Upload featurizers.py

featurizers.py

@@ -0,0 +1,479 @@
+"""
+featurizers.py
+==============
+Utility classes for defining *invertible* feature spaces on top of a model’s
+hidden-state tensors, together with intervention helpers that operate inside
+those spaces.
+
+Key ideas
+---------
+
+* **Featurizer** – a lightweight wrapper holding:
+    • a forward `featurizer` module that maps a tensor **x → (f, error)**  
+      where *error* is the reconstruction residual (useful for lossy
+      featurizers such as sparse auto-encoders);  
+    • an `inverse_featurizer` that re-assembles the original space  
+      **(f, error) → x̂**.
+
+* **Interventions** – three higher-order factory functions build PyVENE
+  interventions that work in the featurized space:
+    - *interchange*
+    - *collect*
+    - *mask* (differential binary masking)
+
+All public classes / functions below carry PEP-257-style doc-strings.
+"""
+
+from typing import Optional, Tuple
+
+import torch
+import pyvene as pv
+
+
+# --------------------------------------------------------------------------- #
+#  Basic identity featurizers                                                 #
+# --------------------------------------------------------------------------- #
+class IdentityFeaturizerModule(torch.nn.Module):
+    """A no-op featurizer: *x → (x, None)*."""
+
+    def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, None]:
+        return x, None
+
+
+class IdentityInverseFeaturizerModule(torch.nn.Module):
+    """Inverse of :class:`IdentityFeaturizerModule`."""
+
+    def forward(self, x: torch.Tensor, error: None) -> torch.Tensor:  # noqa: D401
+        return x
+
+
+# --------------------------------------------------------------------------- #
+#  High-level Featurizer wrapper                                              #
+# --------------------------------------------------------------------------- #
+class Featurizer:
+    """Container object holding paired featurizer and inverse modules.
+
+    Parameters
+    ----------
+    featurizer :
+        A `torch.nn.Module` mapping **x → (features, error)**.
+    inverse_featurizer :
+        A `torch.nn.Module` mapping **(features, error) → x̂**.
+    n_features :
+        Dimensionality of the feature space.  **Required** when you intend to
+        build a *mask* intervention; optional otherwise.
+    id :
+        Human-readable identifier used by `__str__` methods of the generated
+        interventions.
+    """
+
+    # --------------------------------------------------------------------- #
+    #  Construction / public accessors                                      #
+    # --------------------------------------------------------------------- #
+    def __init__(
+        self,
+        featurizer: torch.nn.Module = IdentityFeaturizerModule(),
+        inverse_featurizer: torch.nn.Module = IdentityInverseFeaturizerModule(),
+        *,
+        n_features: Optional[int] = None,
+        id: str = "null",
+    ):
+        self.featurizer = featurizer
+        self.inverse_featurizer = inverse_featurizer
+        self.n_features = n_features
+        self.id = id
+
+    # -------------------- Intervention builders -------------------------- #
+    def get_interchange_intervention(self):
+        if not hasattr(self, "_interchange_intervention"):
+            self._interchange_intervention = build_feature_interchange_intervention(
+                self.featurizer, self.inverse_featurizer, self.id
+            )
+        return self._interchange_intervention
+
+    def get_collect_intervention(self):
+        if not hasattr(self, "_collect_intervention"):
+            self._collect_intervention = build_feature_collect_intervention(
+                self.featurizer, self.id
+            )
+        return self._collect_intervention
+
+    def get_mask_intervention(self):
+        if self.n_features is None:
+            raise ValueError(
+                "`n_features` must be provided on the Featurizer "
+                "to construct a mask intervention."
+            )
+        if not hasattr(self, "_mask_intervention"):
+            self._mask_intervention = build_feature_mask_intervention(
+                self.featurizer,
+                self.inverse_featurizer,
+                self.n_features,
+                self.id,
+            )
+        return self._mask_intervention
+
+    # ------------------------- Convenience I/O --------------------------- #
+    def featurize(self, x: torch.Tensor):
+        return self.featurizer(x)
+
+    def inverse_featurize(self, x: torch.Tensor, error):
+        return self.inverse_featurizer(x, error)
+
+    # --------------------------------------------------------------------- #
+    #  (De)serialisation helpers                                            #
+    # --------------------------------------------------------------------- #
+    def save_modules(self, path: str) -> Tuple[str, str]:
+        """Serialise featurizer & inverse to `<path>_{featurizer, inverse}`.
+
+        Notes
+        -----
+        * **SAE featurizers** are *not* serialisable: a
+          :class:`NotImplementedError` is raised.
+        * Existing files will be *silently overwritten*.
+        """
+        featurizer_class = self.featurizer.__class__.__name__
+
+        if featurizer_class == "SAEFeaturizerModule":
+            #SAE featurizers are to be loaded from sae_lens
+            return None, None
+
+        inverse_featurizer_class = self.inverse_featurizer.__class__.__name__
+
+        # Extra config needed for Subspace featurizers
+        additional_config = {}
+        if featurizer_class == "SubspaceFeaturizerModule":
+            additional_config["rotation_matrix"] = (
+                self.featurizer.rotate.weight.detach().clone()
+            )
+            additional_config["requires_grad"] = (
+                self.featurizer.rotate.weight.requires_grad
+            )
+
+        model_info = {
+            "featurizer_class": featurizer_class,
+            "inverse_featurizer_class": inverse_featurizer_class,
+            "n_features": self.n_features,
+            "additional_config": additional_config,
+        }
+
+        torch.save(
+            {"model_info": model_info, "state_dict": self.featurizer.state_dict()},
+            f"{path}_featurizer",
+        )
+        torch.save(
+            {
+                "model_info": model_info,
+                "state_dict": self.inverse_featurizer.state_dict(),
+            },
+            f"{path}_inverse_featurizer",
+        )
+        return f"{path}_featurizer", f"{path}_inverse_featurizer"
+
+    @classmethod
+    def load_modules(cls, path: str) -> "Featurizer":
+        """Inverse of :meth:`save_modules`.
+
+        Returns
+        -------
+        Featurizer
+            A *new* instance with reconstructed modules and metadata.
+        """
+        featurizer_data = torch.load(f"{path}_featurizer")
+        inverse_data = torch.load(f"{path}_inverse_featurizer")
+
+        model_info = featurizer_data["model_info"]
+        featurizer_class = model_info["featurizer_class"]
+
+        if featurizer_class == "SubspaceFeaturizerModule":
+            rot = model_info["additional_config"]["rotation_matrix"]
+            requires_grad = model_info["additional_config"]["requires_grad"]
+
+            # Re-build a parametrised orthogonal layer with identical shape.
+            in_dim, out_dim = rot.shape
+            rotate_layer = pv.models.layers.LowRankRotateLayer(
+                in_dim, out_dim, init_orth=False
+            )
+            rotate_layer.weight.data.copy_(rot)
+            rotate_layer = torch.nn.utils.parametrizations.orthogonal(rotate_layer)
+            rotate_layer.requires_grad_(requires_grad)
+
+            featurizer = SubspaceFeaturizerModule(rotate_layer)
+            inverse = SubspaceInverseFeaturizerModule(rotate_layer)
+
+            # Sanity-check weight shape
+            assert (
+                featurizer.rotate.weight.shape == rot.shape
+            ), "Rotation-matrix shape mismatch after deserialisation."
+        elif featurizer_class == "IdentityFeaturizerModule":
+            featurizer = IdentityFeaturizerModule()
+            inverse = IdentityInverseFeaturizerModule()
+        else:
+            raise ValueError(f"Unknown featurizer class '{featurizer_class}'.")
+
+        featurizer.load_state_dict(featurizer_data["state_dict"])
+        inverse.load_state_dict(inverse_data["state_dict"])
+
+        return cls(
+            featurizer,
+            inverse,
+            n_features=model_info["n_features"],
+            id=model_info.get("featurizer_id", "loaded"),
+        )
+
+
+# --------------------------------------------------------------------------- #
+#  Intervention factory helpers                                               #
+# --------------------------------------------------------------------------- #
+def build_feature_interchange_intervention(
+    featurizer: torch.nn.Module,
+    inverse_featurizer: torch.nn.Module,
+    featurizer_id: str,
+):
+    """Return a class implementing PyVENE’s TrainableIntervention."""
+
+    class FeatureInterchangeIntervention(
+        pv.TrainableIntervention, pv.DistributedRepresentationIntervention
+    ):
+        """Swap features between *base* and *source* in the featurized space."""
+
+        def __init__(self, **kwargs):
+            super().__init__(**kwargs)
+            self._featurizer = featurizer
+            self._inverse = inverse_featurizer
+
+        def forward(self, base, source, subspaces=None):
+            f_base, base_err = self._featurizer(base)
+            f_src, _ = self._featurizer(source)
+
+            if subspaces is None or _subspace_is_all_none(subspaces):
+                f_out = f_src
+            else:
+                f_out = pv.models.intervention_utils._do_intervention_by_swap(
+                    f_base,
+                    f_src,
+                    "interchange",
+                    self.interchange_dim,
+                    subspaces,
+                    subspace_partition=self.subspace_partition,
+                    use_fast=self.use_fast,
+                )
+            return self._inverse(f_out, base_err).to(base.dtype)
+
+        def __str__(self):  # noqa: D401
+            return f"FeatureInterchangeIntervention(id={featurizer_id})"
+
+    return FeatureInterchangeIntervention
+
+
+def build_feature_collect_intervention(
+    featurizer: torch.nn.Module, featurizer_id: str
+):
+    """Return a `CollectIntervention` operating in feature space."""
+
+    class FeatureCollectIntervention(pv.CollectIntervention):
+        def __init__(self, **kwargs):
+            super().__init__(**kwargs)
+            self._featurizer = featurizer
+
+        def forward(self, base, source=None, subspaces=None):
+            f_base, _ = self._featurizer(base)
+            return pv.models.intervention_utils._do_intervention_by_swap(
+                f_base,
+                source,
+                "collect",
+                self.interchange_dim,
+                subspaces,
+                subspace_partition=self.subspace_partition,
+                use_fast=self.use_fast,
+            )
+
+        def __str__(self):  # noqa: D401
+            return f"FeatureCollectIntervention(id={featurizer_id})"
+
+    return FeatureCollectIntervention
+
+
+def build_feature_mask_intervention(
+    featurizer: torch.nn.Module,
+    inverse_featurizer: torch.nn.Module,
+    n_features: int,
+    featurizer_id: str,
+):
+    """Return a trainable mask intervention."""
+
+    class FeatureMaskIntervention(pv.TrainableIntervention):
+        """Differential-binary masking in the featurized space."""
+
+        def __init__(self, **kwargs):
+            super().__init__(**kwargs)
+            self._featurizer = featurizer
+            self._inverse = inverse_featurizer
+
+            # Learnable parameters
+            self.mask = torch.nn.Parameter(torch.zeros(n_features), requires_grad=True)
+            self.temperature: Optional[torch.Tensor] = None  # must be set by user
+
+        # -------------------- API helpers -------------------- #
+        def get_temperature(self) -> torch.Tensor:
+            if self.temperature is None:
+                raise ValueError("Temperature has not been set.")
+            return self.temperature
+
+        def set_temperature(self, temp: float | torch.Tensor):
+            self.temperature = (
+                torch.as_tensor(temp, dtype=self.mask.dtype).to(self.mask.device)
+            )
+
+        # ------------------------- forward ------------------- #
+        def forward(self, base, source, subspaces=None):
+            if self.temperature is None:
+                raise ValueError("Cannot run forward without a temperature.")
+
+            f_base, base_err = self._featurizer(base)
+            f_src, _ = self._featurizer(source)
+
+            # Align devices / dtypes
+            mask = self.mask.to(f_base.device)
+            temp = self.temperature.to(f_base.device)
+
+            f_base = f_base.to(mask.dtype)
+            f_src = f_src.to(mask.dtype)
+
+            if self.training:
+                gate = torch.sigmoid(mask / temp)
+            else:
+                gate = (torch.sigmoid(mask) > 0.5).float()
+
+            f_out = (1.0 - gate) * f_base + gate * f_src
+            return self._inverse(f_out.to(base.dtype), base_err).to(base.dtype)
+
+        # ---------------- Sparsity regulariser --------------- #
+        def get_sparsity_loss(self) -> torch.Tensor:
+            if self.temperature is None:
+                raise ValueError("Temperature has not been set.")
+            gate = torch.sigmoid(self.mask / self.temperature)
+            return torch.norm(gate, p=1)
+
+        def __str__(self):  # noqa: D401
+            return f"FeatureMaskIntervention(id={featurizer_id})"
+
+    return FeatureMaskIntervention
+
+
+# --------------------------------------------------------------------------- #
+#  Concrete featurizer implementations                                        #
+# --------------------------------------------------------------------------- #
+class SubspaceFeaturizerModule(torch.nn.Module):
+    """Linear projector onto an orthogonal *rotation* sub-space."""
+
+    def __init__(self, rotate_layer: pv.models.layers.LowRankRotateLayer):
+        super().__init__()
+        self.rotate = rotate_layer
+
+    def forward(self, x: torch.Tensor):
+        r = self.rotate.weight.T  # (out, in)ᵀ
+        f = x.to(r.dtype) @ r.T
+        error = x - (f @ r).to(x.dtype)
+        return f, error
+
+
+class SubspaceInverseFeaturizerModule(torch.nn.Module):
+    """Inverse of :class:`SubspaceFeaturizerModule`."""
+
+    def __init__(self, rotate_layer: pv.models.layers.LowRankRotateLayer):
+        super().__init__()
+        self.rotate = rotate_layer
+
+    def forward(self, f, error):
+        r = self.rotate.weight.T
+        return (f.to(r.dtype) @ r).to(f.dtype) + error.to(f.dtype)
+
+
+class SubspaceFeaturizer(Featurizer):
+    """Orthogonal linear sub-space featurizer."""
+
+    def __init__(
+        self,
+        *,
+        shape: Tuple[int, int] | None = None,
+        rotation_subspace: torch.Tensor | None = None,
+        trainable: bool = True,
+        id: str = "subspace",
+    ):
+        assert (
+            shape is not None or rotation_subspace is not None
+        ), "Provide either `shape` or `rotation_subspace`."
+
+        if shape is not None:
+            rotate = pv.models.layers.LowRankRotateLayer(*shape, init_orth=True)
+        else:
+            shape = rotation_subspace.shape
+            rotate = pv.models.layers.LowRankRotateLayer(*shape, init_orth=False)
+            rotate.weight.data.copy_(rotation_subspace)
+
+        rotate = torch.nn.utils.parametrizations.orthogonal(rotate)
+        rotate.requires_grad_(trainable)
+
+        super().__init__(
+            SubspaceFeaturizerModule(rotate),
+            SubspaceInverseFeaturizerModule(rotate),
+            n_features=rotate.weight.shape[1],
+            id=id,
+        )
+
+
+class SAEFeaturizerModule(torch.nn.Module):
+    """Wrapper around a *Sparse Autoencoder*’s encode() / decode() pair."""
+
+    def __init__(self, sae):
+        super().__init__()
+        self.sae = sae
+
+    def forward(self, x):
+        features = self.sae.encode(x.to(self.sae.dtype))
+        error = x - self.sae.decode(features).to(x.dtype)
+        return features.to(x.dtype), error
+
+
+class SAEInverseFeaturizerModule(torch.nn.Module):
+    """Inverse for :class:`SAEFeaturizerModule`."""
+
+    def __init__(self, sae):
+        super().__init__()
+        self.sae = sae
+
+    def forward(self, features, error):
+        return (
+            self.sae.decode(features.to(self.sae.dtype)).to(features.dtype)
+            + error.to(features.dtype)
+        )
+
+
+class SAEFeaturizer(Featurizer):
+    """Featurizer backed by a pre-trained sparse auto-encoder.
+
+    Notes
+    -----
+    Serialisation is *disabled* for SAE featurizers – saving will raise
+    ``NotImplementedError``.
+    """
+
+    def __init__(self, sae, *, trainable: bool = False):
+        sae.requires_grad_(trainable)
+        super().__init__(
+            SAEFeaturizerModule(sae),
+            SAEInverseFeaturizerModule(sae),
+            n_features=sae.cfg.to_dict()["d_sae"],
+            id="sae",
+        )
+
+
+# --------------------------------------------------------------------------- #
+#  Utility helpers                                                            #
+# --------------------------------------------------------------------------- #
+def _subspace_is_all_none(subspaces) -> bool:
+    """Return ``True`` if *every* element of *subspaces* is ``None``."""
+    return subspaces is None or all(
+        inner is None or all(elem is None for elem in inner) for inner in subspaces
+    )