Max out span extractor

CHANGELOG.md

@@ -11,6 +11,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - Added a way to resize the vocabulary in the T5 module
 - Added an argument `reinit_modules` to `cached_transformers.get()` that allows you to re-initialize the pretrained weights of a transformer model, using layer indices or regex strings.
+- Added a `MaxPoolingSpanExtractor`. This `SpanExtractor` represents each span by a component wise max-pooling-operation.
 
 ### Fixed
 

allennlp/modules/span_extractors/__init__.py

@@ -6,3 +6,4 @@
 from allennlp.modules.span_extractors.bidirectional_endpoint_span_extractor import (
     BidirectionalEndpointSpanExtractor,
 )
+from allennlp.modules.span_extractors.max_pooling_span_extractor import MaxPoolingSpanExtractor

allennlp/modules/span_extractors/max_pooling_span_extractor.py

@@ -0,0 +1,131 @@
+import torch
+
+from allennlp.modules.span_extractors.span_extractor import SpanExtractor
+from allennlp.modules.span_extractors.span_extractor_with_span_width_embedding import (
+    SpanExtractorWithSpanWidthEmbedding,
+)
+from allennlp.nn import util
+from allennlp.nn.util import masked_max
+
+
+@SpanExtractor.register("max_pooling")
+class MaxPoolingSpanExtractor(SpanExtractorWithSpanWidthEmbedding):
+    """
+    Represents spans through the application of a dimension-wise max-pooling operation.
+    Given a span x_i, ..., x_j with i,j as span_start and span_end, each dimension d
+    of the resulting span s is computed via s_d = max(x_id, ..., x_jd).
+
+    Elements masked-out by sequence_mask are ignored when max-pooling is computed.
+    Span representations of masked out span_indices by span_mask are set to '0.'
+
+    Registered as a `SpanExtractor` with name "max_pooling".
+
+    # Parameters
+
+    input_dim : `int`, required.
+        The final dimension of the `sequence_tensor`.
+    num_width_embeddings : `int`, optional (default = `None`).
+        Specifies the number of buckets to use when representing
+        span width features.
+    span_width_embedding_dim : `int`, optional (default = `None`).
+        The embedding size for the span_width features.
+    bucket_widths : `bool`, optional (default = `False`).
+        Whether to bucket the span widths into log-space buckets. If `False`,
+        the raw span widths are used.
+
+    # Returns
+
+    max_pooling_text_embeddings : `torch.FloatTensor`.
+        A tensor of shape (batch_size, num_spans, input_dim), which each span representation
+        is the result of a max-pooling operation.
+
+    """
+
+    def __init__(
+        self,
+        input_dim: int,
+        num_width_embeddings: int = None,
+        span_width_embedding_dim: int = None,
+        bucket_widths: bool = False,
+    ) -> None:
+        super().__init__(
+            input_dim=input_dim,
+            num_width_embeddings=num_width_embeddings,
+            span_width_embedding_dim=span_width_embedding_dim,
+            bucket_widths=bucket_widths,
+        )
+
+    def get_output_dim(self) -> int:
+        if self._span_width_embedding is not None:
+            return self._input_dim + self._span_width_embedding.get_output_dim()
+        return self._input_dim
+
+    def _embed_spans(
+        self,
+        sequence_tensor: torch.FloatTensor,
+        span_indices: torch.LongTensor,
+        sequence_mask: torch.BoolTensor = None,
+        span_indices_mask: torch.BoolTensor = None,
+    ) -> torch.FloatTensor:
+
+        if sequence_tensor.size(-1) != self._input_dim:
+            raise ValueError(
+                f"Dimension mismatch expected ({sequence_tensor.size(-1)}) "
+                f"received ({self._input_dim})."
+            )
+
+        if sequence_tensor.shape[1] <= span_indices.max() or span_indices.min() < 0:
+            raise IndexError(
+                f"Span index out of range, max index ({span_indices.max()}) "
+                f"or min index ({span_indices.min()}) "
+                f"not valid for sequence of length ({sequence_tensor.shape[1]})."
+            )
+
+        if (span_indices[:, :, 0] > span_indices[:, :, 1]).any():
+            raise IndexError(
+                "Span start above span end",
+            )
+
+        # Calculate the maximum sequence length for each element in batch.
+        # If span_end indices are above these length, we adjust the indices in adopted_span_indices
+        if sequence_mask is not None:
+            # shape (batch_size)
+            sequence_lengths = util.get_lengths_from_binary_sequence_mask(sequence_mask)
+        else:
+            # shape (batch_size), filled with the sequence length size of the sequence_tensor.
+            sequence_lengths = torch.ones_like(
+                sequence_tensor[:, 0, 0], dtype=torch.long
+            ) * sequence_tensor.size(1)
+
+        adopted_span_indices = torch.tensor(span_indices, device=span_indices.device)
+
+        for b in range(sequence_lengths.shape[0]):
+            adopted_span_indices[b, :, 1][adopted_span_indices[b, :, 1] >= sequence_lengths[b]] = (
+                sequence_lengths[b] - 1
+            )
+
+        # Raise Error if span indices were completly masked by sequence mask.
+        # We only adjust span_end to the last valid index, so if span_end is below span_start,
+        # both were above the max index:
+
+        if (adopted_span_indices[:, :, 0] > adopted_span_indices[:, :, 1]).any():
+            raise IndexError(
+                "Span indices were masked out entirely by sequence mask",
+            )
+
+        # span_vals <- (batch x num_spans x max_span_length x dim)
+        span_vals, span_mask = util.batched_span_select(sequence_tensor, adopted_span_indices)
+
+        # The application of masked_mask requires a mask of the same shape as span_vals
+        # We repeat the mask along the last dimension (embedding dimension)
+        repeat_dim = len(span_vals.shape) - 1
+        repeat_idx = [1] * (repeat_dim) + [span_vals.shape[-1]]
+
+        # ext_span_mask <- (batch x num_spans x max_span_length x dim)
+        # ext_span_mask True for values in span, False for masked out values
+        ext_span_mask = span_mask.unsqueeze(repeat_dim).repeat(repeat_idx)
+
+        # max_out <- (batch x num_spans x dim)
+        max_out = masked_max(span_vals, ext_span_mask, dim=-2)
+
+        return max_out

tests/modules/span_extractors/max_pooling_span_extractor_test.py

@@ -0,0 +1,184 @@
+import pytest
+import torch
+
+from allennlp.common.params import Params
+from allennlp.modules.span_extractors import SpanExtractor
+from allennlp.modules.span_extractors.max_pooling_span_extractor import MaxPoolingSpanExtractor
+
+
+class TestMaxPoolingSpanExtractor:
+    def test_locally_span_extractor_can_build_from_params(self):
+        params = Params(
+            {
+                "type": "max_pooling",
+                "input_dim": 3,
+                "num_width_embeddings": 5,
+                "span_width_embedding_dim": 3,
+            }
+        )
+        extractor = SpanExtractor.from_params(params)
+        assert isinstance(extractor, MaxPoolingSpanExtractor)
+        assert extractor.get_output_dim() == 6
+
+    def test_max_values_extracted(self):
+        # Test if max_pooling is correctly applied
+        # We use a high dimensional random vector and assume that a randomly correct result is too unlikely
+        sequence_tensor = torch.randn([2, 10, 30])
+        extractor = MaxPoolingSpanExtractor(30)
+
+        indices = torch.LongTensor([[[1, 1], [2, 4], [9, 9]], [[0, 1], [4, 4], [0, 9]]])
+        span_representations = extractor(sequence_tensor, indices)
+
+        assert list(span_representations.size()) == [2, 3, 30]
+        assert extractor.get_output_dim() == 30
+        assert extractor.get_input_dim() == 30
+
+        # We iterate over the tensor to compare the span extractors's results
+        # with the results of python max operation over each dimension for each span and for each batch
+        # For each batch
+        for batch, X in enumerate(indices):
+            # For each defined span index
+            for indices_ind, span_def in enumerate(X):
+
+                # original features of current tested span
+                # span_width x embedding dim (30)
+                span_features_complete = sequence_tensor[batch][span_def[0] : span_def[1] + 1]
+
+                # comparison for each dimension
+                for i in range(extractor.get_output_dim()):
+                    # get the features for dimension i of current span
+                    features_from_span = span_features_complete[:, i]
+                    real_max_value = max(features_from_span)
+
+                    extrected_max_value = span_representations[batch, indices_ind, i]
+
+                    assert real_max_value == extrected_max_value, (
+                        f"Error extracting max value for "
+                        f"batch {batch}, span {indices_ind} on dimension {i}."
+                        f"expected {real_max_value} "
+                        f"but got {extrected_max_value} which is "
+                        f"not the maximum element."
+                    )
+
+    def test_sequence_mask_correct_excluded(self):
+        # Check if span indices masked out by the sequence mask are ignored when computing
+        # the span representations. For this test span_start is valid, but span_end is masked out.
+
+        sequence_tensor = torch.randn([2, 6, 30])
+
+        extractor = MaxPoolingSpanExtractor(30)
+        indices = torch.LongTensor([[[1, 1], [3, 5], [2, 5]], [[0, 0], [0, 3], [4, 5]]])
+        # define sequence mak
+        seq_mask = torch.BoolTensor([[True] * 4 + [False] * 2, [True] * 5 + [False] * 1])
+
+        span_representations = extractor(sequence_tensor, indices, sequence_mask=seq_mask)
+
+        # After we computed the representations we set values to -inf
+        # to compute the "real" max-pooling with python's max function.
+        sequence_tensor[torch.logical_not(seq_mask)] = float("-inf")
+
+        # Comparison is similar to test_max_values_extracted
+        for batch, X in enumerate(indices):
+            for indices_ind, span_def in enumerate(X):
+
+                span_features_complete = sequence_tensor[batch][span_def[0] : span_def[1] + 1]
+
+                for i, _ in enumerate(span_features_complete):
+                    features_from_span = span_features_complete[:, i]
+                    real_max_value = max(features_from_span)
+                    extrected_max_value = span_representations[batch, indices_ind, i]
+
+                    assert real_max_value == extrected_max_value, (
+                        f"Error extracting max value for "
+                        f"batch {batch}, span {indices_ind} on dimension {i}."
+                        f"expected {real_max_value} "
+                        f"but got {extrected_max_value} which is "
+                        f"not the maximum element."
+                    )
+
+    def test_span_mask_correct_excluded(self):
+        # All masked out span indices by span_mask should be '0'
+
+        sequence_tensor = torch.randn([2, 6, 10])
+
+        extractor = MaxPoolingSpanExtractor(10)
+        indices = torch.LongTensor([[[1, 1], [3, 5], [2, 5]], [[0, 0], [0, 3], [4, 5]]])
+
+        span_mask = torch.BoolTensor([[True] * 3, [False] * 3])
+
+        span_representations = extractor(
+            sequence_tensor,
+            indices,
+            span_indices_mask=span_mask,
+        )
+
+        # The span-mask masks out all indices in the last batch
+        # We check whether all span representations for this batch are '0'
+        X = indices[-1]
+        batch = -1
+        for indices_ind, span_def in enumerate(X):
+
+            span_features_complete = sequence_tensor[batch][span_def[0] : span_def[1] + 1]
+
+            for i, _ in enumerate(span_features_complete):
+                real_max_value = torch.FloatTensor([0.0])
+                extrected_max_value = span_representations[batch, indices_ind, i]
+
+                assert real_max_value == extrected_max_value, (
+                    f"Error extracting max value for "
+                    f"batch {batch}, span {indices_ind} on dimension {i}."
+                    f"expected {real_max_value} "
+                    f"but got {extrected_max_value} which is "
+                    f"not the maximum element."
+                )
+
+    def test_inconsistent_extractor_dimension_throws_exception(self):
+
+        sequence_tensor = torch.randn([2, 6, 10])
+        indices = torch.LongTensor([[[1, 1], [2, 4], [9, 9]], [[0, 1], [4, 4], [0, 9]]])
+
+        with pytest.raises(ValueError):
+            extractor = MaxPoolingSpanExtractor(9)
+            extractor(sequence_tensor, indices)
+
+        with pytest.raises(ValueError):
+            extractor = MaxPoolingSpanExtractor(11)
+            extractor(sequence_tensor, indices)
+
+    def test_span_indices_outside_sequence(self):
+
+        sequence_tensor = torch.randn([2, 6, 10])
+        indices = torch.LongTensor([[[6, 6], [2, 4]], [[0, 1], [4, 4]]])
+
+        with pytest.raises(IndexError):
+            extractor = MaxPoolingSpanExtractor(10)
+            extractor(sequence_tensor, indices)
+
+        indices = torch.LongTensor([[[5, 6], [2, 4]], [[0, 1], [4, 4]]])
+
+        with pytest.raises(IndexError):
+            extractor = MaxPoolingSpanExtractor(10)
+            extractor(sequence_tensor, indices)
+
+        indices = torch.LongTensor([[[-1, 0], [2, 4]], [[0, 1], [4, 4]]])
+
+        with pytest.raises(IndexError):
+            extractor = MaxPoolingSpanExtractor(10)
+            extractor(sequence_tensor, indices)
+
+    def test_span_start_below_span_end(self):
+
+        sequence_tensor = torch.randn([2, 6, 10])
+        indices = torch.LongTensor([[[4, 2], [2, 4], [1, 1]], [[0, 1], [4, 4], [1, 1]]])
+        with pytest.raises(IndexError):
+            extractor = MaxPoolingSpanExtractor(10)
+            extractor(sequence_tensor, indices)
+
+    def test_span_sequence_complete_masked(self):
+
+        sequence_tensor = torch.randn([2, 6, 10])
+        seq_mask = torch.BoolTensor([[True] * 2 + [False] * 4, [True] * 3 + [False] * 3])
+        indices = torch.LongTensor([[[5, 5]], [[4, 5]]])
+        with pytest.raises(IndexError):
+            extractor = MaxPoolingSpanExtractor(10)
+            extractor(sequence_tensor, indices, sequence_mask=seq_mask)