Gradient-based LLM attribution in tutorial + LLM attribution type annotations (#1333)

Summary:
Pull Request resolved: #1333

Add gradient-based LLM attribution to the tutorial notebook. Addresses #1237. Additionally, add more type annotations to llm_attr.py.

Reviewed By: vivekmig

Differential Revision: D61461521

fbshipit-source-id: 8ae68773cfd32506e797941698a9a08316676279

Author: craymichael

captum/_utils/typing.py

@@ -2,7 +2,7 @@
 
 # pyre-strict
 
-from typing import List, Tuple, TYPE_CHECKING, TypeVar, Union
+from typing import List, Optional, Protocol, Tuple, TYPE_CHECKING, TypeVar, Union
 
 from torch import Tensor
 from torch.nn import Module
@@ -33,3 +33,14 @@
     TensorLikeList4D,
     TensorLikeList5D,
 ]
+
+
+class TokenizerLike(Protocol):
+    """A protocol for tokenizer-like objects that can be used with Captum
+    LLM attribution methods."""
+
+    def encode(
+        self, text: str, return_tensors: Optional[str] = None
+    ) -> Union[List[int], Tensor]: ...
+    def decode(self, token_ids: Tensor) -> str: ...
+    def convert_ids_to_tokens(self, token_ids: Tensor) -> List[str]: ...

captum/attr/_core/llm_attr.py

@@ -1,18 +1,23 @@
 # pyre-strict
 from copy import copy
 
-from typing import Any, Callable, cast, Dict, List, Optional, Union
+from typing import Any, Callable, cast, Dict, List, Optional, Tuple, Union
 
 import matplotlib.pyplot as plt
 import numpy as np
 
 import torch
+from captum._utils.typing import TokenizerLike
 from captum.attr._core.feature_ablation import FeatureAblation
 from captum.attr._core.kernel_shap import KernelShap
 from captum.attr._core.layer.layer_integrated_gradients import LayerIntegratedGradients
 from captum.attr._core.lime import Lime
 from captum.attr._core.shapley_value import ShapleyValues, ShapleyValueSampling
-from captum.attr._utils.attribution import Attribution
+from captum.attr._utils.attribution import (
+    Attribution,
+    GradientAttribution,
+    PerturbationAttribution,
+)
 from captum.attr._utils.interpretable_input import (
     InterpretableInput,
     TextTemplateInput,
@@ -44,11 +49,12 @@ def __init__(
         self.output_tokens = output_tokens
 
     @property
-    def seq_attr_dict(self) -> Dict[str, Any]:
+    def seq_attr_dict(self) -> Dict[str, float]:
         return {k: v for v, k in zip(self.seq_attr.cpu().tolist(), self.input_tokens)}
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def plot_token_attr(self, show: bool = False):
+    def plot_token_attr(
+        self, show: bool = False
+    ) -> Union[None, Tuple[plt.Figure, plt.Axes]]:
         """
         Generate a matplotlib plot for visualising the attribution
         of the output tokens.
@@ -58,7 +64,11 @@ def plot_token_attr(self, show: bool = False):
                 Default: False
         """
 
-        # pyre-fixme[16]: `Optional` has no attribute `cpu`.
+        if self.token_attr is None:
+            raise ValueError(
+                "token_attr is None (no token-level attribution was performed), please "
+                "use plot_seq_attr instead for the sequence-level attribution plot"
+            )
         token_attr = self.token_attr.cpu()  # type: ignore
 
         # maximum absolute attribution value
@@ -83,7 +93,7 @@ def plot_token_attr(self, show: bool = False):
         )
 
         # Create colorbar
-        cbar = ax.figure.colorbar(im, ax=ax)  # type: ignore
+        cbar = fig.colorbar(im, ax=ax)  # type: ignore
         cbar.ax.set_ylabel("Token Attribuiton", rotation=-90, va="bottom")
 
         # Show all ticks and label them with the respective list entries.
@@ -113,11 +123,13 @@ def plot_token_attr(self, show: bool = False):
 
         if show:
             plt.show()
+            return None  # mypy wants this
         else:
             return fig, ax
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def plot_seq_attr(self, show: bool = False):
+    def plot_seq_attr(
+        self, show: bool = False
+    ) -> Union[None, Tuple[plt.Figure, plt.Axes]]:
         """
         Generate a matplotlib plot for visualising the attribution
         of the output sequence.
@@ -150,6 +162,7 @@ def plot_seq_attr(self, show: bool = False):
 
         if show:
             plt.show()
+            return None  # mypy wants this
         else:
             return fig, ax
 
@@ -181,9 +194,8 @@ class LLMAttribution(Attribution):
 
     def __init__(
         self,
-        attr_method: Attribution,
-        # pyre-fixme[2]: Parameter must be annotated.
-        tokenizer,
+        attr_method: PerturbationAttribution,
+        tokenizer: TokenizerLike,
         attr_target: str = "log_prob",  # TODO: support callable attr_target
     ) -> None:
         """
@@ -208,24 +220,19 @@ class created with the llm model that follows huggingface style
         super().__init__(attr_method.forward_func)
 
         # shallow copy is enough to avoid modifying original instance
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.attr_method = copy(attr_method)
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.include_per_token_attr = isinstance(
+        self.attr_method: PerturbationAttribution = copy(attr_method)
+        self.include_per_token_attr: bool = isinstance(
             attr_method, self.SUPPORTED_PER_TOKEN_ATTR_METHODS
         )
 
         self.attr_method.forward_func = self._forward_func
 
         # alias, we really need a model and don't support wrapper functions
         # coz we need call model.forward, model.generate, etc.
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.model = cast(nn.Module, self.forward_func)
+        self.model: nn.Module = cast(nn.Module, self.forward_func)
 
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.tokenizer = tokenizer
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.device = (
+        self.tokenizer: TokenizerLike = tokenizer
+        self.device: torch.device = (
             cast(torch.device, self.model.device)
             if hasattr(self.model, "device")
             else next(self.model.parameters()).device
@@ -239,15 +246,12 @@ class created with the llm model that follows huggingface style
 
     def _forward_func(
         self,
-        # pyre-fixme[2]: Parameter must be annotated.
-        perturbed_tensor,
-        # pyre-fixme[2]: Parameter must be annotated.
-        inp,
-        # pyre-fixme[2]: Parameter must be annotated.
-        target_tokens,
+        perturbed_tensor: Union[None, Tensor],
+        inp: InterpretableInput,
+        target_tokens: Tensor,
         use_cached_outputs: bool = False,
-        _inspect_forward=None,
-    ) -> Union[int, Tensor]:
+        _inspect_forward: Optional[Callable[[str, str, List[float]], None]] = None,
+    ) -> Tensor:
         perturbed_input = self._format_model_input(inp.to_model_input(perturbed_tensor))
         init_model_inp = perturbed_input
 
@@ -279,7 +283,9 @@ def _forward_func(
                 (model_inp, torch.tensor([[target_token]]).to(self.device)), dim=1
             )
 
-        total_log_prob = sum(log_prob_list)
+        # pyre-ignore[9] pyre/mypy thinks sum returns int here, but it will return
+        # Tensor
+        total_log_prob: Tensor = sum(log_prob_list)  # type: ignore
         # 1st element is the total prob, rest are the target tokens
         # add a leading dim for batch even we only support single instance for now
         if self.include_per_token_attr:
@@ -288,8 +294,6 @@ def _forward_func(
             ).unsqueeze(0)
         else:
             target_log_probs = total_log_prob  # type: ignore
-        # pyre-fixme[6]: For 1st argument expected `Tensor` but got `Union[int,
-        #  Tensor]`.
         target_probs = torch.exp(target_log_probs)
 
         if _inspect_forward:
@@ -301,35 +305,31 @@ def _forward_func(
 
         return target_probs if self.attr_target != "log_prob" else target_log_probs
 
-    # pyre-fixme[3]: Return type must be annotated.
-    def _format_model_input(self, model_input: Union[str, Tensor]):
+    def _format_model_input(self, model_input: Union[str, Tensor]) -> Tensor:
         """
         Convert str to tokenized tensor
         to make LLMAttribution work with model inputs of both
         raw text and text token tensors
         """
         # return tensor(1, n_tokens)
         if isinstance(model_input, str):
-            return self.tokenizer.encode(model_input, return_tensors="pt").to(
-                self.device
-            )
+            # pyre-ignore[9] pyre/mypy thinks return type may be List, but it will be
+            # Tensor
+            return self.tokenizer.encode(  # type: ignore
+                model_input, return_tensors="pt"
+            ).to(self.device)
         return model_input.to(self.device)
 
     def attribute(
         self,
         inp: InterpretableInput,
         target: Union[str, torch.Tensor, None] = None,
         num_trials: int = 1,
-        # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use
-        #  `typing.Dict[<key type>, <value type>]` to avoid runtime subscripting
-        #  errors.
-        gen_args: Optional[Dict] = None,
+        gen_args: Optional[Dict[str, Any]] = None,
         use_cached_outputs: bool = True,
         # internal callback hook can be used for logging
-        # pyre-fixme[24]: Generic type `Callable` expects 2 type parameters.
-        _inspect_forward: Optional[Callable] = None,
-        # pyre-fixme[2]: Parameter must be annotated.
-        **kwargs,
+        _inspect_forward: Optional[Callable[[str, str, List[float]], None]] = None,
+        **kwargs: Any,
     ) -> LLMAttributionResult:
         """
         Args:
@@ -380,10 +380,14 @@ def attribute(
                 target_tokens = torch.tensor(target_tokens)
             elif type(target) is torch.Tensor:
                 target_tokens = target
+            else:
+                raise TypeError(
+                    "target must either be str or Tensor, but the type of target is "
+                    "{}".format(type(target))
+                )
 
         attr = torch.zeros(
             [
-                # pyre-fixme[61]: `target_tokens` is undefined, or not always defined.
                 1 + len(target_tokens) if self.include_per_token_attr else 1,
                 inp.n_itp_features,
             ],
@@ -398,8 +402,6 @@ def attribute(
                 attr_input,
                 additional_forward_args=(
                     inp,
-                    # pyre-fixme[61]: `target_tokens` is undefined, or not always
-                    #  defined.
                     target_tokens,
                     use_cached_outputs,
                     _inspect_forward,
@@ -424,7 +426,6 @@ def attribute(
                 attr[1:] if self.include_per_token_attr else None
             ),  # shape(n_output_token, n_input_features)
             inp.values,
-            # pyre-fixme[61]: `target_tokens` is undefined, or not always defined.
             self.tokenizer.convert_ids_to_tokens(target_tokens),
         )
 
@@ -454,14 +455,11 @@ class LLMGradientAttribution(Attribution):
     SUPPORTED_METHODS = (LayerIntegratedGradients,)
     SUPPORTED_INPUTS = (TextTokenInput,)
 
-    # pyre-fixme[3]: Return type must be annotated.
     def __init__(
         self,
-        # pyre-fixme[2]: Parameter must be annotated.
-        attr_method,
-        # pyre-fixme[2]: Parameter must be annotated.
-        tokenizer,
-    ):
+        attr_method: GradientAttribution,
+        tokenizer: TokenizerLike,
+    ) -> None:
         """
         Args:
             attr_method (Attribution): instance of a supported perturbation attribution
@@ -476,19 +474,15 @@ class created with the llm model that follows huggingface style
         super().__init__(attr_method.forward_func)
 
         # shallow copy is enough to avoid modifying original instance
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.attr_method = copy(attr_method)
+        self.attr_method: GradientAttribution = copy(attr_method)
         self.attr_method.forward_func = self._forward_func
 
         # alias, we really need a model and don't support wrapper functions
         # coz we need call model.forward, model.generate, etc.
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.model = cast(nn.Module, self.forward_func)
+        self.model: nn.Module = cast(nn.Module, self.forward_func)
 
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.tokenizer = tokenizer
-        # pyre-fixme[4]: Attribute must be annotated.
-        self.device = (
+        self.tokenizer: TokenizerLike = tokenizer
+        self.device: torch.device = (
             cast(torch.device, self.model.device)
             if hasattr(self.model, "device")
             else next(self.model.parameters()).device
@@ -526,9 +520,7 @@ def _forward_func(
         # the attribution target is limited to the log probability
         return token_log_probs
 
-    # pyre-fixme[3]: Return type must be annotated.
-    # pyre-fixme[2]: Parameter must be annotated.
-    def _format_model_input(self, model_input):
+    def _format_model_input(self, model_input: Tensor) -> Tensor:
         """
         Convert str to tokenized tensor
         """
@@ -538,12 +530,8 @@ def attribute(
         self,
         inp: InterpretableInput,
         target: Union[str, torch.Tensor, None] = None,
-        # pyre-fixme[24]: Generic type `dict` expects 2 type parameters, use
-        #  `typing.Dict[<key type>, <value type>]` to avoid runtime subscripting
-        #  errors.
-        gen_args: Optional[Dict] = None,
-        # pyre-fixme[2]: Parameter must be annotated.
-        **kwargs,
+        gen_args: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
     ) -> LLMAttributionResult:
         """
         Args:
@@ -590,19 +578,21 @@ def attribute(
                 target_tokens = torch.tensor(target_tokens)
             elif type(target) is torch.Tensor:
                 target_tokens = target
+            else:
+                raise TypeError(
+                    "target must either be str or Tensor, but the type of target is "
+                    "{}".format(type(target))
+                )
 
         attr_inp = inp.to_tensor().to(self.device)
 
         attr_list = []
-        # pyre-fixme[61]: `target_tokens` is undefined, or not always defined.
         for cur_target_idx, _ in enumerate(target_tokens):
             # attr in shape(batch_size, input+output_len, emb_dim)
             attr = self.attr_method.attribute(
                 attr_inp,
                 additional_forward_args=(
                     inp,
-                    # pyre-fixme[61]: `target_tokens` is undefined, or not always
-                    #  defined.
                     target_tokens,
                     cur_target_idx,
                 ),
@@ -629,7 +619,7 @@ def attribute(
         # it attributes to all the elements of the output of the specified layer
         # so we need special handling for the inp type which don't care all the elements
         if isinstance(inp, TextTokenInput) and inp.itp_mask is not None:
-            itp_mask = inp.itp_mask.to(self.device)
+            itp_mask = inp.itp_mask.to(attr.device)
             itp_mask = itp_mask.expand_as(attr)
             attr = attr[itp_mask].view(attr.size(0), -1)
 
@@ -642,7 +632,6 @@ def attribute(
             seq_attr,
             attr,  # shape(n_output_token, n_input_features)
             inp.values,
-            # pyre-fixme[61]: `target_tokens` is undefined, or not always defined.
             self.tokenizer.convert_ids_to_tokens(target_tokens),
         )