Fix accuracy of circular convolution.

The previous choice of intercept and eval point distribution was not
sufficiently validated. The choice in this commit performs better
and over a wider set of parameters.

CHANGES.rst

@@ -41,6 +41,9 @@ Release History
   (`#35 <https://github.com/nengo/nengo/pull/35>`_,
   `#32 <https://github.com/nengo/nengo/issues/32>`_,
   `#34 <https://github.com/nengo/nengo/issues/34>`_)
+- Improved accuracy by fixing choice of evaluation point and intercept
+  distributions.
+  (`#39 <https://github.com/nengo/nengo_spa/pull/39>`_)
 
 
 0.1.1 (May 19, 2017)

nengo_spa/networks/circularconvolution.py

@@ -49,7 +49,6 @@ def transform_in(dims, align, invert):
             tr[i] = row.real if i % 4 == 0 or i % 4 == 3 else row.imag
 
     remove_imag_rows(tr)
-    tr /= np.sqrt(dims)
     return tr.reshape((-1, dims))
 
 
@@ -67,6 +66,9 @@ def transform_out(dims):
 
     tr = tr.reshape(4*dims2, dims)
     remove_imag_rows(tr)
+    # IDFT has a 1/D scaling factor
+    tr /= dims
+
     return tr.T
 
 
@@ -193,14 +195,9 @@ def CircularConvolution(n_neurons, dimensions, invert_a=False, invert_b=False,
     with net:
         net.input_a = nengo.Node(size_in=dimensions, label="input_a")
         net.input_b = nengo.Node(size_in=dimensions, label="input_b")
-        with nengo.Config(nengo.Ensemble) as cfg:
-            cfg[nengo.Ensemble].eval_points = nengo.dists.CosineSimilarity(
-                2 * dimensions + 2)
-            cfg[nengo.Ensemble].intercepts = nengo.dists.CosineSimilarity(
-                2 * dimensions + 2)
-            net.product = Product(
-                n_neurons, tr_out.shape[1],
-                input_magnitude=input_magnitude / np.sqrt(2.))
+        net.product = Product(
+            n_neurons, tr_out.shape[1],
+            input_magnitude=2 * input_magnitude / np.sqrt(2.))
         net.output = nengo.Node(size_in=dimensions, label="output")
 
         nengo.Connection(