braindecode 1.8.1.dev181104868__tar.gz → 1.8.1.dev182582231__tar.gz

{braindecode-1.8.1.dev181104868/braindecode.egg-info → braindecode-1.8.1.dev182582231}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev181104868
+Version: 1.8.1.dev182582231
 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.dev181104868 → braindecode-1.8.1.dev182582231}/braindecode/functional/functions.py

@@ -165,6 +165,15 @@ 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
 
     """
+    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:
         x = torch.fft.rfft(x, norm=None, dim=-1)
         x = torch.view_as_real(x)
@@ -177,6 +186,8 @@ def hilbert_freq(x: torch.Tensor, forward_fourier: bool = True) -> torch.Tensor:
     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
 
 

{braindecode-1.8.1.dev181104868 → braindecode-1.8.1.dev182582231}/braindecode/modules/filter.py

@@ -643,16 +643,17 @@ class GeneralizedGaussianFilter(nn.Module):
         """
         min_f_mean = self.clamp_f_mean[0] / (self.sample_rate / 2)
         max_f_mean = self.clamp_f_mean[1] / (self.sample_rate / 2)
-        f_mean = torch.clamp(self.f_mean, min=min_f_mean, max=max_f_mean)
+        f_mean = torch.clamp(self.f_mean.clone(), min=min_f_mean, max=max_f_mean)
         bandwidth = torch.clamp(
-            self.bandwidth, min=1.0 / (self.sample_rate / 2), max=1.0
+            self.bandwidth.clone(), min=1.0 / (self.sample_rate / 2), max=1.0
         )
-        shape = torch.clamp(self.shape, min=2.0, max=3.0)
+        shape = torch.clamp(self.shape.clone(), min=2.0, max=3.0)
 
-        if not torch.jit.is_scripting():
-            self.f_mean.data = f_mean
-            self.bandwidth.data = bandwidth
-            self.shape.data = shape
+        if not torch.jit.is_scripting() and not torch.jit.is_tracing():
+            with torch.no_grad():
+                self.f_mean.copy_(f_mean)
+                self.bandwidth.copy_(bandwidth)
+                self.shape.copy_(shape)
 
         # Create magnitude response with gain=1 -> (channels, freqs)
         mag_response = self.exponential_power(

{braindecode-1.8.1.dev181104868 → braindecode-1.8.1.dev182582231}/braindecode/modules/parametrization.py

@@ -34,5 +34,16 @@ class MaxNormParametrize(nn.Module):
         self.max_norm = max_norm
 
     def forward(self, X: torch.Tensor) -> torch.Tensor:
-        # Renormalize each "row" (dim=0 slice) to have at most self.max_norm L2-norm
-        return X.renorm(p=2, dim=0, maxnorm=self.max_norm)
+        # Renormalize each "row" (dim=0 slice) to have at most self.max_norm
+        # L2-norm. This is ``X.renorm(p=2, dim=0, maxnorm=self.max_norm)``
+        # written out, with the same 1e-7 epsilon: ``renorm`` fails on Intel
+        # Gaudi (HPU), which decomposes it into a broadcast that breaks for
+        # some weight shapes. The rescale uses a mask, not a data-dependent
+        # branch, so it also traces and compiles without graph breaks.
+        # Rows are flattened so the scale broadcasts as (rows, 1); this form
+        # is also TorchScript-compatible.
+        rows = X.reshape(X.shape[0], -1)
+        norm = rows.norm(p=2, dim=1, keepdim=True)
+        over = (norm > self.max_norm).to(X.dtype)
+        scale = (1.0 - over) + over * (self.max_norm / (norm + 1e-7))
+        return (rows * scale).reshape_as(X)

braindecode-1.8.1.dev182582231/braindecode/version.py

@@ -0,0 +1 @@
+__version__ = "1.8.1.dev182582231"

{braindecode-1.8.1.dev181104868 → braindecode-1.8.1.dev182582231/braindecode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev181104868
+Version: 1.8.1.dev182582231
 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.dev181104868 → braindecode-1.8.1.dev182582231}/docs/whats_new.rst

@@ -78,6 +78,19 @@ Bug fixes
   :class:`braindecode.models.Deep4Net` unusable on single-channel data
   (:gh:`1154` by `Julien Gadonneix`_).
 
+- Fix :class:`braindecode.models.EEGMiner` on Intel Gaudi (HPU) and under
+  ``torch.jit.trace``: :class:`braindecode.modules.GeneralizedGaussianFilter`
+  now clamps its parameters in place under ``torch.no_grad()`` instead of
+  reassigning ``.data``, and :func:`braindecode.functional.hilbert_freq`
+  computes the complex FFT step in float32 for bfloat16 inputs, so the phase
+  features work under bfloat16 autocast (:gh:`1183` by `Bruno Aristimunha`_).
+- Fix :class:`braindecode.modules.MaxNormParametrize` failing on Intel Gaudi
+  (HPU), which affects every model with a max-norm weight constraint
+  (e.g. :class:`braindecode.models.EEGNet`,
+  :class:`braindecode.models.ATCNet`). The row rescale is now written out
+  instead of calling ``Tensor.renorm``; values and gradients are unchanged
+  (:gh:`1184` by `Bruno Aristimunha`_).
+
 
 Current 1.8.0 (2026-08-31)
 ===============================

braindecode-1.8.1.dev181104868/braindecode/version.py

@@ -1 +0,0 @@
-__version__ = "1.8.1.dev181104868"