Changeset 9114:203a447f06f0 in orange

Timestamp:

10/17/11 16:56:53 (20 months ago)

Author:

ales_erjavec <ales.erjavec@…>

Branch:

default

Convert:

9c2fc928852e8e548524b1346dedfc1d132c441b

Message:

Fixed the format of return values from the multi-label regression.

Location:

orange/Orange/regression

Files:

• earth.py (modified) (4 diffs)
• pls.py (modified) (2 diffs)

orange/Orange/regression/earth.py

@@ -287 +287 @@
             return vals
         elif result_type == Orange.core.GetBoth:
-            return zip(vals, probs) if self.multi_flag else (vals, probs)
+            return vals, probs
         else:
             return probs
@@ -636 +636 @@
     return subsets, rss_vec
     
+def subset_selection_xtx2(X, Y):
+    """ Another implementation (this uses qr decomp).
+    """
+    from Orange.misc import linalg
+    X = numpy.asfortranarray(X, dtype=ctypes.c_double)
+    Y = numpy.asfortranarray(Y, dtype=ctypes.c_double)
+    col_count = X.shape[1]
+    working_set = range(col_count)
+    subsets = numpy.zeros((col_count, col_count), dtype=int)
+    rss_vec = numpy.zeros((col_count,))
+    QR, k, _, jpvt = linalg.qr_decomp(X)
+    
+    if k < col_count:
+        # remove jpvt[k:] from the work set. Will have zero 
+        # entries in the subsets matrix
+        for i in jpvt[k:]:
+            del working_set[i]
+        col_count = len(working_set)
+        
+    for subset_size in reversed(range(col_count)):
+        subsets[subset_size, :subset_size + 1] = working_set
+        X_work = X[:, working_set]
+        b, rsd, rank = linalg.qr_lstsq(X_work, Y)
+#        print rsd
+        rss_vec[subset_size] = numpy.sum(rsd ** 2)
+        XtX = numpy.dot(X_work.T, X_work)
+        iXtX = numpy.linalg.pinv(XtX)
+        diag = numpy.diag(iXtX)
+        
+        if subset_size == 0:
+            break
+        
+        delta_rss = b ** 2 / diag
+        delete_i = numpy.argmin(delta_rss[1:]) + 1 # Keep the intercept
+        del working_set[delete_i]
+    return subsets, rss_vec
     
 def pruning_pass(bx, y, penalty, pruned_terms=-1):
@@ -1021 +1057 @@
             x, y = data.to_numpy_MA("1A/c")
             try:
-                subsets, rss = subset_selection_xtx(x, y)
+                subsets, rss = subset_selection_xtx2(x, y)
             except numpy.linalg.LinAlgError:
                 subsets, rss = subset_selection_xtx_numpy(x, y)
@@ -1035 +1071 @@
         index = list(data.domain.attributes).index(attr)
         return rss[index]
-        
+    
 
 #from Orange.core import EarthLearner as BaseEarthLearner, \

orange/Orange/regression/pls.py

@@ -442 +442 @@
         ins = [instance[v].native() for v in self.xVars]
         
-        multilabel_flag = sum(data_label_mask(self.domain))
-        
         if "?" in ins: # missing value -> corresponding coefficient omitted
             def miss_2_0(x): return x if x != "?" else 0
@@ -461 +459 @@
                 probs.append(dist)
             if result_type == Orange.core.GetBoth:
-                return zip(y_hat, probs) if self.multilabel_flag else (y_hat[0], probs[0])
+                return (y_hat, probs) if self.multilabel_flag else (y_hat[0], probs[0])
             else:
                 return probs if self.multilabel_flag else probs[0]