Braindecode 1.8.1.dev174671102__tar.gz → 1.8.1.dev174738705__tar.gz

{braindecode-1.8.1.dev174671102/braindecode.egg-info → braindecode-1.8.1.dev174738705}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev174671102
+Version: 1.8.1.dev174738705
 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.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/datasets/base.py

@@ -14,13 +14,16 @@ from __future__ import annotations
 import bisect
 import html as _html
 import json
+import mmap
 import os
 import shutil
 import warnings
 from abc import abstractmethod
 from collections import Counter
-from collections.abc import Callable, Hashable
+from collections.abc import Callable, Hashable, Sequence
 from glob import glob
+from itertools import chain
+from operator import attrgetter
 from pathlib import Path
 from typing import Any, Generic, Iterable, no_type_check
 
@@ -28,6 +31,7 @@ import mne
 import mne.io
 import numpy as np
 import pandas as pd
+from mne.io.constants import FIFF
 from mne.utils import _soft_import
 from mne.utils.docs import deprecated
 from torch.utils.data import ConcatDataset, Dataset, IterableDataset
@@ -220,6 +224,109 @@ def _compute_ch_pos(info, targets_from):
     return pos
 
 
+_TAG_ROW = attrgetter("kind", "type", "size", "next", "pos")
+
+
+def _tag_rows(tags):
+    """(kind, type, size, next, pos) of FIF tags as an int64 array (n, 5)."""
+    flat = chain.from_iterable(map(_TAG_ROW, tags))
+    return np.fromiter(flat, np.int64, 5 * len(tags)).reshape(-1, 5)
+
+
+class _FifTagTable(Sequence):
+    """The data-buffer tags of a lazy FIF raw as one int64 array.
+
+    mne's FIF reader keeps one ``Tag`` object per data buffer in
+    ``raw._raw_extras[i]["ent"]`` (one per second of signal with mne's default
+    buffers), each with its attribute dict and integers once unpickled. A lazy
+    dataset of many recordings holds millions of them in RAM, and every forked
+    DataLoader worker copies the memory pages of those it reads (reference
+    counts). This holds the same fields in 40 bytes per buffer and returns an
+    equal ``Tag`` on access, so mne's own reads are unchanged.
+    """
+
+    def __init__(self, tags):
+        self.tag_class = type(tags[0])
+        self.table = _tag_rows(tags)
+
+    def __len__(self):
+        return len(self.table)
+
+    def __getitem__(self, i):
+        if isinstance(i, slice):
+            return [self.tag_class(*row) for row in self.table[i].tolist()]
+        return self.tag_class(*self.table[i].tolist())
+
+
+_TAG_FIELDS = ("kind", "type", "size", "next", "pos", "data")
+
+
+def _compact_fif_tags(inst):
+    """Swap the ``Tag`` lists of a FIF raw for `_FifTagTable` (in place).
+
+    Only gap-free lists of directory tags (no ``None``, no ``data``) with mne's
+    ``Tag`` fields are swapped; anything else is left as mne built it.
+    """
+    for extra in getattr(inst, "_raw_extras", ()):
+        tags = extra.get("ent")
+        if not isinstance(tags, list) or not tags:
+            continue
+        cls = type(tags[0])
+        if tuple(getattr(cls, "__dataclass_fields__", ())) == _TAG_FIELDS and all(
+            type(t) is cls and t.data is None for t in tags
+        ):
+            extra["ent"] = _FifTagTable(tags)
+    return inst
+
+
+def _read_fif_window(raw, start, stop):
+    """Samples ``start:stop`` of a lazy FIF raw, copied from the file in one go.
+
+    mne reads, casts and calibrates the FIF data buffers of a range one at a
+    time (60 Python steps for a 60-s window saved with 1-s buffers, each through
+    a buffered file whose buffer is 4 MB on network file systems). Here one
+    slice of the memory-mapped file holds the bytes of every buffer in the
+    range and one multiply calibrates them, with mne's arithmetic (float64 data
+    times ``raw._cals``), so the values equal ``raw._getitem``'s. Returns None
+    (the caller uses mne) unless ``raw`` is a lazy, single-file, uncompressed
+    FIF raw of float32 buffers without projector, compensation or gaps.
+    """
+    if (
+        not isinstance(raw, mne.io.Raw)
+        or raw.preload
+        or len(raw._raw_extras) != 1
+        or raw._projector is not None
+        or raw._comp is not None
+    ):
+        return None
+    extra = raw._raw_extras[0]
+    fname, bounds = extra["filename"], extra["bounds"]
+    if not isinstance(fname, Path) or fname.suffix == ".gz":
+        return None
+    start, stop = start + raw._first_samps[0], stop + raw._first_samps[0]
+    first = np.searchsorted(bounds, start, side="right") - 1
+    tags = extra["ent"]
+    if isinstance(tags, _FifTagTable):
+        rows = tags.table[first : np.searchsorted(bounds, stop)]
+    else:
+        tags = tags[first : np.searchsorted(bounds, stop)]
+        if any(t is None for t in tags):
+            return None
+        rows = _tag_rows(tags)
+    if not len(rows) or (rows[:, 1] != FIFF.FIFFT_FLOAT).any():
+        return None
+    size, pos = rows[:, 2].tolist(), rows[:, 4].tolist()
+    # mmap: no copy through a 4 MB read buffer, and a cold file is read around
+    # the window by the kernel (as mne's 4 MB reads did), not window by window
+    with open(fname, "rb") as f, mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ) as m:
+        buf = m[pos[0] : pos[-1] + 16 + size[-1]]
+    one = np.concatenate(
+        [np.frombuffer(buf, ">f4", s // 4, p + 16 - pos[0]) for s, p in zip(size, pos)]
+    ).reshape(-1, extra["orig_nchan"])[start - bounds[first] : stop - bounds[first]]
+    picked = one[:, raw._read_picks[0]].T
+    return np.multiply(picked, raw._cals[:, np.newaxis], order="C")
+
+
 def _window_info(crop_inds, sfreq):
     """Extract window size from crop indices.
 
@@ -823,9 +930,11 @@ class EEGWindowsDataset(_ZarrMixin, RecordDataset):
         if self._zarr_data is not None:
             X = self._zarr_data[:, i_start:i_stop]
         else:
-            X = self.raw._getitem(
-                (slice(None), slice(i_start, i_stop)), return_times=False
-            )
+            X = _read_fif_window(self.raw, i_start, i_stop)
+            if X is None:
+                X = self.raw._getitem(
+                    (slice(None), slice(i_start, i_stop)), return_times=False
+                )
         X = np.array(X, dtype="float32")
         if self.transform is not None:
             X = self.transform(X)
@@ -1672,7 +1781,8 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
                 kwargs_path = os.path.join(path, ".".join([kwarg_name, "json"]))
                 kwargs = getattr(self, kwarg_name)
                 if kwargs is not None:
-                    json.dump(kwargs, open(kwargs_path, "w"))
+                    with open(kwargs_path, "w") as f:
+                        json.dump(kwargs, f)
 
     @property
     def description(self) -> pd.DataFrame:
@@ -1742,7 +1852,10 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
         # Create path if it doesn't exist
         os.makedirs(path, exist_ok=True)
 
-        path_contents = os.listdir(path)
+        # With an offset the caller fills one directory piece by piece (preprocess
+        # saves each recording at its index), so the other entries are expected:
+        # skip the scan that would stat them all on every call and list them in a warning.
+        path_contents = os.listdir(path) if offset == 0 else []
         n_sub_dirs = len(
             [e for e in path_contents if os.path.isdir(os.path.join(path, e))]
         )

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/datasets/tuh.py

@@ -241,7 +241,11 @@ class TUH(BaseConcatDataset):
         # file paths, use this instead as before
         if "year" in description:
             meas_date = meas_date.replace(*description[["year", "month", "day"]])
-        raw.set_meas_date(meas_date)
+        # FIF (e.g. preprocess(..., save_dir=...)) stores dates from 1901-12-13 to
+        # 2038-01-19 only: undated recordings (year 1 above) and de-identified ones
+        # (TUAB has 1899-12-30) keep their date in the description only.
+        in_fif_range = -(2**31) <= meas_date.timestamp() < 2**31
+        raw.set_meas_date(meas_date if in_fif_range else None)
 
         d = {
             "age": int(age),
@@ -251,9 +255,9 @@ class TUH(BaseConcatDataset):
         # if not, get it from meas_date in raw.info and add to description
         # if meas_date is None, create fake one
         if "year" not in description:
-            d["year"] = raw.info["meas_date"].year
-            d["month"] = raw.info["meas_date"].month
-            d["day"] = raw.info["meas_date"].day
+            d["year"] = meas_date.year
+            d["month"] = meas_date.month
+            d["day"] = meas_date.day
 
         # read info relevant for preprocessing from raw without loading it
         if add_physician_reports:

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/datautil/serialization.py

@@ -21,6 +21,7 @@ from ..datasets.base import (
     EEGWindowsDataset,
     RawDataset,
     WindowsDataset,
+    _compact_fif_tags,
 )
 
 
@@ -79,7 +80,8 @@ def _outdated_load_concat_dataset(path, preload, ids_to_load=None, target_name=N
     for path in paths:
         if is_raw and target_name is None:
             target_file_name = path / "target_name.json"
-            target_name = json.load(open(target_file_name, "r"))["target_name"]
+            with open(target_file_name) as f:
+                target_name = json.load(f)["target_name"]
 
         all_signals, description = _load_signals_and_description(
             path=path, preload=preload, is_raw=is_raw, ids_to_load=ids_to_load
@@ -141,7 +143,7 @@ def _load_signals(fif_file, preload, is_raw):
         if all(Path(f).exists() for f in signals.filenames):
             if preload:
                 signals.load_data()
-            return signals
+            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 "
@@ -171,7 +173,8 @@ def _load_signals(fif_file, preload, is_raw):
             if preload:
                 signals._data, signals.preload = data, True
 
-    return signals
+    # after the pickle above, which keeps mne's own layout on disk
+    return _compact_fif_tags(signals)
 
 
 def load_concat_dataset(path, preload, ids_to_load=None, target_name=None, n_jobs=1):
@@ -264,7 +267,8 @@ def _load_parallel(path, i, preload, is_raw, has_stored_windows):
     target_file_path = sub_dir / "target_name.json"
     target_name = None
     if target_file_path.exists():
-        target_name = json.load(open(target_file_path, "r"))["target_name"]
+        with open(target_file_path) as f:
+            target_name = json.load(f)["target_name"]
 
     if is_raw and (not has_stored_windows):
         dataset = RawDataset(signals, description, target_name)
@@ -304,8 +308,8 @@ def _load_kwargs_json(kwargs_name, sub_dir):
     kwargs_file_name = ".".join([kwargs_name, "json"])
     kwargs_file_path = os.path.join(sub_dir, kwargs_file_name)
     if os.path.exists(kwargs_file_path):
-        kwargs = json.load(open(kwargs_file_path, "r"))
-        return kwargs
+        with open(kwargs_file_path) as f:
+            return json.load(f)
 
 
 def _is_outdated_saved(path):

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/attn_sleep.py

@@ -143,6 +143,10 @@ class AttnSleep(EEGModuleMixin, nn.Module):
             sfreq=sfreq,
         )
         del n_outputs, n_chans, chs_info, n_times, input_window_seconds, sfreq
+        if self._n_chans_for_jit not in (None, 1):
+            raise ValueError(
+                f"AttnSleep takes single-channel input; got n_chans={self._n_chans_for_jit}."
+            )
 
         self.mapping = {
             "fc.weight": "final_layer.weight",

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/base.py

@@ -46,6 +46,24 @@ _GEOMETRY_KWARGS = (
 )
 
 
+def _too_short_input_error(exc: RuntimeError, input_shape) -> Optional[ValueError]:
+    """ValueError for a layer that found the input too short, else None."""
+    if not str(exc).endswith(
+        (
+            "Output size is too small",
+            "Kernel size can't be greater than actual input size",
+        )
+    ):
+        return None
+    return ValueError(
+        "During model prediction RuntimeError was thrown showing that at some "
+        f"layer `{str(exc).split('.')[-1]}` (see above in the stacktrace). This "
+        "could be caused by providing too small `n_times`/`input_window_seconds`. "
+        "Model may require longer chunks of signal in the input than "
+        f"{input_shape}."
+    )
+
+
 _HF_INSTALL_HINT = (
     "requires the `huggingface_hub` package. "
     "Install with: pip install 'braindecode[hub]'"
@@ -483,6 +501,18 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
         # Called by torch.jit.script before compiling the model.
         return make_parametrizations_scriptable(self)
 
+    def __call__(self, *args, **kwargs):
+        # An input shorter than the layers need raises ValueError, not a torch
+        # RuntimeError from deep inside forward.
+        try:
+            return super().__call__(*args, **kwargs)
+        except RuntimeError as exc:
+            shape = tuple(args[0].shape) if args and hasattr(args[0], "shape") else None
+            error = _too_short_input_error(exc, shape)
+            if error is None:
+                raise
+            raise error from exc
+
     def get_output_shape(self) -> tuple[int, ...]:
         """Returns shape of neural network output for batch size equal 1.
 
@@ -506,21 +536,10 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
                     .shape
                 )
             except RuntimeError as exc:
-                if str(exc).endswith(
-                    (
-                        "Output size is too small",
-                        "Kernel size can't be greater than actual input size",
-                    )
-                ):
-                    msg = (
-                        "During model prediction RuntimeError was thrown showing that at some "
-                        f"layer `{str(exc).split('.')[-1]}` (see above in the stacktrace). This "
-                        "could be caused by providing too small `n_times`/`input_window_seconds`. "
-                        "Model may require longer chunks of signal in the input than "
-                        f"{self.input_shape}."
-                    )
-                    raise ValueError(msg) from exc
-                raise exc
+                error = _too_short_input_error(exc, self.input_shape)
+                if error is None:
+                    raise
+                raise error from exc
 
     def get_config(self) -> dict:
         """Return a JSON-serializable dict of all ``__init__`` parameters.

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/brainomni.py

@@ -29,15 +29,21 @@ from braindecode.modules.quantization import ResidualVectorQuantizer
 
 
 class _SELU(nn.SELU):
-    """``nn.SELU`` written as ``scale * x`` / ELU: same values and gradients, and
-    trainable on Gaudi, whose ELU backward only supports ``scale == 1``."""
+    """``nn.SELU`` as ``ELU(x, alpha * scale)`` times ``scale`` where ``x > 0``:
+    same values and gradients, and trainable on Gaudi, whose ELU backward only
+    supports ``scale == 1``."""
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         alpha, scale = (
             1.6732632423543772848170429916717,
             1.0507009873554804934193349852946,
         )
-        return torch.where(x > 0, x * scale, F.elu(x, alpha * scale))
+        # ELU is x on the positive side, so this equals the where(x > 0, x * scale,
+        # elu) form bit for bit, without its zero-filled where backward. The
+        # factor is at least float32, like the scalar multiply's compute type.
+        one = x.new_ones((), dtype=torch.promote_types(x.dtype, torch.float32))
+        factor = torch.where(x > 0, one * scale, one)
+        return (F.elu(x, alpha * scale) * factor).to(x.dtype)
 
 
 class BrainTokenizer(EEGModuleMixin, nn.Module, license="mit"):
@@ -863,9 +869,13 @@ class _SEANetConv1d(nn.Module):
         extra = (kernel_size - padding_total - x.shape[-1]) % stride
         right = padding_total // 2
         padding = (padding_total - right, right + extra)
+        if max(padding) == 0:  # 1x1 convs: padding would only copy x
+            return self.conv(x)
         # Zero-extend a signal too short to reflect, as EnCodec does.
         short = max(max(padding) - x.shape[-1] + 1, 0)
-        x = F.pad(F.pad(x, (0, short)), padding, mode="reflect")
+        if short > 0:  # F.pad with nothing to pad still copies x
+            x = F.pad(x, (0, short))
+        x = F.pad(x, padding, mode="reflect")
         return self.conv(x[..., : x.shape[-1] - short])
 
 

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/csbrain.py

@@ -314,6 +314,11 @@ class CSBrain(EEGModuleMixin, nn.Module):
             sfreq=sfreq,
         )
         del n_chans, chs_info, n_times, input_window_seconds, sfreq, n_outputs
+        n_times = self._n_times_or_none()
+        if n_times is not None and n_times % patch_size:
+            raise ValueError(
+                f"n_times ({n_times}) must be a multiple of patch_size ({patch_size})."
+            )
 
         self.rearrange = Rearrange(
             "batch n_chans (n_patch patch_size) -> batch n_chans n_patch patch_size",

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/deep4.py

@@ -170,6 +170,10 @@ class Deep4Net(EEGModuleMixin, nn.Sequential):
         ):
             resolved_n_times = round(self._input_window_seconds * self._sfreq)
 
+        if resolved_n_times is not None and resolved_n_times >= min_n_times:
+            # The bound above ignores stride_before_pool and an integer
+            # final_conv_length; inputs it admits are checked with the exact one.
+            min_n_times = self._get_min_n_times(exact=True)
         if resolved_n_times is not None and resolved_n_times < min_n_times:
             scaling_factor = resolved_n_times / min_n_times
             warn(
@@ -363,13 +367,15 @@ class Deep4Net(EEGModuleMixin, nn.Sequential):
 
         self.train()
 
-    def _get_min_n_times(self) -> int:
+    def _get_min_n_times(self, exact: bool = False) -> int:
         """
         Calculate the minimum number of time samples required for the model
         to work with the given temporal parameters.
         """
         # Start with the minimum valid output length of the network (1)
         min_len = 1
+        if exact and isinstance(self.final_conv_length, int):
+            min_len = self.final_conv_length
 
         # List of conv kernel sizes and pool parameters for the 4 blocks, in reverse order
         # Each tuple: (filter_length, pool_length, pool_stride)
@@ -382,6 +388,9 @@ class Deep4Net(EEGModuleMixin, nn.Sequential):
 
         # Work backward from the last layer to the input
         for filter_len, pool_len, pool_stride in block_params:
+            if exact and self.stride_before_pool:  # strided conv, then pool stride 1
+                min_len = pool_stride * (min_len + pool_len - 2) + filter_len
+                continue
             # Reverse the pooling operation
             # L_in = stride * (L_out - 1) + kernel_size
             min_len = pool_stride * (min_len - 1) + pool_len

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/dgcnn.py

@@ -142,6 +142,11 @@ def _build_initial_adjacency(chs_info, n_chans, n_neighbors=5):
     electrode_positions = extract_channel_locations_from_chs_info(
         chs_info, num_channels=n_chans
     )
+    # MNE stores NaN for a channel without a position.
+    if electrode_positions is not None and not (
+        len(electrode_positions) == n_chans and np.isfinite(electrode_positions).all()
+    ):
+        electrode_positions = None
 
     if electrode_positions is None and chs_info is not None:
         # Try to infer positions from channel names via a standard montage
@@ -162,6 +167,8 @@ def _build_initial_adjacency(chs_info, n_chans, n_neighbors=5):
                 "with 3-D positions or that channel names match a standard montage."
             )
 
+    if n_chans < 2:
+        raise ValueError("DGCNN needs at least 2 channels to build its channel graph.")
     n_neighbors_capped = min(n_neighbors, n_chans - 1)
 
     # Pairwise Euclidean distances

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/eegconformer.py

@@ -7,7 +7,7 @@ import torch
 from einops.layers.torch import Rearrange
 from torch import Tensor, nn
 
-from braindecode.models.base import EEGModuleMixin
+from braindecode.models.base import EEGModuleMixin, _too_short_input_error
 from braindecode.modules import FeedForwardBlock, MultiHeadAttention
 
 
@@ -276,7 +276,10 @@ class EEGConformer(EEGModuleMixin, nn.Module):
         return x
 
     def get_fc_size(self):
-        out = self.patch_embedding(torch.ones((1, 1, self.n_chans, self.n_times)))
+        try:
+            out = self.patch_embedding(torch.ones((1, 1, self.n_chans, self.n_times)))
+        except RuntimeError as exc:
+            raise _too_short_input_error(exc, self.input_shape) or exc
         size_embedding_1 = out.cpu().data.numpy().shape[1]
         size_embedding_2 = out.cpu().data.numpy().shape[2]
 

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/eegnex.py

@@ -9,6 +9,7 @@ import torch
 import torch.nn as nn
 from einops.layers.torch import Rearrange
 
+from braindecode.functional import fft_conv1d, prefer_fft_conv
 from braindecode.models.base import EEGModuleMixin
 from braindecode.modules import Conv2dWithConstraint, LinearWithConstraint
 
@@ -234,7 +235,7 @@ class EEGNeX(EEGModuleMixin, nn.Module):
         )
 
         self.block_2 = nn.Sequential(
-            nn.Conv2d(
+            _TemporalConv(
                 in_channels=self.filter_1,
                 out_channels=self.filter_2,
                 kernel_size=self.kernel_block_1_2,
@@ -352,3 +353,17 @@ class EEGNeX(EEGModuleMixin, nn.Module):
             self.filter_1 * T5
         )  # filter_1 is the number of channels before flatten
         return final_in_features
+
+
+class _TemporalConv(nn.Conv2d):
+    """``(1, k)`` Conv2d with "same" output length; on CPU, through
+    :func:`~braindecode.functional.fft_conv1d` where
+    :func:`~braindecode.functional.prefer_fft_conv` expects it to be faster."""
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not prefer_fft_conv(x, self.kernel_size[1]):
+            return self._conv_forward(x, self.weight, self.bias)
+        b, _, h, t = x.shape  # the conv1d of each of the b * h rows
+        x = x.transpose(1, 2).reshape(b * h, -1, t)
+        x = fft_conv1d(x, self.weight.squeeze(2), self.bias)
+        return x.unflatten(0, [b, h]).transpose(1, 2)

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/labram.py

@@ -1416,9 +1416,7 @@ class _Attention(nn.Module):
         if self.k_norm is not None:
             k = self.k_norm(k).type_as(v)
 
-        q = q * self.scale
-        attn = q @ k.transpose(-2, -1)
-
+        attn_mask: Optional[torch.Tensor] = None
         if self.relative_position_bias_table is not None:
             relative_position_bias = self.relative_position_bias_table[
                 self.relative_position_index.view(-1)
@@ -1427,18 +1425,29 @@ class _Attention(nn.Module):
                 self.window_size[0] * self.window_size[1] + 1,
                 -1,
             )  # Wh*Ww,Wh*Ww,nH
-            relative_position_bias = relative_position_bias.permute(
-                2, 0, 1
-            ).contiguous()  # nH, Wh*Ww, Wh*Ww
-            attn = attn + relative_position_bias.unsqueeze(0)
-
-        attn = attn.softmax(dim=-1)
-        attn = self.attn_drop(attn)
-
-        if return_attention:
-            return attn
+            attn_mask = (
+                relative_position_bias.permute(2, 0, 1).contiguous().unsqueeze(0)
+            )  # 1, nH, Wh*Ww, Wh*Ww
 
-        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        # Fused kernel, same math. Not on Gaudi (Labram trains ~12 % slower there)
+        # nor with attention dropout in training (it would draw other masks).
+        if (
+            not return_attention
+            and q.device.type != "hpu"
+            and (self.attn_drop.p == 0.0 or not self.training)
+        ):
+            x = nn.functional.scaled_dot_product_attention(
+                q, k, v, attn_mask=attn_mask, scale=self.scale
+            )
+        else:
+            attn = (q * self.scale) @ k.transpose(-2, -1)
+            if attn_mask is not None:
+                attn = attn + attn_mask
+            attn = self.attn_drop(attn.softmax(dim=-1))
+            if return_attention:
+                return attn
+            x = attn @ v
+        x = x.transpose(1, 2).reshape(B, N, -1)
 
         x = self.proj(x)
         x = self.proj_drop(x)

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/luna.py

@@ -423,7 +423,10 @@ class LUNA(EEGModuleMixin, nn.Module, license="apache-2.0"):
                 channel_info, num_channels=num_channels
             )
             if locs is not None and len(locs) == num_channels:
-                return torch.from_numpy(locs).float()
+                locs = torch.from_numpy(locs).float()
+                # MNE stores NaN for a channel without a position.
+                if locs.isfinite().all():
+                    return locs
 
         # Fallback: generate default linear spacing along x-axis
         positions = torch.linspace(-1.0, 1.0, steps=num_channels, dtype=torch.float32)

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/medformer.py

@@ -411,7 +411,10 @@ class _ListPatchEmbedding(nn.Module):
         ]
         self.value_embeddings = nn.ModuleList(linear_layers)
         self.position_embedding = _PositionalEmbedding(d_model=d_model)
-        self.channel_embedding = _PositionalEmbedding(d_model=seq_len)
+        # One row per time sample: windows longer than 5000 samples need more.
+        self.channel_embedding = _PositionalEmbedding(
+            d_model=seq_len, max_len=max(5000, enc_in)
+        )
         self.dropout = nn.Dropout(dropout)
 
         self.learnable_embeddings = nn.ParameterList(

{braindecode-1.8.1.dev174671102 → braindecode-1.8.1.dev174738705}/braindecode/models/msvtnet.py

@@ -7,7 +7,7 @@ import torch
 import torch.nn as nn
 from einops.layers.torch import Rearrange
 
-from braindecode.models.base import EEGModuleMixin
+from braindecode.models.base import EEGModuleMixin, _too_short_input_error
 
 
 class MSVTNet(EEGModuleMixin, nn.Module):
@@ -152,7 +152,10 @@ class MSVTNet(EEGModuleMixin, nn.Module):
         self, cat: bool = True
     ) -> Union[torch.Tensor, list[torch.Tensor]]:
         x = torch.randn(1, 1, self.n_chans, self.n_times)
-        x = [tsconv(x) for tsconv in self.mstsconv]
+        try:
+            x = [tsconv(x) for tsconv in self.mstsconv]
+        except RuntimeError as exc:
+            raise _too_short_input_error(exc, self.input_shape) or exc
         if cat:
             x = torch.cat(x, dim=2)
         return x