[Scipy-svn] r2725 - trunk/Lib/sandbox/maskedarray

[scipy-svn at scipy.org]

Author: pierregm
Date: 2007-02-19 00:52:16 -0600 (Mon, 19 Feb 2007)
New Revision: 2725

Modified:
   trunk/Lib/sandbox/maskedarray/CHANGELOG
   trunk/Lib/sandbox/maskedarray/core.py
Log:
core : ensure that self._mask is nomask when setting the mask to nomask

Modified: trunk/Lib/sandbox/maskedarray/CHANGELOG
===================================================================
--- trunk/Lib/sandbox/maskedarray/CHANGELOG	2007-02-19 05:30:05 UTC (rev 2724)
+++ trunk/Lib/sandbox/maskedarray/CHANGELOG	2007-02-19 06:52:16 UTC (rev 2725)
@@ -31,7 +31,10 @@
 be overwritten far too easily. If you want to specify a new value for the mask,
 you should use the _setmask or __setmask__ methods: these methods ensure that
 the new mask has the same shape as the data, and that the _hardmask condition
-is respected.
+is respected. Note that in some particular cases involving subclasses of 
+MaskedArray, the mask is not always propagated properly. It is recommended to
+set the mask of the base object, instead of trying to set the mask through a 
+view of MaskedArray.
 
 Following the suggestions of Reggie Dugard, the class defaults have been
 suppressed. Unfortunately, that required to add some extra definitions in the
@@ -43,7 +46,15 @@
 The previous implementation of MaskedArray is called core_ini.py and it can be
 found in the alternative_versions directory. This folder contains also yet
 another implementation core_alt. This latter is left for documentation purposes,
-and should serve as a template when we'll port the package to C.
+and should serve as a template when we'll port the package to C. It introduces
+yet another attribute, _masklayer. This attribute is always a ndarray of booleans
+with the same shape as the data, that stores the values of the masked. The _mask
+attribute is then a property, that returns _masklayer or nomask depending on 
+the value of the _smallmask flag and the values of _masklayer. This approach 
+seems to solve the anomaly in mask propagation mentioned earlier. However, some
+performance tests show that this approach is significantly slower (from 10% to 
+50%) than the current implementation. It was therfore decided to leave it out of
+the main package.
 
 #...............................................................................
 2007-01-22 : core    : fixed a call to numpy.float128 on 32b machines

Modified: trunk/Lib/sandbox/maskedarray/core.py
===================================================================
--- trunk/Lib/sandbox/maskedarray/core.py	2007-02-19 05:30:05 UTC (rev 2724)
+++ trunk/Lib/sandbox/maskedarray/core.py	2007-02-19 06:52:16 UTC (rev 2725)
@@ -1163,7 +1163,13 @@
             if newmask is not nomask:
                 self._mask.__ior__(newmask)
         else:
-            self._mask.flat = newmask
+            # This one is tricky: if we set the mask that way, we may break the
+            # propagation. But if we don't, we end up with a mask full of False
+            # and a test on nomask fails...
+            if newmask is nomask:
+                self._mask = nomask
+            else:
+                self._mask.flat = newmask
         if self._mask.shape:
             self._mask.shape = self.shape
     _setmask = __setmask__