Braindecode 1.8.1.dev182199177__tar.gz → 1.8.1.dev182241045__tar.gz

{braindecode-1.8.1.dev182199177/braindecode.egg-info → braindecode-1.8.1.dev182241045}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev182199177
+Version: 1.8.1.dev182241045
 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.dev182199177 → braindecode-1.8.1.dev182241045}/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 +224,125 @@ 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, willneed=False):
+    """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.
+
+    With ``willneed=True`` nothing is read and None is returned: the OS is
+    only asked to start reading the window's bytes into the page cache in the
+    background (``posix_fadvise(POSIX_FADV_WILLNEED)``).
+    """
+    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()
+    if willneed:  # only a hint: any error is left to the read itself
+        try:
+            with open(fname, "rb") as f:
+                os.posix_fadvise(
+                    f.fileno(),
+                    pos[0],
+                    pos[-1] + 16 + size[-1] - pos[0],
+                    os.POSIX_FADV_WILLNEED,
+                )
+        except OSError:
+            pass
+        return None
+    # 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 +946,11 @@ 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)
@@ -1253,7 +1378,8 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
     def __len__(self):
         return self._ensure_cumulative_sizes()[-1]
 
-    def _get_single_item(self, idx: int):
+    def _locate(self, idx: int):
+        """The dataset that holds item ``idx`` and the item's index in it."""
         if idx < 0:
             if -idx > len(self):
                 raise ValueError(
@@ -1267,7 +1393,33 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
             sample_idx = idx
         else:
             sample_idx = idx - cumulative_sizes[dataset_idx - 1]
-        return self.datasets[dataset_idx][sample_idx]
+        return self.datasets[dataset_idx], sample_idx
+
+    def _get_single_item(self, idx: int):
+        ds, i = self._locate(idx)
+        return ds[i]
+
+    def __getitems__(self, indices):
+        """Items ``indices``, as ``[self[i] for i in indices]``.
+
+        A :class:`~torch.utils.data.DataLoader` fetches each batch through
+        this. The lazy FIF windows of the batch are first requested from the
+        OS all at once (``posix_fadvise(POSIX_FADV_WILLNEED)``, where it
+        exists), so their reads from disk overlap instead of running one after
+        another: on a network file system a batch read from cold files then
+        costs about one round trip, not one per window.
+        """
+        # ponytail: advises the whole batch at once; chunk it if a batch's
+        # windows ever outgrow the page cache
+        if hasattr(os, "posix_fadvise"):
+            for idx in indices:
+                if not isinstance(idx, (int, np.integer)):
+                    continue  # e.g. a SequenceSampler tuple: read, not advised
+                ds, i = self._locate(idx)
+                if isinstance(ds, EEGWindowsDataset) and ds._zarr_data is None:
+                    _, start, stop = ds.crop_inds[i].tolist()
+                    _read_fif_window(ds.raw, start, stop, willneed=True)
+        return [self[idx] for idx in indices]
 
     def __getitem__(self, idx: int | list):
         """
@@ -1672,7 +1824,8 @@ 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:

{braindecode-1.8.1.dev182199177 → braindecode-1.8.1.dev182241045}/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.dev182199177 → braindecode-1.8.1.dev182241045}/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.dev182199177 → braindecode-1.8.1.dev182241045}/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.dev182199177 → braindecode-1.8.1.dev182241045}/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
     ----------
@@ -82,6 +83,8 @@ 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
@@ -115,8 +118,11 @@ 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 +131,8 @@ 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.dev182199177 → braindecode-1.8.1.dev182241045}/braindecode/models/barista.py

@@ -7,13 +7,13 @@ from __future__ import annotations
 from typing import Optional
 
 import torch
-import torch.nn.functional as F
 from einops.layers.torch import Rearrange
 from torch import nn
 
 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
 
 
 class BaRISTA(EEGModuleMixin, nn.Module, license="other"):
@@ -689,7 +689,7 @@ class _RotarySelfAttention(nn.Module):
 
         # Dropout falls on the attention output rather than on the weights,
         # through dropout_p, because the reference disables the latter outright.
-        attention = F.scaled_dot_product_attention(query, key, value)
+        attention = _scaled_dot_product_attention(query, key, value)
         attention = self.dropout(attention)
         attention = self.merge_heads(attention)
         return self.o_proj(attention)

{braindecode-1.8.1.dev182199177 → braindecode-1.8.1.dev182241045}/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.dev182199177 → braindecode-1.8.1.dev182241045}/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.dev182199177 → braindecode-1.8.1.dev182241045}/braindecode/models/codebrain.py

@@ -15,6 +15,7 @@ from torch import nn
 from torch.nn.utils.parametrizations import weight_norm
 
 from braindecode.functional import spectral_input
+from braindecode.functional.functions import _next_fast_len
 from braindecode.models.base import EEGModuleMixin
 
 
@@ -521,16 +522,21 @@ class _GConv(nn.Module):
             kernel = F.pad(k_fwd, (0, seq_len)) + F.pad(k_bwd.flip(-1), (seq_len, 0))
 
         # FFT-based convolution: O(N log N)
+        # The bidirectional kernel is circular (lags 0..L-1, then -L..-1).
+        # Rolled to lags -L..L-1 it is a linear convolution read from lag 0,
+        # so any FFT length >= 2 * seq_len works: take a 5-smooth (fast) one.
+        start = seq_len if self.bidirectional else 0
+        n_fft = 2 * _next_fast_len(seq_len)
         # kernel_freq: (channels, d_model, freq_bins)
         # spectral_input(x) dtype: x's, at least float32 (the reference's .float()).
         x_in = spectral_input(x)
-        kernel_freq = torch.fft.rfft(kernel.to(x_in), n=2 * seq_len)
+        kernel_freq = torch.fft.rfft(kernel.roll(start, -1).to(x_in), n=n_fft)
         # x_freq: (batch, d_model, freq_bins)
-        x_freq = torch.fft.rfft(x_in, n=2 * seq_len)
+        x_freq = torch.fft.rfft(x_in, n=n_fft)
         # out_freq: (batch, channels, d_model, freq_bins)
         out_freq = torch.einsum("bhl,chl->bchl", x_freq, kernel_freq)
         # out: (batch, channels, d_model, seq_len)
-        out = torch.fft.irfft(out_freq, n=2 * seq_len)[..., :seq_len].to(x)
+        out = torch.fft.irfft(out_freq, n=n_fft)[..., start : start + seq_len].to(x)
 
         # Skip connection via learnable D matrix
         # (batch, channels, d_model, seq_len)
@@ -697,18 +703,15 @@ class _ResidualBlock(nn.Module):
         # Sliding-window attention branch
         # (batch, 2*res_channels, seq_len) -> (batch, seq_len, 2*res_channels)
         h_attn = h_ssm.transpose(1, 2)
-        if (
-            self.swa_window_size == 1
-            and not self.training
-            and not torch.is_grad_enabled()
-            and h_attn.device.type == "cpu"
-        ):
+        if self.swa_window_size == 1 and h_attn.device.type == "cpu":
             # A one-wide window keeps only the diagonal: the softmax weights are
-            # exactly 1 and 0, so the attention returns its value projection.
-            # Same projection calls as nn.MultiheadAttention (on its (seq,
-            # batch, embed) view), so the output is unchanged bit for bit. (In
-            # training the weights carry dropout; on Gaudi the full attention
-            # was faster.)
+            # exactly 1 and 0, so the attention returns its value projection,
+            # times the weights' dropout in training (the same (batch * heads,
+            # L, L) mask nn.MultiheadAttention draws, of which only the
+            # diagonal reaches the output). Same projection calls as
+            # nn.MultiheadAttention (on its (seq, batch, embed) view), so
+            # outputs and gradients are unchanged bit for bit, and so is the
+            # RNG stream. (On Gaudi the full attention was faster.)
             # ponytail: the unused query/key projections are computed too, to
             # keep the rounding; in_proj_weight[2E:] alone is ~2x less GEMM.
             embed_dim = self.attention.embed_dim
@@ -716,6 +719,16 @@ class _ResidualBlock(nn.Module):
             value = F.linear(
                 seq_first, self.attention.in_proj_weight, self.attention.in_proj_bias
             )[..., 2 * embed_dim :]
+            p = self.attention.dropout
+            if self.training and p > 0.0:
+                n_rows = seq_first.shape[1] * self.attention.num_heads
+                keep = (
+                    value.new_empty(n_rows, seq_len, seq_len)
+                    .bernoulli_(1 - p)
+                    .diagonal(dim1=1, dim2=2)
+                    .div(1 - p)
+                )
+                value = value.reshape(seq_len, n_rows, -1) * keep.t().unsqueeze(-1)
             h_attn = (
                 self.attention.out_proj(value.reshape(-1, embed_dim))
                 .view(seq_first.shape)