Add Hopfield and Attention memory models (#134)

* Add hopfield networks

* Add docs

* Torch 2.0.0 now implements prod for bfloat16 on CPU

* Fix documentation references

* Add Hopfield memory class

* Fix documentation links

* Add basic memory tests

Author: mikeheddes

docs/conf.py

@@ -45,6 +45,13 @@
 # This pattern also affects html_static_path and html_extra_path.
 exclude_patterns = []
 
+# maps functions with a class name that is indistinguishable when case is
+# ignore to another filename
+autosummary_filename_map = {
+    "torchhd.memory.Hopfield": "hopfield-class",
+    "torchhd.memory.hopfield": "hopfield-function",
+}
+
 
 # -- Options for HTML output -------------------------------------------------
 

docs/getting_started.rst

@@ -55,7 +55,7 @@ Similar behavior can be achieved using the classes in the :ref:`embeddings` modu
 
 	weight = torch.tensor([149.0])
 	# explicit mapping of the fruit weight to an index
-	w_i = torchhd.functional.value_to_index(weight, 0, 200, 10)
+	w_i = torchhd.value_to_index(weight, 0, 200, 10)
 	weights[w_i]  # select representation of 149
 
 whereas the :ref:`embeddings<embeddings>` have this common behavior built-in:

docs/memory.rst

@@ -10,3 +10,12 @@ torchhd.memory
     :template: class.rst
 
     SparseDistributed
+    Hopfield
+
+.. autosummary:: 
+    :toctree: generated/
+    :template: function.rst
+
+    hopfield
+    modern_hopfield
+    attention

torchhd/functional.py

@@ -771,7 +771,7 @@ def __call__(self, input: VSATensor, shifts: int = 1) -> VSATensor:
 def inverse(input: VSATensor) -> VSATensor:
     r"""Inverse for the binding operation.
 
-    See :func:`~torchhd.functional.bind`.
+    See :func:`~torchhd.bind`.
 
     Args:
         input (VSATensor): input hypervector
@@ -796,7 +796,7 @@ def inverse(input: VSATensor) -> VSATensor:
 def negative(input: VSATensor) -> VSATensor:
     r"""Inverse for the bundling operation.
 
-    See :func:`~torchhd.functional.bundle`.
+    See :func:`~torchhd.bundle`.
 
     Args:
         input (VSATensor): input hypervector