Multiple lib changes

fluents/lib/select_generators.py

@@ -30,22 +30,14 @@ def w_pls_gen(aat,b,n_blocks=None,center=True,index_out=False):
           b_in = b[inn,:]
           b_out = b[out,:]
           if center:
-               # centering projector: I - (1/n)11'
-               # nin = len(inn)
-               # Pc = eye(nin) - outer(ones((nin,)),ones((nin,)))/nin
-               # xxt - x( outer(ones((nin,)),ones((nin,)))/nin ) x.T
-
-               # de jong:
-               h = sum(aat_in,0)[ :,newaxis]
-               h = (h - mean(h)/2)/len(inn)
-               mn_a = h + h.T
-               aat_in = aat_in - mn_a
+               aat_in, mn = outerprod_centering(aat_in)
+               aat_out = aat_out - mn
           if index_out:
                yield aat_in,aat_out,b_in,b_out,out
           else:
                yield aat_in,aat_out,b_in,b_out
 
-def pls_gen(a,b, n_blocks=None, center=False, index_out=False,axis=0):
+def pls_gen(a, b, n_blocks=None, center=False, index_out=False,axis=0, metric=None):
      """Random block crossvalidation
     Leave-one-out is a subset, with n_blocks equals a.shape[-1]
     """
@@ -56,17 +48,38 @@ def pls_gen(a,b, n_blocks=None, center=False, index_out=False,axis=0):
      out_ind_sets = [index[i*n_in_set:(i+1)*n_in_set] for i in range(n_blocks)]
      for out in out_ind_sets:
          inn = [i for i in index if i not in out]
+         acal = a.take(inn, 0)
+         atrue = a.take(out, 0)
+         bcal = b.take(inn, 0)
+         btrue = b.take(out, 0)
          if center:
-              a = a - mean(a,0)[newaxis]
-              b = b - mean(b,0)[newaxis]
+              mn_a = acal.mean(0)[newaxis]
+              acal = acal - mn_a
+              atrue = atrue - mn_a
+              mn_b = bcal.mean(0)[newaxis]
+              bcal = bcal - mn_b
+              btrue = btrue - mn_b
+         if metric!=None:
+              acal = dot(acal, metric)
          if index_out:
-              yield a.take(inn,0),a.take(out,0), b.take(inn,0),b.take(out,0),out
+              yield acal, atrue, bcal, btrue, out
          else:     
-              yield a.take(inn,0),a.take(out,0), b.take(inn,0),b.take(out,0)
+              yield acal, atrue, bcal, btrue
 
          
-def pca_gen(a,n_sets=None, center=False, index_out=False,axis=0):
-     """PCA random block crossval generator.
+def pca_gen(a, n_sets=None, center=False, index_out=False, axis=0):
+     """Returns a generator of crossvalidation sample segments.
+
+     input:
+           -- a, data matrix (m x n)
+           -- n_sets, number of segments/subsets to generate.
+           -- center, bool, choice of centering each subset
+           -- index_out, bool, return subset index
+           -- axis, int, which axis to get subset from
+
+     ouput:
+           -- V, generator with (n_sets) memebers (subsets)
+     
      """
      m = a.shape[axis]
      index = randperm(m)
@@ -76,21 +89,26 @@ def pca_gen(a,n_sets=None, center=False, index_out=False,axis=0):
      out_ind_sets = [index[i*n_in_set:(i+1)*n_in_set] for i in range(n_sets)]
      for out in out_ind_sets:
          inn = [i for i in index if i not in out]
+         acal = a.take(inn, 0)
+         atrue = a.take(out, 0)
          if center:
-              a = a - mean(a,0)[newaxis]
+              mn_a = acal.mean(0)[newaxis]
+              acal = acal - mn_a
+              atrue = atrue - mn_a
          if index_out:
-              yield a.take(inn,0),a.take(out,0),out
-         else:     
-              yield a.take(inn,0),a.take(out,0)
+              yield acal, atrue, out
+         else:
+              yield acal, atrue
 
-def w_pls_gen_jk(a,b,n_sets=None,center=True,index_out=False,axis=0):
+def w_pls_gen_jk(a, b, n_sets=None, center=True,
+                 index_out=False, axis=0):
      """Random block crossvalidation for wide X (m>>n)
      Leave-one-out is a subset, with n_sets equals a.shape[-1]
 
      Returns : X_m and X_m'Y_m
      """
      m = a.shape[axis]
-     ab = dot(a.T,b)
+     ab = dot(a.T, b)
      index = randperm(m)
      if n_sets==None:
           n_sets = m
@@ -103,19 +121,18 @@ def w_pls_gen_jk(a,b,n_sets=None,center=True,index_out=False,axis=0):
          a_in = a[inn,:]
          mn_a = 0
          mAB = 0
-         
          if center:
-              mn_a = mean(a,0)[newaxis]
-              mAin = dot(-ones((1,nout)),a[out,:])/nin
-              mBin = dot(-ones((1,nout)),b[out,:])/nin
-              mAB = dot(mAin.T,(mBin*nin))
-         ab_in = ab - dot(a[out,].T,b[out,:]) - mAB
+              mn_a = a_in.mean(0)[newaxis]
+              mAin = dot(-ones((1,nout)), a[out,:])/nin
+              mBin = dot(-ones((1,nout)), b[out,:])/nin
+              mAB = dot(mAin.T, (mBin*nin))
+         ab_in = ab - dot(a[out,].T, b[out,:]) - mAB
          a_in = a_in - mn_a
 
          if index_out:
-              yield ain,ab, out
+              yield a_in, ab_in, out
          else:     
-              yield a_in, ab
+              yield a_in, ab_in
 
 def shuffle_1d_block(a, n_sets=None, blocks=None, index_out=False, axis=0):
      """Random block shuffling along 1d axis
@@ -185,3 +202,19 @@ def diag_pert(a, n_sets=10, center=True, index_out=False):
              yield a_out, asarray(out)
         else:
              yield a_out
+
+
+def outerprod_centering(aat, ret_mn=True):
+    """Returns mean centered symmetric outerproduct matrix.
+    """
+    n = aat.shape[0]
+    h = aat.sum(0)[:,newaxis]
+    h = (h - mean(h)/2)/n
+    mn_a = h + h.T
+    aatc = aat - mn_a
+    if ret_mn:
+        return aatc, aat.mean(0)
+    return aat - mn_a
+     
+     
+