[Scipy-svn] r2039 - trunk/Lib/sandbox/svm

[scipy-svn at scipy.org]

Author: fullung
Date: 2006-07-05 10:07:14 -0500 (Wed, 05 Jul 2006)
New Revision: 2039

Added:
   trunk/Lib/sandbox/svm/dataset.py
Modified:
   trunk/Lib/sandbox/svm/kernel.py
Log:
New design to hide libsvm data structure from the users.


Added: trunk/Lib/sandbox/svm/dataset.py
===================================================================
--- trunk/Lib/sandbox/svm/dataset.py	2006-07-05 14:53:24 UTC (rev 2038)
+++ trunk/Lib/sandbox/svm/dataset.py	2006-07-05 15:07:14 UTC (rev 2039)
@@ -0,0 +1,98 @@
+__all__ = [
+    'LibSvmRegressionDataSet',
+    'LibSvmClassificationDataSet',
+    'LibSvmOneClassDataSet'
+    ]
+
+import numpy as N
+
+import libsvm
+
+def svm_node_dot(x, y):
+    # associate node indexes with array indexes
+    xidx = dict(zip(x['index'][:-1],range(0,len(x))))
+    yidx = dict(zip(y['index'][:-1],range(0,len(y))))
+    # indexes in either vector
+    indexes = N.unique(N.hstack([x['index'],y['index']]))
+    z = 0.
+    for j in indexes:
+        if j in xidx and j in yidx:
+            # dot if index is present in both vectors
+            z += x['value'][xidx[j]]*y['value'][yidx[j]]
+    return z
+
+class LibSvmDataSet:
+    def __init__(self, data, kernel):
+        self.data = data
+        self.kernel = kernel
+
+    def precompute(self):
+        n = 1 + len(data) + 1
+        a = N.zeros((n,), dtype=svm_node_dtype)
+        a[-1] = -1, 0. # end of record marker
+        grammat = [a.copy() for i in range(len(data))]
+        for i, a in enumerate(grammat):
+            a[0] = 0, i + 1 # id
+        for i in range(len(data)):
+            for j in range(i, len(data)):
+                z = self.kernel(data[i], N.transpose(data[j]))
+                grammat[i][j+1]['value'] = z
+                grammat[j][i+1]['value'] = z
+        return LibSvmPrecomputedDataSet(grammat, self.data, self.kernel)
+
+class LibSvmPrecomputedDataSet:
+    def __init__(self, grammat, data, kernel):
+        self.grammat = grammat
+        self.data = data
+        self.kernel = kernel
+
+    def extend(self, data):
+        raise NotImplementedError
+
+class LibSvmRegressionDataSet(LibSvmDataSet):
+    def __init__(self, data):
+        f = lambda x: (x[0], convert_to_svm_node(x[1]))
+        LibSvmDataSet.__init__(self, map(f, data))
+
+class LibSvmClassificationDataSet(LibSvmDataSet):
+    def __init__(self, data):
+        labels = N.array(map(lambda x: x[0]), dtype=N.intc)
+        assert N.alltrue(labels >= 0), \
+            'labels must be non-negative integers'
+        f = lambda x: (x[0],convert_to_svm_node(x[1]))
+        LibSvmDataSet.__init__(self, map(f, data))
+
+class LibSvmOneClassDataSet(LibSvmDataSet):
+    def __init__(self, data):
+        f = map(lambda x: tuple([0,convert_to_svm_node(x)]))
+        LibSvmDataSet.__init__(self, map(f, data))
+
+def convert_to_svm_node(x):
+    y = N.empty(len(x)+1, dtype=libsvm.svm_node_dtype)
+    y[-1] = (-1, 0.)
+    if isinstance(x, dict):
+        x = x.items()
+    if isinstance(x, list):
+        x.sort(cmp=lambda x,y: cmp(x[0],y[0]))
+        y[:-1] = x
+    else:
+        y['index'][:-1] = N.arange(1,len(x)+1)
+        y['value'][:-1] = x
+    assert N.alltrue(y[:-1]['index'] >= 1), \
+        'indexes must be positive'
+    assert len(x) == len(N.unique(y[:-1]['index'])), \
+        'indexes must be unique'
+    return y
+
+def svm_node_dot(x, y):
+    # associate node indexes with array indexes
+    xidx = dict(zip(x['index'][:-1],range(0,len(x))))
+    yidx = dict(zip(y['index'][:-1],range(0,len(y))))
+    # indexes in either vector
+    indexes = N.unique(N.hstack([x['index'],y['index']]))
+    z = 0.
+    for j in indexes:
+        if j in xidx and j in yidx:
+            # dot if index is present in both vectors
+            z += x['value'][xidx[j]]*y['value'][yidx[j]]
+    return z

Modified: trunk/Lib/sandbox/svm/kernel.py
===================================================================
--- trunk/Lib/sandbox/svm/kernel.py	2006-07-05 14:53:24 UTC (rev 2038)
+++ trunk/Lib/sandbox/svm/kernel.py	2006-07-05 15:07:14 UTC (rev 2039)
@@ -6,28 +6,15 @@
     'CustomKernel'
     ]
 
-def svm_node_dot(x, y):
-    # associate node indexes with array indexes
-    xidx = dict(zip(x['index'][:-1],range(0,len(x))))
-    yidx = dict(zip(y['index'][:-1],range(0,len(y))))
-    # indexes in either vector
-    indexes = N.unique(N.hstack([x['index'],y['index']]))
-    z = 0.
-    for j in indexes:
-        if j in xidx and j in yidx:
-            # dot if index is present in both vectors
-            z += x['value'][xidx[j]]*y['value'][yidx[j]]
-    return z
-
 class LinearKernel:
-    def __init__(self, dot=svm_node_dot):
+    def __init__(self, dot):
         self.dot = dot
 
     def __call__(self, x, y):
         return self.dot(x, y)
 
 class PolynomialKernel:
-    def __init__(self, degree, gamma, coef0, dot=svm_node_dot):
+    def __init__(self, degree, gamma, coef0, dot):
         self.degree = degree
         self.gamma = gamma
         self.coef0 = coef0
@@ -45,7 +32,7 @@
         return ret
 
 class RBFKernel:
-    def __init__(self, gamma, dot=svm_node_dot):
+    def __init__(self, gamma, dot):
         self.gamma = gamma
         self.dot = dot
 
@@ -54,7 +41,7 @@
         return N.exp(-self.gamma*z)
 
 class SigmoidKernel:
-    def __init__(self, gamma, coef0, dot=svm_node_dot):
+    def __init__(self, gamma, coef0, dot):
         self.gamma = gamma
         self.coef0 = coef0
         self.dot = dot
@@ -63,11 +50,7 @@
         return N.tanh(self.gamma*self.dot(x, y)+self.coef0)
 
 class CustomKernel:
-    """
-    XXX example CustomKernel(lambda x, y, d: d(x,y))
-    """
-
-    def __init__(self, f, dot=svm_node_dot):
+    def __init__(self, f, dot):
         self.f = f
         self.dot = dot