Braindecode 1.8.1.dev172484635__tar.gz → 1.8.1.dev172496780__tar.gz

{braindecode-1.8.1.dev172484635/braindecode.egg-info → braindecode-1.8.1.dev172496780}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev172484635
+Version: 1.8.1.dev172496780
 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.dev172484635 → braindecode-1.8.1.dev172496780}/braindecode/models/deep4.py

@@ -186,13 +186,18 @@ class Deep4Net(EEGModuleMixin, nn.Sequential):
         # When padronize all layers,
         # add the old's parameters here
         self.mapping = {
-            "conv_time.weight": "conv_time_spat.conv_time.weight",
-            "conv_spat.weight": "conv_time_spat.conv_spat.weight",
-            "conv_time.bias": "conv_time_spat.conv_time.bias",
-            "conv_spat.bias": "conv_time_spat.conv_spat.bias",
             "conv_classifier.weight": "final_layer.conv_classifier.weight",
             "conv_classifier.bias": "final_layer.conv_classifier.bias",
         }
+        if self.split_first_layer:
+            self.mapping.update(
+                {
+                    "conv_time.weight": "conv_time_spat.conv_time.weight",
+                    "conv_spat.weight": "conv_time_spat.conv_spat.weight",
+                    "conv_time.bias": "conv_time_spat.conv_time.bias",
+                    "conv_spat.bias": "conv_time_spat.conv_spat.bias",
+                }
+            )
 
         if self.stride_before_pool:
             conv_stride = self.pool_time_stride
@@ -319,14 +324,16 @@ class Deep4Net(EEGModuleMixin, nn.Sequential):
 
         # Initialization, xavier is same as in our paper...
         # was default from lasagne
-        init.xavier_uniform_(self.conv_time_spat.conv_time.weight, gain=1)
-        # maybe no bias in case of no split layer and batch norm
-        if self.split_first_layer or (not self.batch_norm):
-            init.constant_(self.conv_time_spat.conv_time.bias, 0)
         if self.split_first_layer:
+            init.xavier_uniform_(self.conv_time_spat.conv_time.weight, gain=1)
+            init.constant_(self.conv_time_spat.conv_time.bias, 0)
             init.xavier_uniform_(self.conv_time_spat.conv_spat.weight, gain=1)
             if not self.batch_norm:
                 init.constant_(self.conv_time_spat.conv_spat.bias, 0)
+        else:
+            init.xavier_uniform_(self.conv_time.weight, gain=1)
+            if not self.batch_norm:
+                init.constant_(self.conv_time.bias, 0)
         if self.batch_norm:
             init.constant_(self.bnorm.weight, 1)
             init.constant_(self.bnorm.bias, 0)

{braindecode-1.8.1.dev172484635 → braindecode-1.8.1.dev172496780}/braindecode/models/eegpt.py

@@ -88,7 +88,8 @@ class EEGPT(EEGModuleMixin, nn.Module):
         - *Role.* Maps raw spatio-temporal EEG patches into a sequence of latent tokens :math:`z`.
     - `EEGPT.chans_id` **(Channel Identification)**
         - *Operations.* A buffer containing channel indices mapped from the standard channel names provided
-          in ``chs_info`` [eegpt]_.
+          in ``chs_info`` [eegpt]_. It is rebuilt from the montage at construction and not stored in
+          checkpoints, so pretrained weights load on any montage.
         - *Role.* Provides the spatial identity for each input channel, allowing the model to look up
           the correct channel embedding vector :math:`\varsigma_i`.
     - **Local Spatio-Temporal Embedding** (Input Processing)
@@ -338,8 +339,12 @@ class EEGPT(EEGModuleMixin, nn.Module):
         else:
             self.channel_names = None  # type: ignore
 
+        # Derived from the montage above, so it is not saved in the state dict:
+        # loading a checkpoint must not replace the IDs of this model's channels.
         self.register_buffer(
-            "chans_id", self.target_encoder.prepare_chan_ids(self.channel_names)
+            "chans_id",
+            self.target_encoder.prepare_chan_ids(self.channel_names),
+            persistent=False,
         )
 
         self.flattened_encoder_output_dim = (
@@ -391,6 +396,13 @@ class EEGPT(EEGModuleMixin, nn.Module):
             "embed_dim": self.embed_dim,
         }
 
+    def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+        # Checkpoints saved before ``chans_id`` became non-persistent (including
+        # the released EEGPT weights) still carry it. Drop it, so the IDs keep
+        # matching this model's montage and a different montage can load.
+        state_dict.pop(prefix + "chans_id", None)
+        super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
     def forward(self, x, return_features=False):
         """
         Forward pass.

{braindecode-1.8.1.dev172484635 → braindecode-1.8.1.dev172496780}/braindecode/models/shallow_fbcsp.py

@@ -124,13 +124,18 @@ class ShallowFBCSPNet(EEGModuleMixin, nn.Sequential):
         self.drop_prob = drop_prob
 
         self.mapping = {
-            "conv_time.weight": "conv_time_spat.conv_time.weight",
-            "conv_spat.weight": "conv_time_spat.conv_spat.weight",
-            "conv_time.bias": "conv_time_spat.conv_time.bias",
-            "conv_spat.bias": "conv_time_spat.conv_spat.bias",
             "conv_classifier.weight": "final_layer.conv_classifier.weight",
             "conv_classifier.bias": "final_layer.conv_classifier.bias",
         }
+        if self.split_first_layer:
+            self.mapping.update(
+                {
+                    "conv_time.weight": "conv_time_spat.conv_time.weight",
+                    "conv_spat.weight": "conv_time_spat.conv_spat.weight",
+                    "conv_time.bias": "conv_time_spat.conv_time.bias",
+                    "conv_spat.bias": "conv_time_spat.conv_spat.bias",
+                }
+            )
 
         self.add_module("ensuredims", Ensure4d())
         pool_class = dict(max=nn.MaxPool2d, mean=nn.AvgPool2d)[self.pool_mode]
@@ -203,14 +208,16 @@ class ShallowFBCSPNet(EEGModuleMixin, nn.Sequential):
         self.add_module("final_layer", module)
 
         # Initialization, xavier is same as in paper...
-        init.xavier_uniform_(self.conv_time_spat.conv_time.weight, gain=1)
-        # maybe no bias in case of no split layer and batch norm
-        if self.split_first_layer or (not self.batch_norm):
-            init.constant_(self.conv_time_spat.conv_time.bias, 0)
         if self.split_first_layer:
+            init.xavier_uniform_(self.conv_time_spat.conv_time.weight, gain=1)
+            init.constant_(self.conv_time_spat.conv_time.bias, 0)
             init.xavier_uniform_(self.conv_time_spat.conv_spat.weight, gain=1)
             if not self.batch_norm:
                 init.constant_(self.conv_time_spat.conv_spat.bias, 0)
+        else:
+            init.xavier_uniform_(self.conv_time.weight, gain=1)
+            if not self.batch_norm:
+                init.constant_(self.conv_time.bias, 0)
         if self.batch_norm:
             init.constant_(self.bnorm.weight, 1)
             init.constant_(self.bnorm.bias, 0)

{braindecode-1.8.1.dev172484635 → braindecode-1.8.1.dev172496780}/braindecode/modules/convolution.py

@@ -48,6 +48,14 @@ class AvgPool2dWithConv(nn.Module):
         # that initializes parameters or something
         self._pool_weights = None
 
+    def _pool_weights_match_input(self, x, weight_shape):
+        return (
+            self._pool_weights is not None
+            and tuple(self._pool_weights.size()) == tuple(weight_shape)
+            and self._pool_weights.device == x.device
+            and self._pool_weights.dtype == x.dtype
+        )
+
     def forward(self, x):
         # Create weights for the convolution on demand:
         # size or type of x changed...
@@ -58,17 +66,10 @@ class AvgPool2dWithConv(nn.Module):
             self.kernel_size[0],
             self.kernel_size[1],
         )
-        if self._pool_weights is None or (
-            (tuple(self._pool_weights.size()) != tuple(weight_shape))
-            or (self._pool_weights.is_cuda != x.is_cuda)
-            or (self._pool_weights.data.type() != x.data.type())
-        ):
+        if not self._pool_weights_match_input(x, weight_shape):
             n_pool = np.prod(self.kernel_size)
             weights = np_to_th(np.ones(weight_shape, dtype=np.float32) / float(n_pool))
-            weights = weights.type_as(x)
-            if x.is_cuda:
-                weights = weights.cuda()
-            self._pool_weights = weights
+            self._pool_weights = weights.to(device=x.device, dtype=x.dtype)
 
         pooled = F.conv2d(
             x,

{braindecode-1.8.1.dev172484635 → braindecode-1.8.1.dev172496780}/braindecode/modules/layers.py

@@ -62,6 +62,8 @@ class Chomp1d(nn.Module):
         return "chomp_size={}".format(self.chomp_size)
 
     def forward(self, x):
+        if self.chomp_size == 0:
+            return x.contiguous()
         return x[:, :, : -self.chomp_size].contiguous()
 
 
@@ -110,8 +112,8 @@ class TimeDistributed(nn.Module):
             Shape (batch_size, seq_len, output_size).
         """
         b, s, c, t = x.shape
-        out = self.module(x.view(b * s, c, t))
-        return out.view(b, s, -1)
+        out = self.module(x.reshape(b * s, c, t))
+        return out.reshape(b, s, -1)
 
 
 class DropPath(nn.Module):

braindecode-1.8.1.dev172496780/braindecode/version.py

@@ -0,0 +1 @@
+__version__ = "1.8.1.dev172496780"

{braindecode-1.8.1.dev172484635 → braindecode-1.8.1.dev172496780/braindecode.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.1.dev172484635
+Version: 1.8.1.dev172496780
 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.dev172484635 → braindecode-1.8.1.dev172496780}/docs/whats_new.rst

@@ -28,6 +28,10 @@ Current 1.8.1 (2026-08-31)
 Enhancements
 ============
 
+- Add registry-wide model contract tests that automatically cover every registered
+  model, checking eval-mode input/state purity, finite batched outputs,
+  batch-permutation equivariance, and ``get_config`` + ``state_dict``
+  reconstruction (:gh:`1208` by `lindicaphxag-tech`_).
 - Generate a version-scoped ``llms.txt`` and selected Markdown documentation
   entry points with source-commit attribution and critical-page coverage checks.
 
@@ -130,6 +134,22 @@ Requirements
 Bug fixes
 ==========
 
+- Fix :class:`braindecode.modules.AvgPool2dWithConv` to invalidate cached pooling weights when the exact input device changes, preventing stale weights from being reused across CUDA devices or non-CUDA backends with the same dtype. By `lindicaphxag-tech`_.
+- Fix :class:`braindecode.models.ShallowFBCSPNet` with
+  ``split_first_layer=False``, which attempted to initialize and remap
+  checkpoint keys through the split-only ``conv_time_spat`` module after the
+  CombinedConv refactor. The unsplit path now initializes and loads its direct
+  temporal convolution as before (:gh:`1212` by `lindicaphxag-tech`_).
+
+- Fix :class:`braindecode.models.Deep4Net` with
+  ``split_first_layer=False``, which attempted to initialize and remap
+  checkpoint keys through the split-only ``conv_time_spat`` module after the
+  CombinedConv refactor. The unsplit path now initializes and loads its direct
+  temporal convolution as before (:gh:`1207` by `lindicaphxag-tech`_).
+- Make :class:`braindecode.modules.TimeDistributed` accept non-contiguous sequence batches by reshaping rather than requiring view-compatible strides. By `lindicaphxag-tech`_.
+
+- Make :class:`braindecode.modules.Chomp1d` preserve the input when ``chomp_size=0`` instead of returning an empty time axis. This restores :class:`braindecode.models.TCN` with ``kernel_size=1``, whose causal padding is zero. By `lindicaphxag-tech`_.
+
 - Preserve ``Deep4Net``'s first-block temporal stride when using the merged
   ``CombinedConv`` path with ``stride_before_pool=True``; dense-prediction
   outputs now retain the historical temporal geometry (:gh:`1205` by `lindicaphxag-tech`_).
@@ -261,6 +281,12 @@ Bug fixes
   intermediate gradients, and rows at or below ``max_norm`` are unchanged.
   Empty tensors pass through and a negative ``max_norm`` raises, as
   ``Tensor.renorm`` does (:gh:`1187` by `Bruno Aristimunha`_).
+- Keep the channel IDs of :class:`braindecode.models.EEGPT` (``chans_id``) out
+  of the state dict. They are rebuilt from ``chs_info``, and a checkpoint that
+  still stores them no longer overrides them: the released weights now load on
+  any montage and with the default channel projection, where they failed with a
+  size mismatch, and a montage with the same channel count no longer silently
+  takes the checkpoint's IDs (:gh:`1195` by `Bruno Aristimunha`_).
 
 - Fix :class:`braindecode.models.Labram` so the released weights keep their
   pretrained time embedding at every window length: it now holds the original

braindecode-1.8.1.dev172484635/braindecode/version.py

@@ -1 +0,0 @@
-__version__ = "1.8.1.dev172484635"