diff --git a/.github/workflows/auto-label.yml b/.github/workflows/auto-label.yml
new file mode 100644
index 00000000..639f3264
--- /dev/null
+++ b/.github/workflows/auto-label.yml
@@ -0,0 +1,15 @@
+name: Auto Label
+
+on:
+  pull_request:
+
+jobs:
+  label:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Label pre-commit PR
+        if: github.actor == 'pre-commit-ci[bot]'
+        uses: actions-ecosystem/action-add-labels@v1
+        with:
+          github_token: ${{ secrets.GITHUB_TOKEN }}
+          labels: pre-commit
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 110f6bb8..ea0739d3 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -87,9 +87,39 @@ jobs:
         uses: actions/checkout@v4
       - name: Build Changelog
         id: changelog
-        uses: mikepenz/release-changelog-builder-action@v4
+        uses: mikepenz/release-changelog-builder-action@v6
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          configurationJson: |
+            {
+              "template": "# Changelog\n\n{{CHANGELOG}}\n\n<details>\n<summary>📦 Other changes</summary>\n\n{{UNCATEGORIZED}}\n\n</details>",
+              "categories": [
+                {
+                  "title": "🚀 Features",
+                  "labels": ["feature","enhancement"]
+                },
+                {
+                  "title": "🐛 Bug Fixes",
+                  "labels": ["bug","fix"]
+                },
+                {
+                  "title": "📚 Documentation",
+                  "labels": ["documentation","docs"]
+                },
+                {
+                  "title": "⬆️ Dependencies",
+                  "labels": ["dependencies"]
+                },
+                {
+                  "title": "🧰 Maintenance",
+                  "labels": ["chore","ci"]
+                }
+              ],
+              "exclude_labels": [
+                "pre-commit"
+              ]
+            }
       - name: Create Release
         id: create_release
         uses: softprops/action-gh-release@v2
diff --git a/CHANGELOG.rst b/CHANGELOG.rst
index a4d9a63a..d0fbf3b6 100644
--- a/CHANGELOG.rst
+++ b/CHANGELOG.rst
@@ -14,9 +14,17 @@ Versioning <https://semver.org/spec/v2.0.0.html>`__.
 Added
 ~~~~~
 
+- Implement a generic `Registry` class and establish registries for backbones, models,
+  attention layers, preprocessors, augmenters, optimizers, schedulers, and losses.
+- Add `forward_features` and `compute_features_output_shape` methods to all CNN
+  backbones to provide a standardized API for SSL and feature extraction.
+
 Changed
 ~~~~~~~
 
+- Refactor `ECG_CRNN`, `PreprocManager`, `AugmenterManager`, and `BaseTrainer` to
+  utilize the new registry system for dynamic component construction and decoupling.
+- Enhance `SizeMixin` to support static shape inference for feature maps.
 - Make the function `remove_spikes_naive` in `torch_ecg.utils.utils_signal`
   support 2D and 3D input signals.
 - Use `save_file` and `load_file` from the `safetensors` package for saving
@@ -36,6 +44,13 @@ Removed
 Fixed
 ~~~~~
 
+- Robustly handle dimension inference and initialization in `ECG_CRNN` models,
+  especially for cases with `None` or `Identity` modules.
+- Address several compatibility issues for Python 3.13, including docstring
+  indentation and `NaN` comparisons in dataclasses.
+- Improve error handling and encoding robustness in `CitationMixin` when
+  reading cache files.
+- Resolve CodeQL warnings regarding incomplete URL substring sanitization in tests.
 - Correctly update the `_df_metadata` attribute of the `PTBXL` database reader
   classes after filtering records.
 - Enhance the `save` method of the `torch_ecg.utils.utils_nn.CkptMixin` class:
diff --git a/pyproject.toml b/pyproject.toml
index 1be2b9df..23b4fd33 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -158,6 +158,10 @@ omit = [
     "torch_ecg/models/grad_cam.py",
     # temporarily ignore torch_ecg/ssl since it's not implemented
     "torch_ecg/ssl/*",
+    # temporarily ignore models that are not implemented completely
+    "torch_ecg/models/cnn/darknet.py",
+    "torch_ecg/models/cnn/efficientnet.py",
+    "torch_ecg/models/cnn/ho_resnet.py",
 ]
 exclude_also = [
     "raise NotImplementedError",
diff --git a/test/test_models/test_backbone_api.py b/test/test_models/test_backbone_api.py
new file mode 100644
index 00000000..7ab4fcba
--- /dev/null
+++ b/test/test_models/test_backbone_api.py
@@ -0,0 +1,82 @@
+"""
+Unit tests for the standardized Backbone API.
+"""
+
+from copy import deepcopy
+
+import pytest
+import torch
+
+from torch_ecg.model_configs import ECG_CRNN_CONFIG
+from torch_ecg.models.registry import BACKBONES
+
+DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# Extract all valid backbone configurations from the central config
+BACKBONE_CONFIGS = []
+for config_key, config_val in ECG_CRNN_CONFIG.cnn.items():
+    if not isinstance(config_val, dict):
+        continue
+
+    # 1. Try to get name from the config dict
+    # 2. If not in dict, use the key if it's in the registry
+    # 3. If key contains a known registry name (e.g. resnet_nature_comm), extract the base name
+    backbone_name = config_val.get("name")
+    if backbone_name is None:
+        if config_key in BACKBONES:
+            backbone_name = config_key
+        else:
+            # Check if config_key contains any registered name as a prefix
+            for registered_name in BACKBONES.list_all():
+                if config_key.startswith(registered_name):
+                    backbone_name = registered_name
+                    break
+
+    if backbone_name:
+        BACKBONE_CONFIGS.append((backbone_name, config_key, config_val))
+
+
+@pytest.mark.parametrize("backbone_name, config_key, config", BACKBONE_CONFIGS)
+def test_backbone_api(backbone_name, config_key, config):
+    n_leads = 12
+    batch_size = 2
+    seq_len = 2000
+
+    # Skip models that are not implemented yet to avoid noisy test failures
+    # These will be implemented in Phase 1.5
+    try:
+        model = BACKBONES.build(backbone_name, in_channels=n_leads, **deepcopy(config)).to(DEVICE)
+    except NotImplementedError:
+        pytest.skip(f"Backbone {backbone_name} (config: {config_key}) is not implemented yet.")
+    except Exception as e:
+        pytest.fail(f"Failed to build backbone {backbone_name} with config {config_key}: {e}")
+
+    model.eval()
+    inp = torch.randn(batch_size, n_leads, seq_len).to(DEVICE)
+
+    # 1. Test forward_features existence
+    assert hasattr(model, "forward_features"), f"Backbone {backbone_name} missing forward_features"
+
+    # 2. Test forward_features output shape
+    features = model.forward_features(inp)
+    assert features.ndim == 3, f"Backbone {backbone_name} forward_features should return 3D tensor, got {features.ndim}D"
+    assert features.shape[0] == batch_size
+
+    # 3. Test compute_features_output_shape consistency
+    # All Backbones in torch_ecg follow (seq_len, batch_size) signature
+    expected_shape = model.compute_features_output_shape(seq_len, batch_size)
+    assert (
+        features.shape[1] == expected_shape[1]
+    ), f"Backbone {backbone_name} feature channels mismatch: {features.shape[1]} vs {expected_shape[1]}"
+    if expected_shape[2] is not None:
+        assert (
+            features.shape[2] == expected_shape[2]
+        ), f"Backbone {backbone_name} feature seq_len mismatch: {features.shape[2]} vs {expected_shape[2]}"
+
+    # 4. Test forward consistency (if model is pure feature extractor)
+    out = model(inp)
+    assert torch.allclose(out, features), f"Backbone {backbone_name} forward and forward_features results differ"
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])
diff --git a/torch_ecg/models/cnn/darknet.py b/torch_ecg/models/cnn/darknet.py
index 3b970fbb..0d4a68d5 100644
--- a/torch_ecg/models/cnn/darknet.py
+++ b/torch_ecg/models/cnn/darknet.py
@@ -10,19 +10,15 @@
 5. Wang, C. Y., Bochkovskiy, A., & Liao, H. Y. M. (2020). Scaled-YOLOv4: Scaling Cross Stage Partial Network. arXiv preprint arXiv:2011.08036.
 """
 
-from typing import List
+from typing import List, Optional, Sequence, Union
 
-from torch import nn
+from torch import Tensor, nn
 
 from ...models._nets import Conv_Bn_Activation, DownSample, GlobalContextBlock, NonLocalBlock, SEBlock  # noqa: F401
 from ...utils import CitationMixin, SizeMixin
 
-__all__ = [
-    "DarkNet",
-]
 
-
-class DarkNet(nn.Sequential, SizeMixin, CitationMixin):
+class DarkNet(SizeMixin, nn.Sequential, CitationMixin):
     """ """
 
     __name__ = "DarkNet"
@@ -32,6 +28,22 @@ def __init__(self, in_channels: int, **config) -> None:
         super().__init__()
         raise NotImplementedError
 
+    def compute_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the model."""
+        raise NotImplementedError
+
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features."""
+        raise NotImplementedError
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features."""
+        raise NotImplementedError
+
     @property
     def doi(self) -> List[str]:
         return list(set(self.config.get("doi", []) + ["10.1109/CVPR.2016.91"]))
diff --git a/torch_ecg/models/cnn/densenet.py b/torch_ecg/models/cnn/densenet.py
index 5fa89995..d0329bee 100644
--- a/torch_ecg/models/cnn/densenet.py
+++ b/torch_ecg/models/cnn/densenet.py
@@ -760,6 +760,44 @@ def compute_output_shape(
         """Compute the output shape of the network."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
diff --git a/torch_ecg/models/cnn/efficientnet.py b/torch_ecg/models/cnn/efficientnet.py
index a54b19b5..caf5bae4 100644
--- a/torch_ecg/models/cnn/efficientnet.py
+++ b/torch_ecg/models/cnn/efficientnet.py
@@ -9,17 +9,13 @@
 
 """
 
-from typing import List
+from typing import List, Optional, Sequence, Union
 
-from torch import nn
+from torch import Tensor, nn
 
 from ...models._nets import Conv_Bn_Activation, DownSample, GlobalContextBlock, NonLocalBlock, SEBlock  # noqa: F401
 from ...utils import CitationMixin, SizeMixin
 
-__all__ = [
-    "EfficientNet",
-]
-
 
 class EfficientNet(nn.Module, SizeMixin, CitationMixin):
     """
@@ -38,10 +34,22 @@ def __init__(self, in_channels: int, **config) -> None:
         super().__init__()
         raise NotImplementedError
 
-    def forward(self):
+    def forward(self, input: Tensor) -> Tensor:
+        raise NotImplementedError
+
+    def compute_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        raise NotImplementedError
+
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features."""
         raise NotImplementedError
 
-    def compute_output_shape(self):
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features."""
         raise NotImplementedError
 
     @property
@@ -67,8 +75,20 @@ def __init__(self, in_channels: int, **config) -> None:
         super().__init__()
         raise NotImplementedError
 
-    def forward(self):
+    def forward(self, input: Tensor) -> Tensor:
+        raise NotImplementedError
+
+    def compute_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        raise NotImplementedError
+
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features."""
         raise NotImplementedError
 
-    def compute_output_shape(self):
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features."""
         raise NotImplementedError
diff --git a/torch_ecg/models/cnn/ho_resnet.py b/torch_ecg/models/cnn/ho_resnet.py
index 8f0f485c..deb5ab95 100644
--- a/torch_ecg/models/cnn/ho_resnet.py
+++ b/torch_ecg/models/cnn/ho_resnet.py
@@ -7,7 +7,9 @@
 
 """
 
-from torch import nn
+from typing import Optional, Sequence, Union
+
+from torch import Tensor, nn
 
 from ...cfg import CFG  # noqa: F401
 from ...models._nets import (  # noqa: F401
@@ -33,6 +35,26 @@ class MidPointResNet(nn.Module, SizeMixin, CitationMixin):
 
     def __init__(self, in_channels: int, **config) -> None:
         """ """
+        super().__init__()
+        raise NotImplementedError
+
+    def forward(self, input: Tensor) -> Tensor:
+        raise NotImplementedError
+
+    def compute_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the model."""
+        raise NotImplementedError
+
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features."""
+        raise NotImplementedError
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features."""
         raise NotImplementedError
 
 
@@ -41,6 +63,26 @@ class RK4ResNet(nn.Module, SizeMixin, CitationMixin):
 
     def __init__(self, in_channels: int, **config) -> None:
         """ """
+        super().__init__()
+        raise NotImplementedError
+
+    def forward(self, input: Tensor) -> Tensor:
+        raise NotImplementedError
+
+    def compute_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the model."""
+        raise NotImplementedError
+
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features."""
+        raise NotImplementedError
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features."""
         raise NotImplementedError
 
 
@@ -49,4 +91,24 @@ class RK8ResNet(nn.Module, SizeMixin, CitationMixin):
 
     def __init__(self, in_channels: int, **config) -> None:
         """ """
+        super().__init__()
+        raise NotImplementedError
+
+    def forward(self, input: Tensor) -> Tensor:
+        raise NotImplementedError
+
+    def compute_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the model."""
+        raise NotImplementedError
+
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features."""
+        raise NotImplementedError
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features."""
         raise NotImplementedError
diff --git a/torch_ecg/models/cnn/mobilenet.py b/torch_ecg/models/cnn/mobilenet.py
index cd6e476b..f50a6ba8 100644
--- a/torch_ecg/models/cnn/mobilenet.py
+++ b/torch_ecg/models/cnn/mobilenet.py
@@ -414,6 +414,44 @@ def compute_output_shape(
             _, _, _seq_len = output_shape
         return output_shape
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
@@ -829,6 +867,44 @@ def compute_output_shape(
         """Compute the output shape of the model."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
@@ -1253,6 +1329,44 @@ def compute_output_shape(
         """Compute the output shape of the model."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
diff --git a/torch_ecg/models/cnn/multi_scopic.py b/torch_ecg/models/cnn/multi_scopic.py
index e999a937..4be1144e 100644
--- a/torch_ecg/models/cnn/multi_scopic.py
+++ b/torch_ecg/models/cnn/multi_scopic.py
@@ -480,6 +480,44 @@ def compute_output_shape(
         output_shape = (batch_size, out_channels, _seq_len)
         return output_shape
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     def assign_weights_lead_wise(self, other: "MultiScopicCNN", indices: Sequence[int]) -> None:
         """Assign weights to the `other` :class:`MultiScopicCNN`
         module in the lead-wise manner
diff --git a/torch_ecg/models/cnn/regnet.py b/torch_ecg/models/cnn/regnet.py
index 0f3cab31..650989f9 100644
--- a/torch_ecg/models/cnn/regnet.py
+++ b/torch_ecg/models/cnn/regnet.py
@@ -11,7 +11,7 @@
 from typing import List, Optional, Sequence, Union
 
 import torch
-from torch import nn
+from torch import Tensor, nn
 
 from ...cfg import CFG
 from ...models._nets import Conv_Bn_Activation, DownSample, SpaceToDepth
@@ -512,6 +512,44 @@ def compute_output_shape(
         """Compute the output shape of the network."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
diff --git a/torch_ecg/models/cnn/resnet.py b/torch_ecg/models/cnn/resnet.py
index 314ed107..b7b72dd7 100644
--- a/torch_ecg/models/cnn/resnet.py
+++ b/torch_ecg/models/cnn/resnet.py
@@ -858,6 +858,44 @@ def compute_output_shape(
         """Compute the output shape of the model."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
diff --git a/torch_ecg/models/cnn/vgg.py b/torch_ecg/models/cnn/vgg.py
index 9157ef15..00f2069b 100644
--- a/torch_ecg/models/cnn/vgg.py
+++ b/torch_ecg/models/cnn/vgg.py
@@ -4,7 +4,7 @@
 from copy import deepcopy
 from typing import List, Optional, Sequence, Union
 
-from torch import nn
+from torch import Tensor, nn
 
 from ...cfg import CFG
 from ...models._nets import Conv_Bn_Activation
@@ -189,6 +189,44 @@ def compute_output_shape(
         """Compute the output shape of the module."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
diff --git a/torch_ecg/models/cnn/xception.py b/torch_ecg/models/cnn/xception.py
index e49e1dab..85308e6b 100644
--- a/torch_ecg/models/cnn/xception.py
+++ b/torch_ecg/models/cnn/xception.py
@@ -888,6 +888,44 @@ def compute_output_shape(
         """Compute the output shape the model."""
         return compute_sequential_output_shape(self, seq_len, batch_size)
 
+    def forward_features(self, input: Tensor) -> Tensor:
+        """Forward pass of the model to extract features.
+
+        Parameters
+        ----------
+        input : torch.Tensor
+            Input signal tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        Returns
+        -------
+        features : torch.Tensor
+            Feature map tensor,
+            of shape ``(batch_size, channels, seq_len)``.
+
+        """
+        return self.forward(input)
+
+    def compute_features_output_shape(
+        self, seq_len: Optional[int] = None, batch_size: Optional[int] = None
+    ) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        Parameters
+        ----------
+        seq_len : int, optional
+            Length of the input signal tensor.
+        batch_size : int, optional
+            Batch size of the input signal tensor.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(seq_len, batch_size)
+
     @property
     def in_channels(self) -> int:
         return self.__in_channels
diff --git a/torch_ecg/utils/utils_nn.py b/torch_ecg/utils/utils_nn.py
index 63f84e53..d01ec960 100644
--- a/torch_ecg/utils/utils_nn.py
+++ b/torch_ecg/utils/utils_nn.py
@@ -1073,6 +1073,28 @@ def dtype_(self) -> str:
     def device_(self) -> str:
         return str(self.device)
 
+    def compute_features_output_shape(self, *args: Any, **kwargs: Any) -> Sequence[Union[int, None]]:
+        """Compute the output shape of the features.
+
+        By default, this is the same as the output shape of the model.
+        For backbones with pooling and classification heads, this should be
+        overridden to return the shape of the features before global pooling.
+
+        Parameters
+        ----------
+        *args : Any
+            Positional arguments passed to `compute_output_shape`.
+        **kwargs : Any
+            Keyword arguments passed to `compute_output_shape`.
+
+        Returns
+        -------
+        output_shape : sequence
+            Output shape of the features.
+
+        """
+        return self.compute_output_shape(*args, **kwargs)
+
 
 def make_safe_globals(obj: Any, remove_paths: bool = True) -> Any:
     """Make a dictionary or a dictionary-like object safe for serialization.