Add Sequential IG method (#222)

README.md

@@ -130,27 +130,29 @@ Use the `inseq.list_feature_attribution_methods` function to list all available
 
 #### Gradient-based attribution
 
-- `saliency`: [Saliency](https://arxiv.org/abs/1312.6034) (Simonyan et al., 2013)
+- `saliency`: [Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps](https://arxiv.org/abs/1312.6034) (Simonyan et al., 2013)
 
-- `input_x_gradient`: [Input x Gradient](https://arxiv.org/abs/1312.6034) (Simonyan et al., 2013)
+- `input_x_gradient`: [Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps](https://arxiv.org/abs/1312.6034) (Simonyan et al., 2013)
 
-- `integrated_gradients`: [Integrated Gradients](https://arxiv.org/abs/1703.01365) (Sundararajan et al., 2017)
+- `integrated_gradients`: [Axiomatic Attribution for Deep Networks](https://arxiv.org/abs/1703.01365) (Sundararajan et al., 2017)
 
-- `deeplift`: [DeepLIFT](https://arxiv.org/abs/1704.02685) (Shrikumar et al., 2017)
+- `deeplift`: [Learning Important Features Through Propagating Activation Differences](https://arxiv.org/abs/1704.02685) (Shrikumar et al., 2017)
 
-- `gradient_shap`: [Gradient SHAP](https://dl.acm.org/doi/10.5555/3295222.3295230) (Lundberg and Lee, 2017)
+- `gradient_shap`: [A unified approach to interpreting model predictions](https://dl.acm.org/doi/10.5555/3295222.3295230) (Lundberg and Lee, 2017)
 
-- `discretized_integrated_gradients`: [Discretized Integrated Gradients](https://aclanthology.org/2021.emnlp-main.805/) (Sanyal and Ren, 2021)
+- `discretized_integrated_gradients`: [Discretized Integrated Gradients for Explaining Language Models](https://aclanthology.org/2021.emnlp-main.805/) (Sanyal and Ren, 2021)
+
+- `sequential_integrated_gradients`: [Sequential Integrated Gradients: a simple but effective method for explaining language models](https://aclanthology.org/2023.findings-acl.477/) (Enguehard, 2023)
 
 #### Internals-based attribution
 
-- `attention`: [Attention Weight Attribution](https://arxiv.org/abs/1409.0473) (Bahdanau et al., 2014)
+- `attention`: Attention Weight Attribution, from [Neural Machine Translation by Jointly Learning to Align and Translate](https://arxiv.org/abs/1409.0473) (Bahdanau et al., 2014)
 
 #### Perturbation-based attribution
 
-- `occlusion`: [Occlusion](https://link.springer.com/chapter/10.1007/978-3-319-10590-1_53) (Zeiler and Fergus, 2014)
+- `occlusion`: [Visualizing and Understanding Convolutional Networks](https://link.springer.com/chapter/10.1007/978-3-319-10590-1_53) (Zeiler and Fergus, 2014)
 
-- `lime`: [LIME](https://arxiv.org/abs/1602.04938) (Ribeiro et al., 2016)
+- `lime`: ["Why Should I Trust You?": Explaining the Predictions of Any Classifier](https://arxiv.org/abs/1602.04938) (Ribeiro et al., 2016)
 
 #### Step functions
 
@@ -262,9 +264,10 @@ Inseq has been used in various research projects. A list of known publications t
 <details>
   <summary><b>2023</b></summary>
   <ol>
-    <li> <a href="https://arxiv.org/abs/2302.13942">Inseq: An Interpretability Toolkit for Sequence Generation Models</a> (Sarti et al., 2023) </li>
+    <li> <a href="https://aclanthology.org/2023.acl-demo.40/">Inseq: An Interpretability Toolkit for Sequence Generation Models</a> (Sarti et al., 2023) </li>
     <li> <a href="https://arxiv.org/abs/2302.14220">Are Character-level Translations Worth the Wait? Comparing Character- and Subword-level Models for Machine Translation</a> (Edman et al., 2023) </li>
-    <li> <a href="https://arxiv.org/abs/2305.15908">Response Generation in Longitudinal Dialogues: Which Knowledge Representation Helps?</a> (Mousavi et al., 2023)  </li>
+    <li> <a href="https://aclanthology.org/2023.nlp4convai-1.1/">Response Generation in Longitudinal Dialogues: Which Knowledge Representation Helps?</a> (Mousavi et al., 2023)  </li>
+    <li> <a href="https://arxiv.org/abs/2310.01188">Quantifying the Plausibility of Context Reliance in Neural Machine Translation</a> (Sarti et al., 2023)</li>
   </ol>
 
 </details>

docs/source/main_classes/feature_attribution.rst

@@ -28,9 +28,6 @@ Gradient Attribution Methods
     :members:
 
 
-.. warning::
-    The DiscretizedIntegratedGradientsAttribution class is currently exhibiting inconsistent behavior, so usage should be limited until further notice. See PR `# 114 <https://github.com/inseq-team/inseq/pull/114>`__ for additional info.
-
 .. autoclass:: inseq.attr.feat.DiscretizedIntegratedGradientsAttribution
     :members:
 
@@ -50,6 +47,10 @@ Gradient Attribution Methods
     :members:
 
 
+.. autoclass:: inseq.attr.feat.SequentialIntegratedGradientsAttribution
+    :members:
+
+
 Layer Attribution Methods
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

inseq/attr/feat/__init__.py

@@ -11,6 +11,7 @@
     LayerGradientXActivationAttribution,
     LayerIntegratedGradientsAttribution,
     SaliencyAttribution,
+    SequentialIntegratedGradientsAttribution,
 )
 from .internals_attribution import AttentionWeightsAttribution, InternalsAttributionRegistry
 from .perturbation_attribution import (
@@ -37,4 +38,5 @@
     "AttentionWeightsAttribution",
     "OcclusionAttribution",
     "LimeAttribution",
+    "SequentialIntegratedGradientsAttribution",
 ]

inseq/attr/feat/gradient_attribution.py

@@ -32,7 +32,7 @@
 from ..attribution_decorators import set_hook, unset_hook
 from .attribution_utils import get_source_target_attributions
 from .feature_attribution import FeatureAttribution
-from .ops import DiscretetizedIntegratedGradients
+from .ops import DiscretetizedIntegratedGradients, SequentialIntegratedGradients
 
 logger = logging.getLogger(__name__)
 
@@ -212,6 +212,22 @@ def __init__(self, attribution_model):
         self.method = Saliency(self.attribution_model)
 
 
+class SequentialIntegratedGradientsAttribution(GradientAttributionRegistry):
+    """Sequential Integrated Gradients attribution method.
+
+    Reference: https://aclanthology.org/2023.findings-acl.477/
+
+    Original implementation: https://github.com/josephenguehard/time_interpret/blob/main/tint/attr/seq_ig.py
+    """
+
+    method_name = "sequential_integrated_gradients"
+
+    def __init__(self, attribution_model, multiply_by_inputs: bool = True, **kwargs):
+        super().__init__(attribution_model)
+        self.method = SequentialIntegratedGradients(self.attribution_model, multiply_by_inputs)
+        self.use_baselines = True
+
+
 # Layer methods
 
 

inseq/attr/feat/internals_attribution.py

@@ -18,7 +18,6 @@
 
 from captum._utils.typing import TensorOrTupleOfTensorsGeneric
 from captum.attr._utils.attribution import Attribution
-from captum.log import log_usage
 
 from ...data import MultiDimensionalFeatureAttributionStepOutput
 from ...utils import Registry
@@ -43,7 +42,6 @@ class AttentionWeights(Attribution):
         def has_convergence_delta() -> bool:
             return False
 
-        @log_usage()
         def attribute(
             self,
             inputs: TensorOrTupleOfTensorsGeneric,

inseq/attr/feat/ops/__init__.py

@@ -1,9 +1,11 @@
 from .discretized_integrated_gradients import DiscretetizedIntegratedGradients
 from .lime import Lime
 from .monotonic_path_builder import MonotonicPathBuilder
+from .sequential_integrated_gradients import SequentialIntegratedGradients
 
 __all__ = [
     "DiscretetizedIntegratedGradients",
     "MonotonicPathBuilder",
     "Lime",
+    "SequentialIntegratedGradients",
 ]

inseq/attr/feat/ops/discretized_integrated_gradients.py

@@ -31,7 +31,6 @@
 from captum.attr._core.integrated_gradients import IntegratedGradients
 from captum.attr._utils.batching import _batch_attribution
 from captum.attr._utils.common import _format_input_baseline, _reshape_and_sum, _validate_input
-from captum.log import log_usage
 from torch import Tensor
 
 from ....utils import INSEQ_ARTIFACTS_CACHE
@@ -87,7 +86,6 @@ def get_inputs_baselines(scaled_features_tpl: Tuple[Tensor, ...], n_steps: int)
         )
         return inputs, baselines
 
-    @log_usage()
     def attribute(  # type: ignore
         self,
         inputs: MultiStepEmbeddingsTensor,

inseq/attr/feat/ops/lime.py

@@ -62,7 +62,6 @@ def __init__(
         )
         self.attribution_model = attribution_model
 
-    # @log_usage
     def attribute(
         self,
         inputs: TensorOrTupleOfTensorsGeneric,