Braindecode 1.8.1.dev183301946__tar.gz → 1.8.1.dev183301950__tar.gz

{braindecode-1.8.1.dev183301946 → braindecode-1.8.1.dev183301950}/NOTICE.txt

@@ -54,6 +54,7 @@ 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`
 
 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.dev183301946/braindecode.egg-info → braindecode-1.8.1.dev183301950}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev183301946
+Version: 1.8.1.dev183301950
 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.dev183301946 → braindecode-1.8.1.dev183301950}/braindecode/datasets/base.py

@@ -279,7 +279,7 @@ def _compact_fif_tags(inst):
     return inst
 
 
-def _read_fif_window(raw, start, stop):
+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
@@ -290,6 +290,10 @@ def _read_fif_window(raw, start, stop):
     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)
@@ -316,6 +320,18 @@ def _read_fif_window(raw, start, stop):
     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:
@@ -1362,7 +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(
@@ -1376,7 +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):
         """
@@ -1475,7 +1518,9 @@ 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(
             [
@@ -1501,9 +1546,15 @@ 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:
@@ -1513,6 +1564,7 @@ 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)

{braindecode-1.8.1.dev183301946 → braindecode-1.8.1.dev183301950}/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.dev183301946 → braindecode-1.8.1.dev183301950}/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.dev183301946 → braindecode-1.8.1.dev183301950}/braindecode/models/__init__.py

@@ -36,6 +36,7 @@ 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
@@ -134,6 +135,7 @@ __all__ = [
     "EEGTCNet",
     "DGCNN",
     "EMG2QwertyNet",
+    "EpiNT",
     "NeuroPose",
     "SensingDynamics",
     "VEMG2Pose",

{braindecode-1.8.1.dev183301946 → braindecode-1.8.1.dev183301950}/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(

{braindecode-1.8.1.dev183301946 → braindecode-1.8.1.dev183301950}/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.dev183301946 → braindecode-1.8.1.dev183301950}/braindecode/models/brainmodule.py

@@ -20,6 +20,7 @@ from torch.nn import functional as F
 from braindecode.models.base import EEGModuleMixin
 from braindecode.models.util import has_valid_locations, positions_from_chs_info
 from braindecode.modules import ChannelMerger
+from braindecode.modules.attention import _hpu_lazy
 from braindecode.modules.layers import SubjectLayers
 
 __all__ = ["BrainModule"]
@@ -712,7 +713,7 @@ class _ConvSequence(nn.Module):
                         nn.Conv1d(
                             chout, 2 * chout, 1 + 2 * glu_context, padding=glu_context
                         ),
-                        nn.GLU(dim=1),
+                        _GLU(dim=1),
                     )
                 )
             else:
@@ -727,6 +728,17 @@ class _ConvSequence(nn.Module):
         return x
 
 
+class _GLU(nn.GLU):
+    """``nn.GLU`` that Intel Gaudi's lazy mode can compile."""
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if not torch.jit.is_scripting() and x.device.type == "hpu" and _hpu_lazy():
+            # Its graph compiler fails on ``aten.glu`` after a convolution.
+            a, b = x.chunk(2, dim=self.dim)
+            return a * torch.sigmoid(b)
+        return F.glu(x, self.dim)
+
+
 def _pad_multiple(x: torch.Tensor, base: int) -> torch.Tensor:
     """Pad tensor to be a multiple of base."""
     length = x.shape[-1]

{braindecode-1.8.1.dev183301946 → braindecode-1.8.1.dev183301950}/braindecode/models/codebrain.py

@@ -15,7 +15,9 @@ 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
+from braindecode.modules.attention import _hpu_lazy
 
 
 class CodeBrain(EEGModuleMixin, nn.Module):
@@ -521,16 +523,21 @@ 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 +704,15 @@ 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 +720,16 @@ 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)
@@ -818,10 +832,20 @@ class _ResidualGroup(nn.Module):
     def forward(self, input_data: torch.Tensor) -> torch.Tensor:
         noise = input_data
         h = noise
-        skip = torch.zeros((), dtype=h.dtype, device=h.device)
-        for block in self.residual_blocks:
-            h, skip_n = block((h, noise))
-            skip = skip_n + skip
+        if not torch.jit.is_scripting() and h.device.type == "hpu" and _hpu_lazy():
+            # Gaudi's lazy-mode graph compiler fails on the backward of the
+            # sequential skip sum (one skip_conv dedw node lands in two
+            # pipeline bundles); one n-ary stack-sum reduction compiles.
+            skip_list = []
+            for block in self.residual_blocks:
+                h, skip_n = block((h, noise))
+                skip_list.append(skip_n)
+            skip = torch.stack(skip_list, dim=0).sum(dim=0)
+        else:
+            skip = torch.zeros((), dtype=h.dtype, device=h.device)
+            for block in self.residual_blocks:
+                h, skip_n = block((h, noise))
+                skip = skip_n + skip
 
         return skip * math.sqrt(1.0 / self.num_res_layers)
 

{braindecode-1.8.1.dev183301946 → braindecode-1.8.1.dev183301950}/braindecode/models/contrawr.py

@@ -4,7 +4,7 @@ import torch
 import torch.nn as nn
 from mne.utils import warn
 
-from braindecode.functional import spectral_input
+from braindecode.functional import _real_dft, spectral_input
 from braindecode.models.base import EEGModuleMixin
 from braindecode.models.util import _disable_batch_norm_training_if_batch_size_one
 
@@ -297,6 +297,8 @@ class _STFTModule(nn.Module):
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
         # x: (B, C, T)
+        if _real_dft.needs_real_dft(x) and self.center and self.one_sided:
+            return self._forward_real_dft(x).to(x)
         B, C, T = x.shape
         # flatten batch & channel into one dim
         x_flat = spectral_input(x.reshape(B * C, T))
@@ -319,3 +321,23 @@ class _STFTModule(nn.Module):
         spec = spec_flat.view(B, C, F, L)
 
         return torch.abs(spec).to(x)
+
+    @torch.jit.unused
+    @_real_dft.fp32_island
+    def _forward_real_dft(self, x):
+        # The magnitude above without complex tensors, on the device (Intel
+        # Gaudi / HPU), instead of a host STFT and a host-to-device copy of
+        # the whole spectrogram: reflect-padded frames (gathered, not a
+        # strided unfold) times the one-sided DFT basis. No annotations: with
+        # postponed annotations torch.jit.unused cannot resolve them.
+        pad = self.n_fft // 2
+        x = nn.functional.pad(x, (pad, pad), mode="reflect")
+        n_frames = 1 + (x.shape[-1] - self.n_fft) // self.hop_length
+        starts = torch.arange(n_frames, device=x.device) * self.hop_length
+        index = (starts[:, None] + torch.arange(self.n_fft, device=x.device)).flatten()
+        frames = x.index_select(-1, index).unflatten(-1, (n_frames, self.n_fft))
+        real, imag = _real_dft.real_rfft(frames)
+        magnitude = torch.sqrt(real * real + imag * imag)
+        if self.normalized:
+            magnitude = magnitude / self.n_fft**0.5
+        return magnitude.transpose(-1, -2)