Add complex hypervector types (#81)

* WIP implement complex hypervectors

* Implement complex level and circular hypervector sets

* Add complex tests

* Simplify plotting utility

* Use unbind function

* Add unbind to docs

docs/functional.rst

@@ -28,6 +28,7 @@ Operations
     :template: function.rst
 
     bind
+    unbind
     bundle
     permute
     cleanup

torchhd/__init__.py

@@ -10,6 +10,7 @@
     level_hv,
     circular_hv,
     bind,
+    unbind,
     bundle,
     permute,
 )
@@ -27,6 +28,7 @@
     "level_hv",
     "circular_hv",
     "bind",
+    "unbind",
     "bundle",
     "permute",
 ]

torchhd/structures.py

@@ -212,10 +212,10 @@ def contains(self, input: Tensor) -> Tensor:
         Examples::
 
             >>> M.contains(letters_hv[0])
-            tensor([0.4575])
+            tensor(0.4575)
 
         """
-        return functional.cosine_similarity(input, self.value.unsqueeze(0))
+        return functional.cosine_similarity(input, self.value)
 
     def __len__(self) -> int:
         """Returns the size of the multiset.
@@ -363,7 +363,7 @@ def get(self, key: Tensor) -> Tensor:
             tensor([ 1., -1.,  1.,  ..., -1.,  1., -1.])
 
         """
-        return functional.bind(self.value, key)
+        return functional.unbind(self.value, key)
 
     def replace(self, key: Tensor, old: Tensor, new: Tensor) -> None:
         """Replace the value from key-value pair in the hash table.
@@ -711,7 +711,7 @@ def pop(self, input: Tensor) -> None:
 
         """
         self.size -= 1
-        self.value = functional.bind(self.value, input)
+        self.value = functional.unbind(self.value, input)
         self.value = functional.permute(self.value, shifts=-1)
 
     def popleft(self, input: Tensor) -> None:
@@ -727,7 +727,7 @@ def popleft(self, input: Tensor) -> None:
         """
         self.size -= 1
         rotated_input = functional.permute(input, shifts=len(self))
-        self.value = functional.bind(self.value, rotated_input)
+        self.value = functional.unbind(self.value, rotated_input)
 
     def replace(self, index: int, old: Tensor, new: Tensor) -> None:
         """Replace the old hypervector value from the given index, for the new hypervector value.
@@ -744,7 +744,7 @@ def replace(self, index: int, old: Tensor, new: Tensor) -> None:
 
         """
         rotated_old = functional.permute(old, shifts=self.size - index - 1)
-        self.value = functional.bind(self.value, rotated_old)
+        self.value = functional.unbind(self.value, rotated_old)
 
         rotated_new = functional.permute(new, shifts=self.size - index - 1)
         self.value = functional.bind(self.value, rotated_new)
@@ -880,13 +880,13 @@ def node_neighbors(self, input: Tensor, outgoing=True) -> Tensor:
         """
         if self.is_directed:
             if outgoing:
-                permuted_neighbors = functional.bind(self.value, input)
+                permuted_neighbors = functional.unbind(self.value, input)
                 return functional.permute(permuted_neighbors, shifts=-1)
             else:
                 permuted_node = functional.permute(input, shifts=1)
-                return functional.bind(self.value, permuted_node)
+                return functional.unbind(self.value, permuted_node)
         else:
-            return functional.bind(self.value, input)
+            return functional.unbind(self.value, input)
 
     def contains(self, input: Tensor) -> Tensor:
         """Returns the cosine similarity of the input vector against the graph.
@@ -898,9 +898,9 @@ def contains(self, input: Tensor) -> Tensor:
 
             >>> e = G.encode_edge(letters_hv[0], letters_hv[1])
             >>> G.contains(e)
-            tensor([1.])
+            tensor(1.)
         """
-        return functional.cosine_similarity(input, self.value.unsqueeze(0))
+        return functional.cosine_similarity(input, self.value)
 
     def clear(self) -> None:
         """Empties the graph.
@@ -1012,7 +1012,7 @@ def get_leaf(self, path: List[str]) -> Tensor:
                         hv_path, functional.permute(self.right, shifts=idx)
                     )
 
-        return functional.bind(hv_path, self.value)
+        return functional.unbind(self.value, hv_path)
 
     def clear(self) -> None:
         """Empties the tree.
@@ -1084,8 +1084,8 @@ def transition(self, state: Tensor, action: Tensor) -> Tensor:
             tensor([ 1.,  1., -1.,  ..., -1., -1.,  1.])
 
         """
-        next_state = functional.bind(self.value, state)
-        next_state = functional.bind(next_state, action)
+        next_state = functional.unbind(self.value, state)
+        next_state = functional.unbind(next_state, action)
         return functional.permute(next_state, shifts=-1)
 
     def clear(self) -> None:

torchhd/tests/basis_hv/test_circular_hv.py

@@ -47,30 +47,51 @@ def test_value(self, dtype):
             assert torch.all(
                 (hv == True) | (hv == False)
             ).item(), "values are either 1 or 0"
+        elif dtype in torch_complex_dtypes:
+            magnitudes= hv.abs()
+            assert torch.allclose(magnitudes, torch.tensor(1.0, dtype=magnitudes.dtype)), "magnitude must be 1"
         else:
             assert torch.all(
                 (hv == -1) | (hv == 1)
             ).item(), "values are either -1 or +1"
 
+
         hv = functional.circular_hv(8, 1000000, generator=generator, dtype=dtype)
-        sims = functional.hamming_similarity(hv[0], hv).float() / 1000000
-        sims_diff = sims[:-1] - sims[1:]
+        if dtype in torch_complex_dtypes:
+            sims = functional.cosine_similarity(hv[0], hv)
+            sims_diff = sims[:-1] - sims[1:]
 
-        assert torch.all(
-            sims_diff.sign() == torch.tensor([1, 1, 1, 1, -1, -1, -1])
-        ), "second half must get more similar"
+            assert torch.all(
+                sims_diff.sign() == torch.tensor([1, 1, 1, 1, -1, -1, -1])
+            ), "second half must get more similar"
 
-        abs_sims_diff = sims_diff.abs()
-        assert torch.all(
-            (0.124 < abs_sims_diff) & (abs_sims_diff < 0.126)
-        ).item(), "similarity changes linearly"
+            abs_sims_diff = sims_diff.abs()
+            assert torch.all(
+                (0.248 < abs_sims_diff) & (abs_sims_diff < 0.252)
+            ).item(), "similarity changes linearly"
+        else:
+            sims = functional.hamming_similarity(hv[0], hv).float() / 1000000
+            sims_diff = sims[:-1] - sims[1:]
+
+            assert torch.all(
+                sims_diff.sign() == torch.tensor([1, 1, 1, 1, -1, -1, -1])
+            ), "second half must get more similar"
+
+            abs_sims_diff = sims_diff.abs()
+            assert torch.all(
+                (0.124 < abs_sims_diff) & (abs_sims_diff < 0.126)
+            ).item(), "similarity changes linearly"
 
     @pytest.mark.parametrize("sparsity", [0.0, 0.1, 0.756, 1.0])
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_sparsity(self, sparsity, dtype):
         if not supported_dtype(dtype):
             return
 
+        if dtype in torch_complex_dtypes:
+            # Complex hypervectors don't support sparsity.
+            return
+
         generator = torch.Generator()
         generator.manual_seed(seed)
 
@@ -96,12 +117,6 @@ def test_device(self, dtype):
 
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_dtype(self, dtype):
-        if dtype in torch_complex_dtypes:
-            with pytest.raises(NotImplementedError):
-                functional.circular_hv(3, 26, dtype=dtype)
-
-            return
-
         if dtype == torch.uint8:
             with pytest.raises(ValueError):
                 functional.circular_hv(3, 26, dtype=dtype)

torchhd/tests/basis_hv/test_identity_hv.py

@@ -47,12 +47,6 @@ def test_device(self, dtype):
 
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_dtype(self, dtype):
-        if dtype in torch_complex_dtypes:
-            with pytest.raises(NotImplementedError):
-                functional.identity_hv(3, 26, dtype=dtype)
-
-            return
-
         if dtype == torch.uint8:
             with pytest.raises(ValueError):
                 functional.identity_hv(3, 26, dtype=dtype)

torchhd/tests/basis_hv/test_level_hv.py

@@ -47,29 +47,48 @@ def test_value(self, dtype):
             assert torch.all(
                 (hv == True) | (hv == False)
             ).item(), "values are either 1 or 0"
+        elif dtype in torch_complex_dtypes:
+            magnitudes= hv.abs()
+            assert torch.allclose(magnitudes, torch.tensor(1.0, dtype=magnitudes.dtype)), "magnitude must be 1"
         else:
             assert torch.all(
                 (hv == -1) | (hv == 1)
             ).item(), "values are either -1 or +1"
 
         # look at the similarity profile w.r.t. the first hypervector
-        sims = functional.hamming_similarity(hv[0], hv).float() / 10000
-        sims_diff = sims[:-1] - sims[1:]
-        assert torch.all(sims_diff > 0).item(), "similarity must be decreasing"
+        if dtype in torch_complex_dtypes:
+            sims = functional.cosine_similarity(hv[0], hv)
+            sims_diff = sims[:-1] - sims[1:]
+            assert torch.all(sims_diff > 0).item(), "similarity must be decreasing"
 
-        hv = functional.level_hv(5, 1000000, generator=generator, dtype=dtype)
-        sims = functional.hamming_similarity(hv[0], hv).float() / 1000000
-        sims_diff = sims[:-1] - sims[1:]
-        assert torch.all(
-            (0.124 < sims_diff) & (sims_diff < 0.126)
-        ).item(), "similarity decreases linearly"
+            hv = functional.level_hv(5, 1000000, generator=generator, dtype=dtype)
+            sims = functional.cosine_similarity(hv[0], hv)
+            sims_diff = sims[:-1] - sims[1:]
+            assert torch.all(
+                (0.248 < sims_diff) & (sims_diff < 0.252)
+            ).item(), "similarity decreases linearly"
+        else:
+            sims = functional.hamming_similarity(hv[0], hv).float() / 10000
+            sims_diff = sims[:-1] - sims[1:]
+            assert torch.all(sims_diff > 0).item(), "similarity must be decreasing"
+
+            hv = functional.level_hv(5, 1000000, generator=generator, dtype=dtype)
+            sims = functional.hamming_similarity(hv[0], hv).float() / 1000000
+            sims_diff = sims[:-1] - sims[1:]
+            assert torch.all(
+                (0.124 < sims_diff) & (sims_diff < 0.126)
+            ).item(), "similarity decreases linearly"
 
     @pytest.mark.parametrize("sparsity", [0.0, 0.1, 0.756, 1.0])
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_sparsity(self, sparsity, dtype):
         if not supported_dtype(dtype):
             return
 
+        if dtype in torch_complex_dtypes:
+            # Complex hypervectors don't support sparsity.
+            return
+
         generator = torch.Generator()
         generator.manual_seed(seed)
 
@@ -95,12 +114,6 @@ def test_device(self, dtype):
 
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_dtype(self, dtype):
-        if dtype in torch_complex_dtypes:
-            with pytest.raises(NotImplementedError):
-                functional.level_hv(3, 26, dtype=dtype)
-
-            return
-
         if dtype == torch.uint8:
             with pytest.raises(ValueError):
                 functional.level_hv(3, 26, dtype=dtype)

torchhd/tests/basis_hv/test_random_hv.py

@@ -42,21 +42,26 @@ def test_value(self, dtype):
         generator = torch.Generator()
         generator.manual_seed(seed)
 
+        hv = functional.random_hv(100, 10000, dtype=dtype, generator=generator)
+
         if dtype == torch.bool:
-            hv = functional.random_hv(100, 10000, dtype=dtype, generator=generator)
             assert torch.all((hv == False) | (hv == True)).item()
-
-            return
-
-        hv = functional.random_hv(100, 10000, dtype=dtype, generator=generator)
-        assert torch.all((hv == -1) | (hv == 1)).item()
+        elif dtype in torch_complex_dtypes:
+            magnitudes= hv.abs()
+            assert torch.allclose(magnitudes, torch.tensor(1.0, dtype=magnitudes.dtype)), "magnitude must be 1"
+        else:
+            assert torch.all((hv == -1) | (hv == 1)).item()
 
     @pytest.mark.parametrize("sparsity", [0.0, 0.1, 0.756, 1.0])
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_sparsity(self, sparsity, dtype):
         if not supported_dtype(dtype):
             return
 
+        if dtype in torch_complex_dtypes:
+            # Complex hypervectors don't support sparsity.
+            return
+
         generator = torch.Generator()
         generator.manual_seed(seed)
 
@@ -83,14 +88,24 @@ def test_orthogonality(self, dtype):
         generator = torch.Generator()
         generator.manual_seed(seed)
 
-        sims = [None] * 100
-        for i in range(100):
-            hv = functional.random_hv(2, 10000, dtype=dtype, generator=generator)
-            sims[i] = functional.hamming_similarity(hv[0], hv[1].unsqueeze(0))
-
-        sims = torch.cat(sims).float() / 10000
-        assert within(sims.mean().item(), 0.5, 0.001)
-        assert sims.std().item() < 0.01
+        if dtype in torch_complex_dtypes:
+            sims = [None] * 100
+            for i in range(100):
+                hv = functional.random_hv(2, 10000, dtype=dtype, generator=generator)
+                sims[i] = functional.cosine_similarity(hv[0], hv[1])
+
+            sims = torch.stack(sims).float() / 10000
+            assert within(sims.mean().item(), 0.0, 0.001)
+            assert sims.std().item() < 0.01
+        else:
+            sims = [None] * 100
+            for i in range(100):
+                hv = functional.random_hv(2, 10000, dtype=dtype, generator=generator)
+                sims[i] = functional.hamming_similarity(hv[0], hv[1].unsqueeze(0))
+
+            sims = torch.stack(sims).float() / 10000
+            assert within(sims.mean().item(), 0.5, 0.001)
+            assert sims.std().item() < 0.01
 
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_device(self, dtype):
@@ -103,12 +118,6 @@ def test_device(self, dtype):
 
     @pytest.mark.parametrize("dtype", torch_dtypes)
     def test_dtype(self, dtype):
-        if dtype in torch_complex_dtypes:
-            with pytest.raises(NotImplementedError):
-                functional.random_hv(3, 26, dtype=dtype)
-
-            return
-
         if dtype == torch.uint8:
             with pytest.raises(ValueError):
                 functional.random_hv(3, 26, dtype=dtype)