braindecode 1.8.1.dev169274909__tar.gz → 1.8.1.dev170227516__tar.gz

{braindecode-1.8.1.dev169274909 → braindecode-1.8.1.dev170227516}/NOTICE.txt

@@ -42,6 +42,7 @@ The following components are licensed under the MIT License:
 - `braindecode/models/medformer.py`
 - `braindecode/models/tcformer.py`
 - `braindecode/models/ifnet.py`
+- `braindecode/models/mirepnet.py`
 
 ### Apache-2.0 Licensed Files
 
@@ -50,16 +51,16 @@ The following components are licensed under the Apache License 2.0:
 - `braindecode/models/mvpformer.py`
 - `braindecode/models/zuna.py`
 - `braindecode/models/luna.py`
+- `braindecode/models/brant.py`
+- `braindecode/models/diver1.py`
+- `braindecode/models/mscformer.py`
 
 ### USC Academic License Files
 
-The following component is a derivative of `ShanechiLab/BaRISTA`, released by
-the University of Southern California under an academic license that permits
-use, copying, modification and distribution for educational, research and
-non-profit purposes only. Commercial use requires an agreement with the USC
-Stevens Center for Innovation.
+[License](https://github.com/ShanechiLab/BaRISTA/blob/master/LICENSE.md)
 
 - `braindecode/models/barista.py`
+- `docs/_static/model/barista.jpeg`
 
 ## License Links
 

{braindecode-1.8.1.dev169274909/braindecode.egg-info → braindecode-1.8.1.dev170227516}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev169274909
+Version: 1.8.1.dev170227516
 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>
@@ -22,8 +22,8 @@ Requires-Python: >=3.11
 Description-Content-Type: text/x-rst
 License-File: LICENSE.txt
 License-File: NOTICE.txt
-Requires-Dist: torch>=2.0
-Requires-Dist: torchaudio>=2.0
+Requires-Dist: torch>=2.4
+Requires-Dist: torchaudio>=2.4
 Requires-Dist: mne>=1.11.0
 Requires-Dist: mne_bids>=0.18
 Requires-Dist: h5py

{braindecode-1.8.1.dev169274909 → braindecode-1.8.1.dev170227516}/braindecode/datasets/base.py

@@ -5,6 +5,7 @@
 #          Simon Brandt <simonbrandt@protonmail.com>
 #          David Sabbagh <dav.sabbagh@gmail.com>
 #          Robin Schirrmeister <robintibor@gmail.com>
+#          Sarthak Tayal <sarthaktayal2@gmail.com>
 #
 # License: BSD (3-clause)
 
@@ -1398,6 +1399,9 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
                 df = ds._windows.metadata
             else:
                 df = ds.metadata
+            if isinstance(df, pd.DataFrame):
+                # copy so the description columns stay out of the dataset
+                df = df.copy()
             for k, v in ds.description.items():
                 df[k] = v
             all_dfs.append(df)

{braindecode-1.8.1.dev169274909 → braindecode-1.8.1.dev170227516}/braindecode/datautil/serialization.py

@@ -1,6 +1,7 @@
 """Convenience functions for storing and loading of windows datasets."""
 
 # Authors: Lukas Gemein <l.gemein@gmail.com>
+#          Sarthak Tayal <sarthaktayal2@gmail.com>
 #
 # License: BSD (3-clause)
 
@@ -270,7 +271,9 @@ def _load_parallel(path, i, preload, is_raw, has_stored_windows):
     else:
         window_kwargs = _load_kwargs_json("window_kwargs", sub_dir)
         windows_ds_kwargs = [
-            kwargs[1] for kwargs in window_kwargs if kwargs[0] == "WindowsDataset"
+            kwargs[1]
+            for kwargs in window_kwargs
+            if kwargs[0] in ("WindowsDataset", "EEGWindowsDataset")
         ]
         windows_ds_kwargs = windows_ds_kwargs[0] if len(windows_ds_kwargs) == 1 else {}
         if is_raw:

{braindecode-1.8.1.dev169274909 → braindecode-1.8.1.dev170227516}/braindecode/functional/__init__.py

@@ -1,6 +1,5 @@
 from .functions import (
     _get_gaussian_kernel1d,
-    apply_rotary,
     daubechies_filters,
     drop_path,
     dwt_max_level,
@@ -9,8 +8,7 @@ from .functions import (
     iou_1d,
     pairwise_iou_1d,
     plv_time,
-    rotary_positional_encoding,
-    rotate_half,
+    rotate_pairs,
     safe_log,
     sinusoidal_positional_encoding,
     square,
@@ -20,7 +18,6 @@ from .initialization import glorot_weight_zero_bias, rescale_parameter
 
 __all__ = [
     "_get_gaussian_kernel1d",
-    "apply_rotary",
     "daubechies_filters",
     "drop_path",
     "dwt_max_level",
@@ -29,8 +26,7 @@ __all__ = [
     "iou_1d",
     "pairwise_iou_1d",
     "plv_time",
-    "rotary_positional_encoding",
-    "rotate_half",
+    "rotate_pairs",
     "safe_log",
     "sinusoidal_positional_encoding",
     "square",

braindecode-1.8.1.dev170227516/braindecode/functional/_real_dft.py

@@ -0,0 +1,138 @@
+"""Real-valued DFT helpers for accelerators without complex tensors.
+
+Intel Gaudi (HPU) has no complex dtype, so ``torch.fft`` with complex outputs
+cannot run there. These helpers express the one-sided DFT, and the inverse
+transforms built on top of it, as matmuls against a cosine/sine basis. Under
+bfloat16 autocast those matmuls would run in bfloat16, whereas ``torch.fft`` is
+never downcast, so callers wrap the numerically sensitive parts with
+:func:`fp32_island`.
+"""
+
+# Authors: Bruno Aristimunha <b.aristimunha@gmail.com>
+#
+# License: BSD (3-clause)
+import functools
+
+import torch
+
+
+def needs_real_dft(x: torch.Tensor) -> bool:
+    """Whether ``x`` lives on a device without complex-tensor FFT support."""
+    return x.device.type == "hpu"
+
+
+def fp32_island(fn):
+    """Run ``fn`` with autocast disabled, in at least float32 precision.
+
+    Any bfloat16/float16 tensor argument is upcast to float32 before ``fn``
+    runs; float32 and float64 arguments are left untouched, so the same code
+    path can also be checked against a float64 ``torch.fft`` reference.
+    """
+
+    @functools.wraps(fn)
+    def wrapped(*args, **kwargs):
+        ref = next(a for a in args if isinstance(a, torch.Tensor))
+        with torch.autocast(device_type=ref.device.type, enabled=False):
+            args = tuple(
+                a.float()
+                if isinstance(a, torch.Tensor)
+                and a.is_floating_point()
+                and torch.finfo(a.dtype).bits < 32
+                else a
+                for a in args
+            )
+            return fn(*args, **kwargs)
+
+    return wrapped
+
+
+def real_dft_basis(n, *, device, dtype):
+    """One-sided cosine/sine basis of an ``n``-point DFT, shape ``(n // 2 + 1, n)``.
+
+    The phase index is reduced modulo ``n`` in integers before scaling, so the
+    DC row is exact (``sin`` is exactly zero there); the Nyquist row (present
+    for even ``n``) is exact to floating-point precision.
+    """
+    if n <= 0:
+        raise ValueError("n must be positive")
+    if (n // 2) * (n - 1) >= 2**31:
+        # The int32 phase index below would overflow and give wrong phases.
+        raise ValueError(f"n={n} exceeds the 65536 samples of the real-valued DFT")
+    freqs = torch.arange(n // 2 + 1, device=device, dtype=torch.int32)
+    samples = torch.arange(n, device=device, dtype=torch.int32)
+    phase_index = (freqs[:, None] * samples[None, :]) % n
+    angles = 2 * torch.pi * phase_index.to(dtype) / n
+    return angles.cos(), angles.sin()
+
+
+def real_rfft(x):
+    """Real and imaginary parts of ``torch.fft.rfft(x, dim=-1)`` without complex tensors."""
+    n = x.shape[-1]
+    cosine, sine = real_dft_basis(n, device=x.device, dtype=x.dtype)
+    return x @ cosine.transpose(0, 1), -(x @ sine.transpose(0, 1))
+
+
+def real_irfft(real, imag, *, n):
+    """``torch.fft.irfft(torch.complex(real, imag), n=n, dim=-1)`` without complex tensors."""
+    if real.shape[-1] != n // 2 + 1:
+        raise ValueError(
+            f"expected {n // 2 + 1} one-sided bins for n={n}, got {real.shape[-1]}"
+        )
+    cosine, sine = real_dft_basis(n, device=real.device, dtype=real.dtype)
+    output = real[..., :1] @ cosine[:1, :]
+    if n % 2 == 0:
+        interior = slice(1, -1)
+        output = output + real[..., -1:] @ cosine[-1:, :]
+    else:
+        interior = slice(1, None)
+    if real.shape[-1] > (2 if n % 2 == 0 else 1):
+        output = output + 2 * (
+            real[..., interior] @ cosine[interior, :]
+            - imag[..., interior] @ sine[interior, :]
+        )
+    return output / n
+
+
+def real_analytic_ifft(real, imag, *, n):
+    """Inverse ``n``-point FFT of a one-sided spectrum, without complex tensors.
+
+    ``real``/``imag`` hold the bins ``0 .. n // 2`` of a length-``n`` spectrum
+    whose remaining (negative-frequency) bins are zero, so this is a plain
+    (non-Hermitian) inverse DFT, not :func:`real_irfft`. Returns the real and
+    imaginary parts of the generally complex time-domain signal, each of shape
+    ``(..., n)``. This is the last step of :func:`hilbert_freq_real`.
+    """
+    if real.shape[-1] != n // 2 + 1:
+        raise ValueError(
+            f"expected {n // 2 + 1} one-sided bins for n={n}, got {real.shape[-1]}"
+        )
+    cosine, sine = real_dft_basis(n, device=real.device, dtype=real.dtype)
+    out_real = (real @ cosine - imag @ sine) / n
+    out_imag = (real @ sine + imag @ cosine) / n
+    return out_real, out_imag
+
+
+@torch.jit.unused
+def hilbert_freq_real(x: torch.Tensor, forward_fourier: bool) -> torch.Tensor:
+    """Real-valued equivalent of :func:`braindecode.functional.hilbert_freq`.
+
+    Lives here rather than in ``functions.py`` because ``torch.jit.unused`` needs
+    real annotations, and ``functions.py`` uses ``from __future__ import annotations``.
+    """
+
+    @fp32_island
+    def compute(x):
+        if forward_fourier:
+            seq_len = x.shape[-1]
+            real, imag = real_rfft(x)
+        else:
+            seq_len = 2 * (x.shape[-2] - 1)
+            real, imag = x[..., 0], x[..., 1]
+        scale = torch.full((real.shape[-1],), 2.0, dtype=real.dtype, device=real.device)
+        scale[0] = 1.0  # Don't multiply the DC-term by 2
+        if forward_fourier and seq_len % 2 == 0:
+            scale[-1] = 1.0  # Nor the Nyquist term, which only exists for even lengths
+        out_real, out_imag = real_analytic_ifft(real * scale, imag * scale, n=seq_len)
+        return torch.stack((out_real, out_imag), dim=-1)
+
+    return compute(x)

{braindecode-1.8.1.dev169274909 → braindecode-1.8.1.dev170227516}/braindecode/functional/functions.py

@@ -1,7 +1,6 @@
 # Authors: Robin Schirrmeister <robintibor@gmail.com>
 #          Bruno Aristimunha <b.aristimunha@gmail.com>
 #          Sarthak Tayal <sarthaktayal2@gmail.com>
-#          Julien Gadonneix <juliengado.2001@gmail.com>
 #
 # License: BSD (3-clause)
 
@@ -13,6 +12,8 @@ import numpy as np
 import torch
 import torch.nn.functional as F
 
+from . import _real_dft
+
 
 def square(x):
     return x * x
@@ -166,18 +167,40 @@ def hilbert_freq(x: torch.Tensor, forward_fourier: bool = True) -> torch.Tensor:
     https://github.com/scipy/scipy/blob/v1.14.1/scipy/signal/_signaltools.py#L2287-L2394
 
     """
+    if _real_dft.needs_real_dft(x):
+        return _real_dft.hilbert_freq_real(x, forward_fourier).to(x.dtype)
+
+    input_dtype = x.dtype
+    # ``view_as_complex`` does not accept bfloat16 real/imaginary pairs.  The
+    # HPU path can produce bfloat16 Fourier coefficients under autocast, so do
+    # the complex-valued part in float32 and restore the real-valued contract
+    # at the boundary.  Promoting before the FFT also covers backends that do
+    # not implement bfloat16 FFT kernels.
+    if input_dtype == torch.bfloat16:
+        x = x.float()
+
     if forward_fourier:
+        seq_len = x.shape[-1]
         x = torch.fft.rfft(x, norm=None, dim=-1)
         x = torch.view_as_real(x)
+    else:
+        # The one-sided spectrum does not record whether the signal length
+        # was odd or even, so assume an even length here.
+        seq_len = 2 * (x.shape[-2] - 1)
     x = x * 2.0
     x[..., 0, :] = x[..., 0, :] / 2.0  # Don't multiply the DC-term by 2
+    if forward_fourier and seq_len % 2 == 0:
+        # Nor the Nyquist term, which only exists for even lengths
+        x[..., -1, :] = x[..., -1, :] / 2.0
     x = F.pad(
-        x, [0, 0, 0, x.shape[-2] - 2]
+        x, [0, 0, 0, seq_len - x.shape[-2]]
     )  # Fill Fourier coefficients to retain shape
     x = torch.view_as_complex(x)
     x = torch.fft.ifft(x, norm=None, dim=-1)  # returns complex signal
     x = torch.view_as_real(x)
 
+    if input_dtype == torch.bfloat16:
+        x = x.to(input_dtype)
     return x
 
 
@@ -449,109 +472,19 @@ def sinusoidal_positional_encoding(n_positions: int, dim: int) -> torch.Tensor:
     return pe[:, :dim].contiguous()
 
 
-def rotary_positional_encoding(
-    n_positions: int, head_dim: int, base: float = 10000.0
-) -> tuple[torch.Tensor, torch.Tensor]:
-    r"""Cosine and sine tables of a half-split rotary embedding (RoPE).
-
-    RoPE [Su2021]_ encodes position by rotating each attention head in the
-    :math:`d/2` planes formed by pairing up its dimensions, the plane of index
-    :math:`j` turning by :math:`p \cdot base^{-2j/d}` at position :math:`p`.
-
-    Pairs are formed *half-split*, as :math:`(x_j, x_{j + d/2})`, which is the
-    convention of GPT-NeoX, Llama and the transformers library, rather than the
-    *interleaved* :math:`(x_{2j}, x_{2j+1})` of the original paper. The two
-    differ only by a fixed permutation of the head dimension, so a model trained
-    from scratch is indifferent, but a pre-trained checkpoint is not. Use
-    :func:`apply_rotary`, whose pairing matches.
-
-    Each of the :math:`d/2` angles is therefore tabulated twice, at ``j`` and at
-    ``j + head_dim // 2``, so the tables broadcast against a full-width head.
-
-    Parameters
-    ----------
-    n_positions : int
-        Number of positions to tabulate.
-    head_dim : int
-        Dimension of an attention head. Must be even.
-    base : float, default=10000.0
-        Base of the geometric progression of frequencies. Larger values lengthen
-        the wavelengths, extending the range of positions that stay
-        distinguishable.
-
-    Returns
-    -------
-    tuple of torch.Tensor
-        Two ``(n_positions, head_dim)`` float tables, the cosines and the sines.
-        Callers index them with their own positions and add broadcast axes.
-
-    References
-    ----------
-    .. [Su2021] Su, J., Lu, Y., Pan, S., Murtadha, A., Wen, B., & Liu, Y.
-       (2021). RoFormer: Enhanced Transformer with Rotary Position Embedding.
-       arXiv preprint arXiv:2104.09864.
-
-    Examples
-    --------
-    >>> import torch
-    >>> from braindecode.functional import rotary_positional_encoding
-    >>> cos, sin = rotary_positional_encoding(n_positions=16, head_dim=8)
-    >>> cos.shape
-    torch.Size([16, 8])
-    """
-    if head_dim % 2:
-        raise ValueError(
-            f"head_dim must be even to pair up its dimensions, got {head_dim}."
-        )
-    inv_freq = 1.0 / torch.pow(
-        base, torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim
-    )
-    angles = torch.outer(torch.arange(n_positions, dtype=torch.float32), inv_freq)
-    # Concatenating rather than interleaving is what places the angle of plane j
-    # at j and j + head_dim // 2, which is the pairing rotate_half applies.
-    angles = torch.cat((angles, angles), dim=-1)
-    return angles.cos(), angles.sin()
-
-
-def rotate_half(x: torch.Tensor) -> torch.Tensor:
-    """Swap the two halves of the last axis of ``x``, negating the second one.
-
-    The companion of :func:`rotary_positional_encoding`: pairing dimension
-    :math:`j` with :math:`j + d/2` makes this the quarter-turn that
-    :func:`apply_rotary` mixes into the unrotated tensor.
-
-    Parameters
-    ----------
-    x : torch.Tensor
-        Tensor whose last axis is even.
-
-    Returns
-    -------
-    torch.Tensor
-        Tensor of the same shape, holding ``(-x[..., d/2:], x[..., :d/2])``.
-    """
-    first, second = x.chunk(2, dim=-1)
-    return torch.cat((-second, first), dim=-1)
-
-
-def apply_rotary(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
-    r"""Rotate ``x`` by the angles of a half-split rotary embedding.
-
-    Computes :math:`x \cos\theta + \mathrm{rotate\_half}(x) \sin\theta`, which
-    turns each pair :math:`(x_j, x_{j + d/2})` by :math:`\theta_j` in its own
-    plane. Being a rotation, it leaves the norm of every head unchanged.
+def rotate_pairs(x: torch.Tensor) -> torch.Tensor:
+    """Rotate adjacent feature pairs from ``(a, b)`` to ``(-b, a)``.
 
     Parameters
     ----------
     x : torch.Tensor
-        Queries or keys, with the head dimension last.
-    cos, sin : torch.Tensor
-        Tables from :func:`rotary_positional_encoding`, indexed by the positions
-        of ``x`` and carrying whatever axes are needed to broadcast against it.
+        Tensor of shape ``(..., features)`` with an even final dimension.
 
     Returns
     -------
     torch.Tensor
-        Rotated tensor, of the shape and dtype of ``x``.
+        Same shape, device and dtype as ``x``. Frequencies, positional indexing
+        and mixed-precision policy are left to the caller.
     """
-    return x * cos.to(x.dtype) + rotate_half(x) * sin.to(x.dtype)
+    pairs = x.unflatten(-1, (-1, 2))
+    return torch.stack((-pairs[..., 1], pairs[..., 0]), dim=-1).flatten(-2)

{braindecode-1.8.1.dev169274909 → braindecode-1.8.1.dev170227516}/braindecode/models/__init__.py

@@ -7,7 +7,9 @@ from .barista import BaRISTA
 from .base import EEGModuleMixin
 from .bendr import BENDR, InterpolatedBENDR
 from .biot import BIOT, InterpolatedBIOT
+from .brainbert import BrainBERT
 from .brainmodule import BrainModule
+from .brant import Brant
 from .cbramod import CBraMod
 from .codebrain import CodeBrain
 from .contrawr import ContraWR
@@ -16,6 +18,7 @@ from .dance import DANCE
 from .deep4 import Deep4Net
 from .deepsleepnet import DeepSleepNet
 from .dgcnn import DGCNN
+from .diver1 import DIVER1
 from .eegconformer import EEGConformer
 from .eegdino import EEGDINO
 from .eeginception_erp import EEGInceptionERP
@@ -39,6 +42,8 @@ from .labram import InterpolatedLaBraM, Labram
 from .luna import LUNA
 from .medformer import MEDFormer
 from .meta_neuromotor import MetaNeuromotorHand
+from .mirepnet import MIRepNet
+from .mscformer import MSCFormer
 from .msvtnet import MSVTNet
 from .mvpformer import MVPFormer
 from .neuropose import NeuroPose
@@ -89,6 +94,7 @@ __all__ = [
     "EEGModuleMixin",
     "BIOT",
     "BENDR",
+    "BrainBERT",
     "CBraMod",
     "CodeBrain",
     "ContraWR",
@@ -96,7 +102,9 @@ __all__ = [
     "DANCE",
     "Deep4Net",
     "DeepSleepNet",
+    "DIVER1",
     "BrainModule",
+    "Brant",
     "EEGConformer",
     "EEGDINO",
     "EEGPT",
@@ -131,6 +139,8 @@ __all__ = [
     "extract_channel_locations_from_chs_info",
     "positions_from_chs_info",
     "MEDFormer",
+    "MIRepNet",
+    "MSCFormer",
     "MSVTNet",
     "MVPFormer",
     "PBT",