braindecode 1.8.0.dev580__tar.gz → 1.8.0.dev1058__tar.gz

{braindecode-1.8.0.dev580/braindecode.egg-info → braindecode-1.8.0.dev1058}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: braindecode
-Version: 1.8.0.dev580
+Version: 1.8.0.dev1058
 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-1.8.0.dev580 → braindecode-1.8.0.dev1058}/braindecode/augmentation/base.py

@@ -3,6 +3,7 @@
 #          Bruno Aristimunha <b.aristimunha@gmail.com>
 #          Martin Wimpff <martin.wimpff@iss.uni-stuttgart.de>
 #          Valentin Iovene <val@too.gy>
+#          Sarthak Tayal <sarthaktayal2@gmail.com>
 # License: BSD (3-clause)
 
 from numbers import Real
@@ -176,6 +177,14 @@ class Compose(Transform):
         return X, y
 
 
+def _as_mixed_target(y, lam_dtype):
+    # an untouched target is the same as being mixed with itself with lam of one
+    if isinstance(y, (tuple, list)) and len(y) == 3:
+        return tuple(y)
+    lam = torch.ones(y.shape[0], device=y.device, dtype=lam_dtype)
+    return y, y, lam
+
+
 class _AugmentationCollate:
     """Collate that applies a transform to each batch, with optional expansion.
 
@@ -193,7 +202,10 @@ class _AugmentationCollate:
         ``0`` (default) applies the transform in place (batch size unchanged).
         ``> 0`` keeps the clean originals and appends ``n_augmentation``
         independently transformed copies, returning ``(X, y)`` of
-        ``(1 + n_augmentation)`` times the original size.
+        ``(1 + n_augmentation)`` times the original size. When the transform
+        mixes targets, as :class:`braindecode.augmentation.Mixup` does, ``y``
+        stays the ``(y_a, y_b, lam)`` triple and the clean originals get a
+        mixing coefficient of one.
     """
 
     def __init__(self, transform, device=None, n_augmentation=0):
@@ -216,6 +228,18 @@ class _AugmentationCollate:
             aug_X, aug_y = self.transform(X, y)
             xs.append(aug_X)
             ys.append(aug_y)
+        mixed_ys = [
+            aug_y
+            for aug_y in ys
+            if isinstance(aug_y, (tuple, list)) and len(aug_y) == 3
+        ]
+        if mixed_ys:
+            # a target-mixing transform such as Mixup returns (y_a, y_b, lam),
+            # so the parts are concatenated one by one and the untouched copies
+            # are given a mixing coefficient of one
+            lam_dtype = mixed_ys[0][2].dtype
+            ys = [_as_mixed_target(aug_y, lam_dtype) for aug_y in ys]
+            return torch.cat(xs), tuple(torch.cat(part) for part in zip(*ys))
         return torch.cat(xs), torch.cat(ys)
 
 

{braindecode-1.8.0.dev580 → braindecode-1.8.0.dev1058}/braindecode/augmentation/functional.py

@@ -1065,13 +1065,11 @@ def mixup(
     batch_size, n_channels, n_times = X.shape
 
     X_mix = torch.zeros((batch_size, n_channels, n_times)).to(device)
-    y_a = torch.arange(batch_size).to(device)
-    y_b = torch.arange(batch_size).to(device)
+    y_a = y.clone()
+    y_b = y[idx_perm].clone()
 
     for idx in range(batch_size):
         X_mix[idx] = lam[idx] * X[idx] + (1 - lam[idx]) * X[idx_perm[idx]]
-        y_a[idx] = y[idx]
-        y_b[idx] = y[idx_perm[idx]]
 
     return X_mix, (y_a, y_b, lam)
 

{braindecode-1.8.0.dev580 → braindecode-1.8.0.dev1058}/braindecode/augmentation/transforms.py

@@ -1068,16 +1068,18 @@ class Mixup(Transform):
         device = X.device
         batch_size, _, _ = X.shape
 
+        # lam follows the dtype of X, numpy draws float64 and that would leak
+        # into the mixed signal and into the loss returned by mixup_criterion
         if self.alpha > 0:
             if self.beta_per_sample:
                 lam = torch.as_tensor(
                     self.rng.beta(self.alpha, self.alpha, batch_size)
-                ).to(device)
+                ).to(device=device, dtype=X.dtype)
             else:
-                lam = torch.ones(batch_size).to(device)
+                lam = torch.ones(batch_size, dtype=X.dtype).to(device)
                 lam *= self.rng.beta(self.alpha, self.alpha)
         else:
-            lam = torch.ones(batch_size).to(device)
+            lam = torch.ones(batch_size, dtype=X.dtype).to(device)
 
         idx_perm = torch.as_tensor(
             self.rng.permutation(

{braindecode-1.8.0.dev580 → braindecode-1.8.0.dev1058}/braindecode/datasets/bids/hub.py

@@ -28,6 +28,7 @@ The format follows a BIDS-inspired sourcedata structure:
 # License: BSD (3-clause)
 
 import contextlib
+import copy
 import json
 import logging
 import tempfile
@@ -56,6 +57,7 @@ from .hub_io import (
     _load_eegwindows_from_zarr,
     _load_raw_from_zarr,
     _load_windows_from_zarr,
+    _prepare_info_for_json,
     _save_eegwindows_to_zarr,
     _save_raw_to_zarr,
     _save_windows_to_zarr,
@@ -72,6 +74,33 @@ log = logging.getLogger(__name__)
 _LOCK_FILE = "format_info.json"
 
 
+def _normalize_kwargs_for_json(value, field_name):
+    """Return one preprocessing-kwargs value as native strict JSON."""
+
+    def _convert(obj):
+        if isinstance(obj, np.ndarray):
+            return _convert(obj.tolist())
+        if isinstance(obj, dict):
+            return {key: _convert(item) for key, item in obj.items()}
+        if isinstance(obj, (list, tuple)):
+            return [_convert(item) for item in obj]
+        if isinstance(obj, np.generic):
+            return _convert(obj.item())
+        return obj
+
+    try:
+        converted = _convert(value)
+        if isinstance(converted, str):
+            raise ValueError(
+                "a root string is ambiguous with the legacy encoded format"
+            )
+        return json.loads(json.dumps(converted, allow_nan=False))
+    except (OverflowError, RecursionError, TypeError, ValueError) as error:
+        raise ValueError(
+            f"{field_name} must contain only finite JSON-serializable values"
+        ) from error
+
+
 class HubDatasetMixin:
     """
     Mixin class for Hugging Face Hub integration with EEG datasets.
@@ -685,14 +714,57 @@ class HubDatasetMixin:
                 f"{output_path} already exists. Set overwrite=True to replace it."
             )
 
-        # Create zarr store (zarr v3 API)
-        root = zarr.open(str(output_path), mode="w")
-
         # Validate uniformity across all datasets using shared validation
         dataset_type, _, _ = hub_validation.validate_dataset_uniformity(self.datasets)
 
-        # Keep reference to first dataset for preprocessing kwargs
-        first_ds = self.datasets[0]
+        # Normalize every JSON value before opening the output store. The cached
+        # values below are the exact values handed to the write helpers.
+        prepared_infos = []
+        for ds in self.datasets:
+            if dataset_type == "WindowsDataset":
+                info = ds.windows.info
+            elif dataset_type in ("EEGWindowsDataset", "RawDataset"):
+                info = ds.raw.info
+            prepared_infos.append(_prepare_info_for_json(info.to_json_dict()))
+
+        prepared_kwargs = {}
+        for kwarg_name in [
+            "raw_preproc_kwargs",
+            "window_kwargs",
+            "window_preproc_kwargs",
+        ]:
+            expected_present = hasattr(self.datasets[0], kwarg_name)
+            expected_value = None
+            expected_token = None
+            for i_ds, ds in enumerate(self.datasets):
+                present = hasattr(ds, kwarg_name)
+                if present != expected_present:
+                    raise ValueError(
+                        f"{kwarg_name} on dataset {i_ds} has inconsistent presence; "
+                        "the Zarr format stores one global value"
+                    )
+                if not present:
+                    continue
+                value = _normalize_kwargs_for_json(
+                    getattr(ds, kwarg_name), f"{kwarg_name} on dataset {i_ds}"
+                )
+                value_token = json.dumps(
+                    value, sort_keys=True, separators=(",", ":"), allow_nan=False
+                )
+                if i_ds == 0:
+                    expected_value = value
+                    expected_token = value_token
+                elif value_token != expected_token:
+                    raise ValueError(
+                        f"{kwarg_name} on dataset {i_ds} differs from dataset 0; "
+                        "the Zarr format stores one global value"
+                    )
+            if expected_present:
+                prepared_kwargs[kwarg_name] = expected_value
+
+        # Create compressor and zarr store (zarr v3 API) only after preflight.
+        compressor = _create_compressor(compression, compression_level)
+        root = zarr.open(str(output_path), mode="w")
 
         # Store global metadata
         root.attrs["n_datasets"] = len(self.datasets)
@@ -706,21 +778,10 @@ class HubDatasetMixin:
         root.attrs["zarr_version"] = zarr.__version__
         root.attrs["scipy_version"] = scipy.__version__
 
-        # Save preprocessing kwargs (check first dataset, assuming uniform preprocessing)
-        # These are typically set by windowing functions on individual datasets
-        for kwarg_name in [
-            "raw_preproc_kwargs",
-            "window_kwargs",
-            "window_preproc_kwargs",
-        ]:
-            # Check first dataset for these attributes
-            if hasattr(first_ds, kwarg_name):
-                kwargs = getattr(first_ds, kwarg_name)
-                if kwargs:
-                    root.attrs[kwarg_name] = json.dumps(kwargs)
-
-        # Create compressor
-        compressor = _create_compressor(compression, compression_level)
+        # Save preprocessing kwargs from the preflight cache. These are
+        # typically set by windowing functions on individual datasets.
+        for kwarg_name, kwargs in prepared_kwargs.items():
+            root.attrs[kwarg_name] = kwargs
 
         # Save each recording
         for i_ds, ds in enumerate(self.datasets):
@@ -731,7 +792,7 @@ class HubDatasetMixin:
                 data = ds.windows.get_data()
                 metadata = ds.windows.metadata
                 description = ds.description
-                info_dict = ds.windows.info.to_json_dict()
+                info_dict = prepared_infos[i_ds]
                 target_name = ds.target_name if hasattr(ds, "target_name") else None
 
                 # Save using inlined function
@@ -750,7 +811,7 @@ class HubDatasetMixin:
                 raw = ds.raw
                 metadata = ds.metadata
                 description = ds.description
-                info_dict = ds.raw.info.to_json_dict()
+                info_dict = prepared_infos[i_ds]
                 targets_from = ds.targets_from
                 last_target_only = ds.last_target_only
 
@@ -771,7 +832,7 @@ class HubDatasetMixin:
                 # Get continuous raw data from RawDataset
                 raw = ds.raw
                 description = ds.description
-                info_dict = ds.raw.info.to_json_dict()
+                info_dict = prepared_infos[i_ds]
                 target_name = ds.target_name if hasattr(ds, "target_name") else None
 
                 # Save using inlined function
@@ -956,10 +1017,14 @@ class HubDatasetMixin:
             "window_preproc_kwargs",
         ]:
             if kwarg_name in root.attrs:
-                kwargs = json.loads(root.attrs[kwarg_name])
+                kwargs = root.attrs[kwarg_name]
+                if isinstance(kwargs, str):
+                    # Stores written by older braindecode versions kept
+                    # these attributes as double-encoded JSON strings.
+                    kwargs = json.loads(kwargs)
                 # Set on each individual dataset (where they were originally stored)
                 for ds in datasets:
-                    setattr(ds, kwarg_name, kwargs)
+                    setattr(ds, kwarg_name, copy.deepcopy(kwargs))
 
         return concat_ds
 

{braindecode-1.8.0.dev580 → braindecode-1.8.0.dev1058}/braindecode/datasets/bids/hub_io.py

@@ -8,6 +8,7 @@ These functions keep the Zarr serialization details isolated from hub.py.
 from __future__ import annotations
 
 import json
+from numbers import Real
 from pathlib import Path
 
 import numpy as np
@@ -17,17 +18,78 @@ from mne.utils import _soft_import
 zarr = _soft_import("zarr", purpose="hugging face integration", strict=False)
 
 
+def _is_non_bool_real(value):
+    return not isinstance(value, (bool, np.bool_)) and isinstance(
+        value, (Real, np.integer, np.floating)
+    )
+
+
+def _prepare_info_for_json(obj, path="info", *, _in_numeric_sequence=False):
+    """Normalize an MNE Info value to strict JSON without losing sequence NaNs."""
+    if isinstance(obj, np.ndarray):
+        obj = obj.tolist()
+
+    if isinstance(obj, dict):
+        return {
+            key: _prepare_info_for_json(value, f"{path}.{key}")
+            for key, value in obj.items()
+        }
+
+    if isinstance(obj, (list, tuple)):
+        values = list(obj)
+        has_none = any(value is None for value in values)
+        has_number = any(_is_non_bool_real(value) for value in values)
+        all_none = bool(values) and all(value is None for value in values)
+        if all_none or (has_none and has_number):
+            raise ValueError(
+                f"{path} is ambiguous: numeric sequences cannot contain JSON null"
+            )
+
+        is_numeric_sequence = bool(values) and all(
+            _is_non_bool_real(value) for value in values
+        )
+        return [
+            _prepare_info_for_json(
+                value,
+                f"{path}[{index}]",
+                _in_numeric_sequence=is_numeric_sequence,
+            )
+            for index, value in enumerate(values)
+        ]
+
+    if isinstance(obj, (bool, np.bool_)):
+        return bool(obj)
+    if isinstance(obj, (int, np.integer)):
+        return int(obj)
+    if _is_non_bool_real(obj):
+        value = float(obj)
+        if np.isnan(value):
+            if _in_numeric_sequence:
+                return None
+            raise ValueError(f"{path} contains unsupported NaN")
+        if np.isposinf(value):
+            raise ValueError(f"{path} contains positive infinity")
+        if np.isneginf(value):
+            raise ValueError(f"{path} contains negative infinity")
+        return value
+    if obj is None or isinstance(obj, str):
+        return obj
+    raise ValueError(f"{path} contains a non-JSON-serializable value")
+
+
 def _restore_nan_from_json(obj):
-    """Restore NaN values from None in legacy zarr stores.
+    """Restore NaN values from None in JSON-loaded attributes.
 
-    Datasets saved before zarr v3 native NaN support used
-    ``_sanitize_for_json`` to convert NaN/Inf → None. This restores them
-    on load so ``mne.Info.from_json_dict`` gets proper NaN arrays.
+    JSON null is reserved for NaN only in non-boolean numeric sequences.
+    All-null lists are therefore the representation of validated all-NaN
+    sequences, while ordinary null values elsewhere remain unchanged.
     """
     if isinstance(obj, dict):
         return {k: _restore_nan_from_json(v) for k, v in obj.items()}
     if isinstance(obj, list):
-        if len(obj) > 0 and all(isinstance(x, (int, float, type(None))) for x in obj):
+        if len(obj) > 0 and all(
+            value is None or _is_non_bool_real(value) for value in obj
+        ):
             return [np.nan if x is None else x for x in obj]
         return [_restore_nan_from_json(v) for v in obj]
     return obj

{braindecode-1.8.0.dev580 → braindecode-1.8.0.dev1058}/braindecode/models/base.py

@@ -198,7 +198,34 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
     See :ref:`load-pretrained-models` for a complete tutorial.
     """
 
+    #: Attributes that :func:`torch.jit.script` must not introspect. The
+    #: signal-related properties raise :class:`ValueError` when their value was
+    #: neither given nor inferable, and ``mapping`` carries a postponed
+    #: annotation TorchScript cannot resolve. Either one aborts scripting
+    #: before ``forward`` is ever compiled.
+    __jit_ignored_attributes__ = [
+        *sorted(_EEG_PARAMS),
+        "_chs_info",
+        "input_shape",
+        "mapping",
+    ]
+    #: Rich MNE channel dictionaries are intentionally unavailable in scripted
+    #: forwards; they remain unchanged on eager models and in saved configs.
+    __jit_unused_properties__ = ["chs_info"]
+
     def __init_subclass__(cls, **kwargs):
+        # TorchScript only honours ``__jit_ignored_attributes__`` for
+        # properties defined directly on the concrete class: it collects them
+        # with ``vars(type(module))``, which skips inherited ones. Rebinding
+        # the very same property objects on each subclass makes them visible
+        # there without changing any runtime behaviour.
+        for name in cls.__jit_ignored_attributes__:
+            if name in cls.__dict__:
+                continue
+            prop = getattr(cls, name, None)
+            if isinstance(prop, property):
+                setattr(cls, name, prop)
+
         # Append model-specific Hub integration notes to the docstring.
         # This runs before the metaclass __init__, so the Hub notes will
         # be included in the docstring that the metaclass processes.
@@ -286,6 +313,12 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
         self._chs_info = chs_info  # type: ignore[assignment]
         self._n_outputs = n_outputs  # type: ignore[assignment]
         self._n_chans = n_chans  # type: ignore[assignment]
+        # TorchScript cannot represent the rich MNE dictionaries in _chs_info.
+        # Keep the original eager state above and expose only the derived scalar
+        # to scripted signal-property getters.
+        self._n_chans_for_jit = (
+            len(chs_info) if n_chans is None and chs_info is not None else n_chans
+        )
         self._n_times = n_times  # type: ignore[assignment]
         self._sfreq = sfreq  # type: ignore[assignment]
 
@@ -310,6 +343,12 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
 
     @property
     def n_chans(self) -> int:
+        if torch.jit.is_scripting():
+            if self._n_chans_for_jit is None:
+                raise ValueError(
+                    "n_chans could not be inferred. Either specify n_chans or chs_info."
+                )
+            return self._n_chans_for_jit
         if self._n_chans is None and self._chs_info is not None:
             return len(self._chs_info)
         elif self._n_chans is None:
@@ -319,7 +358,7 @@ class EEGModuleMixin(_BaseHubMixin, metaclass=_BraindecodeDocstringMeta):
         return self._n_chans
 
     @property
-    def chs_info(self) -> list[str]:
+    def chs_info(self) -> list[dict]:
         if self._chs_info is None:
             raise ValueError("chs_info not specified.")
         return self._chs_info