+            cls._fill_legacy_defaults(model_kwargs)
             resolve_type_kwargs(cls, model_kwargs)
 
             # Read saved n_outputs from config.json to detect when the

{braindecode-1.8.1.dev180679353 → braindecode-1.8.1.dev180761562}/braindecode/models/brant.py

@@ -13,6 +13,7 @@ from einops.layers.torch import Rearrange, Reduce
 from braindecode.functional import spectral_input
 from braindecode.models.base import EEGModuleMixin
 from braindecode.modules import PatchTokenizer
+from braindecode.modules.blocks import _check_on_non_divisible, _patch_count
 
 # Standard rhythmic-activity bands (Hz) used by Brant's frequency encoding,
 # from the paper (§ Frequency encoding): theta, alpha, beta, gamma1-5.
@@ -96,6 +97,8 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
     ``return_features=True`` returns ``cls_token=None``. The learned temporal
     positions require the runtime signal length to equal the configured
     ``n_times``, and the input must have the configured ``n_chans`` channels.
+    A window that is not a multiple of ``patch_size`` is right zero-padded to
+    ``ceil(n_times / patch_size)`` patches by default (``on_non_divisible``).
 
     The upstream model operates on signals down-sampled to **250 Hz**, but its
     band-power features are computed at a fixed 256 Hz (upstream
@@ -127,6 +130,11 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
     patch_size : int, optional
         Number of time samples per patch fed to the encoders. Default 250
         (1 s at 250 Hz). The paper's large architecture uses 1500 (see above).
+    on_non_divisible : {"pad", "crop", "error"}, default="pad"
+        How :class:`~braindecode.modules.PatchTokenizer` treats a time axis that
+        is not a multiple of ``patch_size`` (right zero-pad, crop, or raise).
+        The reference implementation crops (``"crop"``); both give the same
+        model and output when ``n_times`` is a multiple of ``patch_size``.
     embed_dim : int, optional
         Model width ``D`` (patch embedding size). Default 256.
     ffn_dim : int, optional
@@ -156,6 +164,9 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
        NeurIPS. Code: https://github.com/yzz673/Brant (Apache-2.0).
     """
 
+    # Configs and checkpoints saved before ``on_non_divisible`` existed cropped.
+    _legacy_on_non_divisible = "crop"
+
     def __init__(
         self,
         # braindecode parameters
@@ -168,6 +179,7 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
         *,
         # model-specific parameters
         patch_size: int = 250,
+        on_non_divisible: str = "pad",
         embed_dim: int = 256,
         ffn_dim: int = 384,
         temporal_n_layers: int = 4,
@@ -192,6 +204,7 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
         del n_outputs, n_chans, chs_info, n_times, input_window_seconds, sfreq
 
         self.patch_size = patch_size
+        self.on_non_divisible = on_non_divisible
         self.embed_dim = embed_dim
         self.ffn_dim = ffn_dim
         self.temporal_n_layers = temporal_n_layers
@@ -207,12 +220,10 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
             )
         # Number of patches per channel, fixed by the input length. The learnable
         # temporal positional encoding is sized to it, hence n_times is required.
-        self.seq_len = self.n_times // self.patch_size
-        if self.seq_len < 1:
-            raise ValueError(
-                f"n_times ({self.n_times}) must be >= patch_size "
-                f"({self.patch_size}) to form at least one patch."
-            )
+        # ceil for "pad", floor for "crop" (raises below one patch), exact for
+        # "error" (raises otherwise): the count the tokenizer produces.
+        _check_on_non_divisible(on_non_divisible)
+        self.seq_len = _patch_count(self.n_times, self.patch_size, on_non_divisible)
 
         # Shared non-overlapping patching (same tokenizer as the other
         # transformer foundation models); non-learnable, so it is a pure reshape
@@ -222,7 +233,7 @@ class Brant(EEGModuleMixin, nn.Module, license="apache-2.0"):
             patch_size=self.patch_size,
             n_times=self.n_times,
             learnable=False,
-            on_non_divisible="crop",
+            on_non_divisible=on_non_divisible,
         )
         # braindecode-native: band-power computed inside forward (see module).
         self.band_power = _BandPowerFeatures(