--- a/pylearn/algorithms/mcRBM.py	Thu Oct 21 11:43:36 2010 -0400
+++ b/pylearn/algorithms/mcRBM.py	Thu Oct 21 14:47:49 2010 -0400
@@ -507,13 +507,28 @@
         rval._params = [rval.U, rval.W, rval.a, rval.b, rval.c]
         return rval
 
-def topological_connectivity(out_shape=(12,12), window_shape=(3,3), window_stride=(2,2),
-        **kwargs):
+def topological_connectivity(out_shape=(12,12), window_shape=(3,3), window_stride=(2,2), dtype=None):
+    """Return numpy R x C matrix with connection weights from R inputs (filters) to C output (covariance) units.
+    """
+    if window_stride[0] != window_stride[1]:
+        raise ValueError()
+    if out_shape[0] != out_shape[1]:
+        raise ValueError()
+    if window_shape[0] != window_shape[1]:
+        raise ValueError()
+
+    if dtype is None:
+        dtype = theano.config.floatX
+
+    ker = numpy.asarray([1., 2., 1.], dtype=dtype)
+    ker = numpy.outer(ker,ker)
+    ker /= ker.sum()
 
     in_shape = (window_stride[0] * out_shape[0],
             window_stride[1] * out_shape[1])
 
-    rval = numpy.zeros(in_shape + out_shape, dtype=theano.config.floatX)
+    tmp = numpy.zeros(in_shape, dtype=dtype)
+    rval = numpy.zeros(in_shape + out_shape, dtype=dtype)
     A,B,C,D = rval.shape
 
     # for each output position (out_r, out_c)
@@ -523,17 +538,10 @@
             for win_r in range(window_shape[0]):
                 for win_c in range(window_shape[1]):
                     # add 1 to the corresponding input location
-                    in_r = out_r * window_stride[0] + win_r
-                    in_c = out_c * window_stride[1] + win_c
-                    rval[in_r%A, in_c%B, out_r%C, out_c%D] += 1
-
-    # This normalization algorithm is a guess, based on inspection of the matrix loaded from 
-    # see CVPR2010paper_material/topo2D_3x3_stride2_576filt.mat
-    rval = rval.reshape((A*B, C*D))
-    rval = (rval.T / rval.sum(axis=1)).T
-
-    rval /= rval.sum(axis=0)
-    return rval
+                    in_r = out_r * window_stride[0] + win_r - (window_shape[0]-1)//2
+                    in_c = out_c * window_stride[1] + win_c - (window_shape[1]-1)//2
+                    rval[(in_c+A)%A, (in_r+B)%B, out_r%C, out_c%D] = ker[win_r, win_c]
+    return rval.reshape((A*B,C*D))
 
 class mcRBM_withP(mcRBM):
     """Light-weight class that provides the math related to inference
@@ -598,8 +606,10 @@
     @classmethod
     def alloc_topo_P(cls, n_I, n_J, p_out_shape=(12,12), p_win_shape=(3,3), p_win_stride=(2,2),
             **kwargs):
+        # the factor of 2 is for some degree of compatibility with Marc'Aurelio's code the 'w2'
+        # in his code is -0.5 * self.P
         return cls.alloc_with_P(
-                -topological_connectivity(p_out_shape, p_win_shape, p_win_stride),
+                -2*topological_connectivity(p_out_shape, p_win_shape, p_win_stride),
                 n_I=n_I, n_J=n_J, **kwargs)
 
     @classmethod