HRR documentation update

Author: mikeheddes

torchhd/fhrr.py

@@ -175,9 +175,8 @@ def random_hv(
         size = (num_vectors, dimensions)
         angle = torch.empty(size, dtype=dtype, device=device)
         angle.uniform_(-math.pi, math.pi, generator=generator)
-        magnitude = torch.ones_like(angle)
 
-        result = torch.polar(magnitude, angle)
+        result = torch.complex(angle.cos(), angle.sin())
         result.requires_grad = requires_grad
         return result.as_subclass(cls)
 

torchhd/functional.py

@@ -438,6 +438,13 @@ def circular_hv(
     Implements circular-hypervectors based on level-hypervectors as described in `An Extension to Basis-Hypervectors for Learning from Circular Data in Hyperdimensional Computing <https://arxiv.org/abs/2205.07920>`_.
     Any hypervector is quasi-orthogonal to the hypervector opposite site of the circle.
 
+    .. note::
+        Circular hypervectors cannot be created directly with Holographic Reduced Representations (HRR) because of imprecisions inherent to HRR. 
+        One way around this is to use FHRR for the creation of circular hypervectors and then transform them to HRR vectors. Example:
+
+        >>> hv = torchhd.circular_hv(10, 6, torchhd.FHRR)
+        >>> hv = torch.real(torch.fft.ifft(hv)).as_subclass(torchhd.HRR)
+
     Args:
         num_vectors (int): the number of hypervectors to generate.
         dimensions (int): the dimensionality of the hypervectors.

torchhd/hrr.py

@@ -1,8 +1,8 @@
 import torch
 from torch import Tensor
-import torch.nn.functional as F
 from torch.fft import fft, ifft
 from typing import Set
+import math
 
 from torchhd.base import VSA_Model
 
@@ -164,7 +164,7 @@ def random_hv(
 
         size = (num_vectors, dimensions)
         result = torch.empty(size, dtype=dtype, device=device)
-        result.normal_(0, 1.0 / dimensions, generator=generator)
+        result.normal_(0, 1.0 / math.sqrt(dimensions), generator=generator)
 
         result.requires_grad = requires_grad
         return result.as_subclass(cls)

torchhd/tests/basis_hv/test_circular_hv.py

@@ -69,12 +69,6 @@ def test_value(self, dtype, model):
         elif model == torchhd.MAP:
             assert torch.all((hv == -1) | (hv == 1)).item()
 
-        elif model == torchhd.HRR:
-            std, mean = torch.std_mean(hv)
-            assert torch.allclose(
-                mean, torch.tensor(0.0, dtype=mean.dtype), atol=0.0001
-            )
-
         elif model == torchhd.FHRR:
             mag = hv.abs()
             assert torch.allclose(mag, torch.tensor(1.0, dtype=mag.dtype))
@@ -140,14 +134,10 @@ def test_uses_default_dtype(self):
         torch.set_default_dtype(torch.float32)
         hv = functional.circular_hv(3, 52, torchhd.MAP)
         assert hv.dtype == torch.float32
-        # hv = functional.circular_hv(3, 52, torchhd.HRR)
-        # assert hv.dtype == torch.float32
 
         torch.set_default_dtype(torch.float64)
         hv = functional.circular_hv(3, 52, torchhd.MAP)
         assert hv.dtype == torch.float64
-        # hv = functional.circular_hv(3, 52, torchhd.HRR)
-        # assert hv.dtype == torch.float64
 
         hv = functional.circular_hv(3, 52, torchhd.FHRR)
         assert hv.dtype == torch.complex64
@@ -156,8 +146,5 @@ def test_requires_grad(self):
         hv = functional.circular_hv(3, 52, torchhd.MAP, requires_grad=True)
         assert hv.requires_grad == True
 
-        # hv = functional.circular_hv(3, 52, torchhd.HRR, requires_grad=True)
-        # assert hv.requires_grad == True
-
         hv = functional.circular_hv(3, 52, torchhd.FHRR, requires_grad=True)
         assert hv.requires_grad == True