braindecode 1.8.1.dev178298903__tar.gz → 1.8.1.dev178305276__tar.gz

{braindecode-1.8.1.dev178298903 → braindecode-1.8.1.dev178305276}/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`
@@ -54,6 +55,8 @@ The following components are licensed under the MIT License:
 - `docs/_static/model/seizure_transformer_arch.png`
 - `braindecode/models/guetschel2026.py`
 - `braindecode/models/tmsanet.py`
+- `braindecode/models/epint.py`
+- `braindecode/models/gcnsnet.py`
 
 The `brainomni.py` file is a derivative of `OpenTSLab/BrainOmni` (MIT), whose
 SEANet codec derives from Meta's `facebookresearch/encodec` (MIT). The

{braindecode-1.8.1.dev178298903/braindecode.egg-info → braindecode-1.8.1.dev178305276}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev178298903
+Version: 1.8.1.dev178305276
 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.dev178298903 → braindecode-1.8.1.dev178305276}/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.dev178298903 → braindecode-1.8.1.dev178305276}/braindecode/augmentation/functional.py

@@ -580,6 +580,7 @@ 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)
@@ -1075,6 +1076,54 @@ 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.dev178298903 → braindecode-1.8.1.dev178305276}/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)
 
@@ -24,6 +25,7 @@ from .functional import (
     channels_permute,
     channels_rereference,
     channels_shuffle,
+    cutmix,
     frequency_shift,
     ft_surrogate,
     gaussian_noise,
@@ -1095,6 +1097,55 @@ 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 +1334,7 @@ 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.dev178298903 → braindecode-1.8.1.dev178305276}/braindecode/datasets/base.py

@@ -1281,6 +1281,49 @@ class WindowsDataset(_ZarrMixin, RecordDataset):
         )
 
 
+def _lazy_source_files(ds) -> list:
+    """Return the files ``ds`` still reads its signal from (none once preloaded)."""
+    signal = ds.raw if hasattr(ds, "raw") else getattr(ds, "windows", None)
+    if signal is None or getattr(signal, "preload", True):
+        return []
+    files = list(
+        getattr(signal, "filenames", None) or [getattr(signal, "filename", None)]
+    )
+    raw = getattr(signal, "_raw", None)
+    if isinstance(raw, mne.io.BaseRaw) and not raw.preload:
+        files += raw.filenames  # mne.Epochs of a lazily read Raw read through it
+    return [f for f in files if f is not None]
+
+
+def _dir_key(directory):
+    """Identify a directory by device and inode, whatever its spelling."""
+    stat = os.stat(directory)
+    return stat.st_dev, stat.st_ino
+
+
+def _overwrite_conflicts(datasets, path, offset):
+    """Yield ``(i_ds, i_target, target_dir)`` for each dataset ``i_ds`` that
+    lazily reads from inside ``target_dir = path/<i_target + offset>``, which
+    ``save(overwrite=True)`` deletes before writing dataset ``i_target``.
+    """
+    targets = {}
+    for i in range(len(datasets)):
+        target_dir = os.path.join(path, str(i + offset))
+        if os.path.isdir(target_dir):
+            targets[_dir_key(target_dir)] = (i, target_dir)
+    if not targets:
+        return
+    for i_ds, ds in enumerate(datasets):
+        for file in _lazy_source_files(ds):
+            # ancestors of the name and of the file it links to (DataLad)
+            parents = {
+                *Path(os.path.abspath(file)).parents,
+                *Path(os.path.realpath(file)).parents,
+            }
+            for key in {_dir_key(p) for p in parents if p.is_dir()} & targets.keys():
+                yield (i_ds, *targets[key])
+
+
 @register_dataset
 class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
     """A base class for concatenated datasets.
@@ -1518,7 +1561,9 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
         metadata : pd.DataFrame
             DataFrame containing as many rows as there are windows in the
             BaseConcatDataset, with the metadata and description information
-            for each window.
+            for each window, and an ``i_dataset`` column giving the position
+            in ``datasets`` of the dataset (recording) each window comes from.
+            Samplers use it to keep recordings apart.
         """
         if not all(
             [
@@ -1544,9 +1589,15 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
                     "Dataset descriptions cannot contain the reserved window "
                     "metadata key 'i_trial_in_dataset'."
                 )
+            if "i_dataset" in df.columns or "i_dataset" in ds.description:
+                raise ValueError(
+                    "Window metadata and dataset descriptions cannot contain the "
+                    "key 'i_dataset', which get_metadata() reserves for the "
+                    "position of the dataset each window comes from. Rename it."
+                )
 
         all_dfs = list()
-        for ds in self.datasets:
+        for i_ds, ds in enumerate(self.datasets):
             if hasattr(ds, "_windows") and ds._windows is not None:
                 df = ds._windows.metadata
             else:
@@ -1556,6 +1607,7 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
                 df = df.copy()
             for k, v in ds.description.items():
                 df[k] = v
+            df["i_dataset"] = i_ds
             all_dfs.append(df)
 
         return pd.concat(all_dfs)
@@ -1848,6 +1900,11 @@ 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)
@@ -1878,7 +1935,9 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
              -raw.fif | -epo.fif and .json files to.
         overwrite : bool
             Whether to delete old subdirectories that will be saved to in this
-            call.
+            call. If a dataset still reads its data lazily from one of them, a
+            ``ValueError`` is raised before anything is deleted: load the data
+            into memory first (``preload=True``) or save to another directory.
         offset : int
             If provided, the integer is added to the id of the dataset in the
             concat. This is useful in the setting of very large datasets, where
@@ -1891,6 +1950,13 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
             hasattr(self.datasets[0], "raw") or hasattr(self.datasets[0], "windows")
         ):
             raise ValueError("dataset should have either raw or windows attribute")
+        if overwrite:
+            for _, _, target_dir in _overwrite_conflicts(self.datasets, path, offset):
+                raise ValueError(
+                    f"Saving with overwrite=True would delete {target_dir} while "
+                    "a dataset still reads its data from it. Load the data into "
+                    "memory first (preload=True) or save to another directory."
+                )
 
         # Create path if it doesn't exist
         os.makedirs(path, exist_ok=True)

{braindecode-1.8.1.dev178298903 → braindecode-1.8.1.dev178305276}/braindecode/eegneuralnet.py

@@ -201,12 +201,17 @@ class _EEGNeuralNet(NeuralNet, abc.ABC):
         if classes is None:
             classes = getattr(self, "classes", None)
         signal_kwargs = infer_signal_properties(X, y, mode=self.mode, classes=classes)
+        module = _get_model(self.module)
+        # Model-specific values computed from the data (e.g. GCNsNet's graph);
+        # the signal properties take precedence.
+        infer_model_kwargs = getattr(module, "_infer_model_kwargs", None)
+        if infer_model_kwargs is not None:
+            signal_kwargs = {**infer_model_kwargs(X, y), **signal_kwargs}
         if not signal_kwargs:
             return
 
         # kick out missing kwargs:
         module_kwargs = dict()
-        module = _get_model(self.module)
         all_module_kwargs = inspect.signature(module.__init__).parameters.keys()
         for k, v in signal_kwargs.items():
             if v is None:

{braindecode-1.8.1.dev178298903 → braindecode-1.8.1.dev178305276}/braindecode/functional/_real_dft.py

@@ -12,8 +12,10 @@ never downcast, so callers wrap the numerically sensitive parts with
 #
 # License: BSD (3-clause)
 import functools
+from typing import Optional
 
 import torch
+from torch import nn
 
 
 def needs_real_dft(x: torch.Tensor) -> bool:
@@ -136,3 +138,56 @@ def hilbert_freq_real(x: torch.Tensor, forward_fourier: bool) -> torch.Tensor:
         return torch.stack((out_real, out_imag), dim=-1)
 
     return compute(x)
+
+
+@torch.jit.unused
+def fft_conv1d_real(
+    x: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor]
+) -> torch.Tensor:
+    """Real-valued equivalent of :func:`braindecode.functional.fft_conv1d`.
+
+    Overlap-save: the input is cut into blocks of ``n`` samples (the power of
+    two ``>= 2 * kernel_size``) that advance by ``n - kernel_size + 1`` outputs,
+    each block's DFT is a matmul against the cosine/sine basis, and per
+    frequency the complex product with the kernels, summed over the input
+    channels, is one real matmul.
+    """
+
+    @fp32_island
+    def compute(x, weight):
+        n_times, kernel_size = x.shape[-1], weight.shape[-1]
+        n = 1 << (2 * kernel_size - 1).bit_length()
+        hop = n - kernel_size + 1
+        n_blocks = -(-n_times // hop)
+        n_freqs = n // 2 + 1
+        # "same" alignment: torch pads the extra zero of an even kernel on the right
+        left = (kernel_size - 1) // 2
+        x = nn.functional.pad(x, (left, (n_blocks + 1) * hop - n_times - left))
+        hops = x.unflatten(-1, (n_blocks + 1, hop))
+        # block i = hop i + the first kernel_size - 1 samples of hop i + 1
+        blocks = torch.cat([hops[:, :, :-1], hops[:, :, 1:, : kernel_size - 1]], -1)
+        cosine, sine = real_dft_basis(n, device=x.device, dtype=x.dtype)
+        basis = torch.cat([cosine, -sine]).transpose(0, 1)  # -> [Re | Im] of rfft
+        x_f = blocks @ basis  # (batch, in, blocks, 2 * n_freqs)
+        w_f = nn.functional.pad(weight.to(x), (0, n - kernel_size)) @ basis
+        w_re, w_im = w_f[..., :n_freqs], w_f[..., n_freqs:]
+        # conv1d is a cross-correlation: x_f * conj(w_f), summed over the input
+        # channels. Per frequency [Re x, Im x] @ [[Re w, -Im w], [Im w, Re w]].
+        w_f = torch.cat([torch.cat([w_re, -w_im]), torch.cat([w_im, w_re])], 1)
+        x_f = x_f.unflatten(-1, (2, n_freqs)).permute(4, 0, 2, 3, 1).flatten(1, 2)
+        y_f = torch.bmm(x_f.flatten(2), w_f.permute(2, 1, 0))  # (freq, ., 2 * out)
+        y_f = y_f.unflatten(1, (x.shape[0], n_blocks)).unflatten(-1, (2, -1))
+        y_f = y_f.permute(1, 4, 2, 3, 0).flatten(
+            -2
+        )  # (batch, out, blocks, 2 * n_freqs)
+        # inverse DFT of the first `hop` samples, the ones without wrap-around
+        scale = torch.full((n_freqs, 1), 2.0 / n, dtype=x.dtype, device=x.device)
+        scale[0] = 1.0 / n
+        scale[-1] = 1.0 / n  # Nyquist (n is even)
+        inverse = torch.cat([scale * cosine[:, :hop], -scale * sine[:, :hop]])
+        return (y_f @ inverse).flatten(-2)[..., :n_times]
+
+    y = compute(x, weight)
+    if bias is not None:
+        y = y + bias.to(y).unsqueeze(-1)
+    return y.to(x)

{braindecode-1.8.1.dev178298903 → braindecode-1.8.1.dev178305276}/braindecode/functional/functions.py

@@ -62,9 +62,10 @@ def fft_conv1d(
 
     Equal to the direct convolution up to float rounding: the FFT is long
     enough for a linear (not circular) convolution. Its cost barely grows with
-    the kernel size, so it is faster than the direct one for long kernels on
-    CPU (see :func:`prefer_fft_conv`). float16/bfloat16 inputs are computed in
-    float32 (:func:`spectral_input`) and cast back.
+    the kernel size, so it is faster than the direct one for long kernels
+    (see :func:`prefer_fft_conv`). float16/bfloat16 inputs are computed in
+    float32 (:func:`spectral_input`) and cast back. On Intel Gaudi (HPU, no
+    complex dtype) the DFTs are real matmuls over overlapping blocks.
 
     Parameters
     ----------
@@ -82,22 +83,35 @@ def fft_conv1d(
         ``x``. For an even ``kernel_size`` it is aligned like torch's
         ``padding="same"`` (the extra zero padded on the right).
     """
+    if _real_dft.needs_real_dft(x):
+        return _real_dft.fft_conv1d_real(x, weight, bias)
     n_times, kernel_size = x.shape[-1], weight.shape[-1]
     # >= n_times + kernel_size - 1 (no circular wrap-around), even and
     # 5-smooth: the fast cases of a real FFT
     n_fft = 2 * _next_fast_len((n_times + kernel_size) // 2)
     xs = spectral_input(x)
     # conv1d is a cross-correlation: a convolution with the flipped kernel
-    x_f = torch.fft.rfft(xs, n=n_fft).permute(2, 0, 1)  # (freq, batch, in)
+    x_f = torch.fft.rfft(xs, n=n_fft)  # (batch, in, freq)
     w_f = torch.fft.rfft(weight.to(xs).flip(-1), n=n_fft).permute(2, 1, 0)
     # Per frequency (batch, in) @ (in, out) in complex numbers, as one real
-    # matmul (a complex one is slower on CPU): [Re x; Im x] @ [Re w, Im w]
-    # (columns interleaved) holds the four real products.
-    p = torch.cat([x_f.real, x_f.imag], 1) @ torch.view_as_real(w_f).flatten(2)
-    p = p.unflatten(1, [2, x.shape[0]]).unflatten(3, [-1, 2])  # x re/im, w re/im
-    y_f = torch.complex(
-        p[:, 0, :, :, 0] - p[:, 1, :, :, 1], p[:, 0, :, :, 1] + p[:, 1, :, :, 0]
-    )
+    # matmul (a complex one is slower on CPU).
+    needs_grad = torch.is_grad_enabled() and (x.requires_grad or weight.requires_grad)
+    if needs_grad or x.shape[0] >= 2 * x.shape[1]:
+        # [Re x, Im x] @ [[Re w, Im w], [-Im w, Re w]] (rows of i * w): the
+        # output columns are the (Re, Im) pairs of the result, read as complex
+        # without a copy (and without a costly backward).
+        w_mat = torch.view_as_real(torch.stack([w_f, w_f * 1j], 1)).flatten(1, 2)
+        y_f = torch.view_as_real(x_f).permute(2, 0, 3, 1).flatten(2) @ w_mat.flatten(2)
+        y_f = torch.view_as_complex(y_f.unflatten(-1, [-1, 2]))
+    else:
+        # Forward only with fewer rows than twice the input channels: the
+        # doubled weight costs more than combining [Re x; Im x] @ [Re w, Im w].
+        x_f = x_f.permute(2, 0, 1)
+        p = torch.cat([x_f.real, x_f.imag], 1) @ torch.view_as_real(w_f).flatten(2)
+        p = p.unflatten(1, [2, x.shape[0]]).unflatten(3, [-1, 2])
+        y_f = torch.complex(
+            p[:, 0, :, :, 0] - p[:, 1, :, :, 1], p[:, 0, :, :, 1] + p[:, 1, :, :, 0]
+        )
     y = torch.fft.irfft(y_f.permute(1, 2, 0), n=n_fft)
     # "same" output t is sample t + kernel_size // 2 of the full convolution
     y = y[..., kernel_size // 2 : kernel_size // 2 + n_times]
@@ -115,8 +129,11 @@ def prefer_fft_conv(x: torch.Tensor, kernel_size: int) -> bool:
     and from ``batch_size * kernel_size >= 600`` if also
     ``in_channels * kernel_size >= 120``: with one or two input channels a
     float32 direct convolution stays cheaper up to k = 85 / 31 (13 long
-    temporal convolutions of the model zoo, 22 channels x batch 32). GPUs and
-    HPUs (no complex dtype) keep the direct convolution.
+    temporal convolutions of the model zoo, 22 channels x batch 32). On Intel
+    Gaudi (HPU) a direct convolution costs about ``batch_size * n_times *
+    kernel_size`` and the FFT about a fixed few milliseconds in eager mode: the
+    FFT wins from 8 input channels and ``batch_size * n_times * kernel_size >=
+    8e6``. GPUs keep the direct convolution.
 
     Parameters
     ----------
@@ -125,6 +142,8 @@ def prefer_fft_conv(x: torch.Tensor, kernel_size: int) -> bool:
     kernel_size : int
         Kernel length in samples.
     """
+    if x.device.type == "hpu":
+        return x.shape[1] >= 8 and x.shape[0] * x.shape[-1] * kernel_size >= 8e6
     return x.device.type == "cpu" and (
         x.element_size() == 2
         or (x.shape[0] * kernel_size >= 600 and x.shape[1] * kernel_size >= 120)

{braindecode-1.8.1.dev178298903 → braindecode-1.8.1.dev178305276}/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
@@ -36,9 +37,11 @@ from .eegsimpleconv import EEGSimpleConv
 from .eegsym import EEGSym
 from .eegtcnet import EEGTCNet
 from .emg2qwerty import EMG2QwertyNet
+from .epint import EpiNT
 from .fbcnet import FBCNet
 from .fblightconvnet import FBLightConvNet
 from .fbmsnet import FBMSNet
+from .gcnsnet import GCNsNet
 from .guetschel2026 import Guetschel2026
 from .hybrid import HybridNet
 from .ifnet import IFNet
@@ -117,6 +120,7 @@ __all__ = [
     "Deep4Net",
     "DeepSleepNet",
     "DIVER1",
+    "DuIN",
     "BrainModule",
     "Brant",
     "EEGCLIP",
@@ -134,12 +138,14 @@ __all__ = [
     "EEGTCNet",
     "DGCNN",
     "EMG2QwertyNet",
+    "EpiNT",
     "NeuroPose",
     "SensingDynamics",
     "VEMG2Pose",
     "FBCNet",
     "FBLightConvNet",
     "FBMSNet",
+    "GCNsNet",
     "Guetschel2026",
     "MetaNeuromotorHand",
     "HybridNet",

{braindecode-1.8.1.dev178298903 → braindecode-1.8.1.dev178305276}/braindecode/models/axon.py

@@ -14,6 +14,7 @@ from torch import nn
 from braindecode.functional import sinusoidal_positional_encoding
 from braindecode.models.base import EEGModuleMixin
 from braindecode.models.util import warn_if_sfreq_differs
+from braindecode.modules.attention import _scaled_dot_product_attention
 from braindecode.modules.channels import _resolve
 
 
@@ -462,11 +463,33 @@ class _AxisAttention(nn.Module):
             .unbind(0)
         )
         drop = self.att_drop_prob if self.training else 0.0
-        attended = F.scaled_dot_product_attention(
+        attended = _scaled_dot_product_attention(
             query, key, value, attn_mask=mask, dropout_p=drop
         )
         return attended.transpose(1, 2).reshape(n_seq, length, -1)
 
+    def _attend_two_scales(
+        self, projected: torch.Tensor, band: torch.Tensor, weights: torch.Tensor
+    ) -> torch.Tensor:
+        """Banded and full attention from one score matrix, mixed per query.
+
+        ``weights`` (N, L, 2) are the scale-gate weights. Both softmaxes share
+        the scores and one product with the values serves the mixture.
+        """
+        n_seq, length, _ = projected.shape
+        query, key, value = (
+            projected.reshape(n_seq, length, 3, self.num_heads, self.head_dim)
+            .permute(2, 0, 3, 1, 4)
+            .unbind(0)
+        )
+        scores = torch.matmul(query, key.transpose(-2, -1)) * self.head_dim**-0.5
+        weights = weights.reshape(n_seq, 1, length, 2)
+        probs = weights[..., 0:1] * scores.masked_fill(~band, float("-inf")).softmax(
+            -1
+        ) + weights[..., 1:2] * scores.softmax(-1)
+        attended = torch.matmul(probs, value)
+        return attended.transpose(1, 2).reshape(n_seq, length, -1)
+
     def forward(self, x: torch.Tensor, band: Optional[torch.Tensor]) -> torch.Tensor:
         batch, n_chans, n_patches, dim = x.shape
         axis_weights = self.axis_gate(x)  # (B, C, P, 2)
@@ -476,15 +499,26 @@ class _AxisAttention(nn.Module):
         # the restricted branch is the full one; attend once unless the two
         # branches draw their own attention dropout.
         per_channel = self.temporal_qkv(x.reshape(batch * n_chans, n_patches, dim))
-        restricted = self.temporal_out(self._attend(per_channel, band)).reshape(
-            batch, n_chans, n_patches, dim
-        )
-        full = restricted
-        if band is not None or (self.training and self.att_drop_prob > 0):
-            full = self.temporal_out(self._attend(per_channel, None)).reshape(
+        dropout = self.training and self.att_drop_prob > 0
+        if band is not None and not dropout:
+            # The scale-gate weights sum to one, so the two branches share the
+            # scores, the product with the values and temporal_out.
+            attended = self._attend_two_scales(per_channel, band, scale_weights)
+            temporal = self.temporal_out(attended).reshape(
                 batch, n_chans, n_patches, dim
             )
-        temporal = scale_weights[..., 0:1] * restricted + scale_weights[..., 1:2] * full
+        else:
+            restricted = self.temporal_out(self._attend(per_channel, band)).reshape(
+                batch, n_chans, n_patches, dim
+            )
+            full = restricted
+            if dropout:
+                full = self.temporal_out(self._attend(per_channel, None)).reshape(
+                    batch, n_chans, n_patches, dim
+                )
+            temporal = (
+                scale_weights[..., 0:1] * restricted + scale_weights[..., 1:2] * full
+            )
 
         # Spatial path: one sequence of C tokens per time step.
         per_step = self.spatial_qkv(