Merge pull request #747 from mateiz/improved-lr

Update the Python logistic regression example

1 files changed, 26 insertions, 27 deletions

diff --git a/python/examples/logistic_regression.py b/python/examples/logistic_regression.py
index 3ac1bae4e9..1117dea538 100755
--- a/python/examples/logistic_regression.py
+++ b/python/examples/logistic_regression.py
@@ -16,7 +16,8 @@
 #
 
 """
-This example requires numpy (http://www.numpy.org/)
+A logistic regression implementation that uses NumPy (http://www.numpy.org) to act on batches
+of input data using efficient matrix operations.
 """
 from collections import namedtuple
 from math import exp
@@ -27,47 +28,45 @@ import numpy as np
 from pyspark import SparkContext
 
 
-N = 100000  # Number of data points
 D = 10  # Number of dimensions
-R = 0.7   # Scaling factor
-ITERATIONS = 5
-np.random.seed(42)
 
 
-DataPoint = namedtuple("DataPoint", ['x', 'y'])
-from logistic_regression import DataPoint  # So that DataPoint is properly serialized
-
-
-def generateData():
-    def generatePoint(i):
-        y = -1 if i % 2 == 0 else 1
-        x = np.random.normal(size=D) + (y * R)
-        return DataPoint(x, y)
-    return [generatePoint(i) for i in range(N)]
-
+# Read a batch of points from the input file into a NumPy matrix object. We operate on batches to
+# make further computations faster.
+# The data file contains lines of the form <label> <x1> <x2> ... <xD>. We load each block of these
+# into a NumPy array of size numLines * (D + 1) and pull out column 0 vs the others in gradient().
+def readPointBatch(iterator):
+    strs = list(iterator)
+    matrix = np.zeros((len(strs), D + 1))
+    for i in xrange(len(strs)):
+        matrix[i] = np.fromstring(strs[i].replace(',', ' '), dtype=np.float32, sep=' ')
+    return [matrix]
 
 if __name__ == "__main__":
-    if len(sys.argv) == 1:
-        print >> sys.stderr, "Usage: logistic_regression <master> [<slices>]"
+    if len(sys.argv) != 4:
+        print >> sys.stderr, "Usage: logistic_regression <master> <file> <iters>"
         exit(-1)
     sc = SparkContext(sys.argv[1], "PythonLR", pyFiles=[realpath(__file__)])
-    slices = int(sys.argv[2]) if len(sys.argv) > 2 else 2
-    points = sc.parallelize(generateData(), slices).cache()
+    points = sc.textFile(sys.argv[2]).mapPartitions(readPointBatch).cache()
+    iterations = int(sys.argv[3])
 
     # Initialize w to a random value
     w = 2 * np.random.ranf(size=D) - 1
     print "Initial w: " + str(w)
 
+    # Compute logistic regression gradient for a matrix of data points
+    def gradient(matrix, w):
+        Y = matrix[:,0]    # point labels (first column of input file)
+        X = matrix[:,1:]   # point coordinates
+        # For each point (x, y), compute gradient function, then sum these up
+        return ((1.0 / (1.0 + np.exp(-Y * X.dot(w))) - 1.0) * Y * X.T).sum(1)
+
     def add(x, y):
         x += y
         return x
 
-    for i in range(1, ITERATIONS + 1):
-        print "On iteration %i" % i
-
-        gradient = points.map(lambda p:
-            (1.0 / (1.0 + exp(-p.y * np.dot(w, p.x)))) * p.y * p.x
-        ).reduce(add)
-        w -= gradient
+    for i in range(iterations):
+        print "On iteration %i" % (i + 1)
+        w -= points.map(lambda m: gradient(m, w)).reduce(add)
 
     print "Final w: " + str(w)