ODIn/spark-instrumented-optimizer
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Tests for the Python side of the mllib bindings.

Browse source
This commit is contained in:
Tor Myklebust 2013-12-20 01:33:32 -05:00
parent 73e17064c6
commit 2940201ad8
1 changed files with 172 additions and 52 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

224
python/pyspark/mllib.py
View file

@ -1,4 +1,5 @@
from numpy import *
from pyspark import SparkContext
# Double vector format:
#
@ -7,44 +8,106 @@ from numpy import *
# Double matrix format:
#
# [8-byte 2] [8-byte rows] [8-byte cols] [rows*cols*8 bytes of data]
#
#
# This is all in machine-endian. That means that the Java interpreter and the
# Python interpreter must agree on what endian the machine is.
def _deserialize_byte_array(shape, ba, offset):
"""Wrapper around ndarray aliasing hack.
>>> x = array([1.0, 2.0, 3.0, 4.0, 5.0])
>>> array_equal(x, _deserialize_byte_array(x.shape, x.data, 0))
True
>>> x = array([1.0, 2.0, 3.0, 4.0]).reshape(2,2)
>>> array_equal(x, _deserialize_byte_array(x.shape, x.data, 0))
True
"""
ar = ndarray(shape=shape, buffer=ba, offset=offset, dtype="float64",
order='C')
return ar.copy()
def _serialize_double_vector(v):
if (type(v) == ndarray and v.dtype == float64 and v.ndim == 1):
length = v.shape[0]
ba = bytearray(16 + 8*length)
header = ndarray(shape=[2], buffer=ba, dtype="int64")
header[0] = 1
header[1] = length
copyto(ndarray(shape=[length], buffer=ba, offset=16,
dtype="float64"), v)
return ba
else:
raise TypeError("_serialize_double_vector called on a "
"non-double-vector")
"""Serialize a double vector into a mutually understood format.
>>> _serialize_double_vector(array([]))
bytearray(b'\\x01\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00')
>>> _serialize_double_vector(array([0.0, 1.0]))
bytearray(b'\\x01\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x02\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\xf0?')
>>> _serialize_double_vector("hello, world")
Traceback (most recent call last):
File "/usr/lib/python2.7/doctest.py", line 1289, in __run
compileflags, 1) in test.globs
File "<doctest __main__._serialize_double_vector[1]>", line 1, in <module>
_serialize_double_vector("hello, world")
File "python/pyspark/mllib.py", line 41, in _serialize_double_vector
raise TypeError("_serialize_double_vector called on a %s; wanted ndarray" % type(v))
TypeError: _serialize_double_vector called on a <type 'str'>; wanted ndarray
>>> _serialize_double_vector(array([0, 1]))
Traceback (most recent call last):
File "/usr/lib/python2.7/doctest.py", line 1289, in __run
compileflags, 1) in test.globs
File "<doctest __main__._serialize_double_vector[2]>", line 1, in <module>
_serialize_double_vector(array([0, 1]))
File "python/pyspark/mllib.py", line 51, in _serialize_double_vector
"wanted ndarray of float64" % v.dtype)
TypeError: _serialize_double_vector called on an ndarray of int64; wanted ndarray of float64
>>> _serialize_double_vector(array([0.0, 1.0, 2.0, 3.0]).reshape(2,2))
Traceback (most recent call last):
File "/usr/lib/python2.7/doctest.py", line 1289, in __run
compileflags, 1) in test.globs
File "<doctest __main__._serialize_double_vector[3]>", line 1, in <module>
_serialize_double_vector(array([0.0, 1.0, 2.0, 3.0]).reshape(2,2))
File "python/pyspark/mllib.py", line 62, in _serialize_double_vector
"wanted a 1darray" % v.ndim)
TypeError: _serialize_double_vector called on a 2darray; wanted a 1darray
"""
if type(v) != ndarray:
raise TypeError("_serialize_double_vector called on a %s; "
"wanted ndarray" % type(v))
if v.dtype != float64:
raise TypeError("_serialize_double_vector called on an ndarray of %s; "
"wanted ndarray of float64" % v.dtype)
if v.ndim != 1:
raise TypeError("_serialize_double_vector called on a %ddarray; "
"wanted a 1darray" % v.ndim)
length = v.shape[0]
ba = bytearray(16 + 8*length)
header = ndarray(shape=[2], buffer=ba, dtype="int64")
header[0] = 1
header[1] = length
copyto(ndarray(shape=[length], buffer=ba, offset=16,
dtype="float64"), v)
return ba
def _deserialize_double_vector(ba):
if (type(ba) == bytearray and len(ba) >= 16 and (len(ba) & 7 == 0)):
header = ndarray(shape=[2], buffer=ba, dtype="int64")
if (header[0] != 1):
raise TypeError("_deserialize_double_vector called on bytearray "
"with wrong magic")
length = header[1]
if (len(ba) != 8*length + 16):
raise TypeError("_deserialize_double_vector called on bytearray "
"with wrong length")
return _deserialize_byte_array([length], ba, 16)
else:
raise TypeError("_deserialize_double_vector called on a non-bytearray")
"""Deserialize a double vector from a mutually understood format.
>>> x = array([1.0, 2.0, 3.0, 4.0, -1.0, 0.0, -0.0])
>>> array_equal(x, _deserialize_double_vector(_serialize_double_vector(x)))
True
"""
if type(ba) != bytearray:
raise TypeError("_deserialize_double_vector called on a %s; "
"wanted bytearray" % type(ba))
if len(ba) < 16:
raise TypeError("_deserialize_double_vector called on a %d-byte array, "
"which is too short" % len(ba))
if (len(ba) & 7) != 0:
raise TypeError("_deserialize_double_vector called on a %d-byte array, "
"which is not a multiple of 8" % len(ba))
header = ndarray(shape=[2], buffer=ba, dtype="int64")
if header[0] != 1:
raise TypeError("_deserialize_double_vector called on bytearray "
"with wrong magic")
length = header[1]
if len(ba) != 8*length + 16:
raise TypeError("_deserialize_double_vector called on bytearray "
"with wrong length")
return _deserialize_byte_array([length], ba, 16)
def _serialize_double_matrix(m):
"""Serialize a double matrix into a mutually understood format.
"""
if (type(m) == ndarray and m.dtype == float64 and m.ndim == 2):
rows = m.shape[0]
cols = m.shape[1]
@ -61,22 +124,31 @@ def _serialize_double_matrix(m):
"non-double-matrix")
def _deserialize_double_matrix(ba):
if (type(ba) == bytearray and len(ba) >= 24 and (len(ba) & 7 == 0)):
header = ndarray(shape=[3], buffer=ba, dtype="int64")
if (header[0] != 2):
raise TypeError("_deserialize_double_matrix called on bytearray "
"with wrong magic")
rows = header[1]
cols = header[2]
if (len(ba) != 8*rows*cols + 24):
raise TypeError("_deserialize_double_matrix called on bytearray "
"with wrong length")
return _deserialize_byte_array([rows, cols], ba, 24)
else:
raise TypeError("_deserialize_double_matrix called on a non-bytearray")
"""Deserialize a double matrix from a mutually understood format.
"""
if type(ba) != bytearray:
raise TypeError("_deserialize_double_matrix called on a %s; "
"wanted bytearray" % type(ba))
if len(ba) < 24:
raise TypeError("_deserialize_double_matrix called on a %d-byte array, "
"which is too short" % len(ba))
if (len(ba) & 7) != 0:
raise TypeError("_deserialize_double_matrix called on a %d-byte array, "
"which is not a multiple of 8" % len(ba))
header = ndarray(shape=[3], buffer=ba, dtype="int64")
if (header[0] != 2):
raise TypeError("_deserialize_double_matrix called on bytearray "
"with wrong magic")
rows = header[1]
cols = header[2]
if (len(ba) != 8*rows*cols + 24):
raise TypeError("_deserialize_double_matrix called on bytearray "
"with wrong length")
return _deserialize_byte_array([rows, cols], ba, 24)
def _linear_predictor_typecheck(x, coeffs):
"""Predict the class of the vector x."""
"""Check that x is a one-dimensional vector of the right shape.
This is a temporary hackaround until I actually implement bulk predict."""
if type(x) == ndarray:
if x.ndim == 1:
if x.shape == coeffs.shape:
@ -98,12 +170,17 @@ class LinearModel(object):
self._intercept = intercept
class LinearRegressionModelBase(LinearModel):
"""A linear regression model."""
"""A linear regression model.
>>> lrmb = LinearRegressionModelBase(array([1.0, 2.0]), 0.1)
>>> abs(lrmb.predict(array([-1.03, 7.777])) - 14.624) < 1e-6
True
"""
def predict(self, x):
"""Predict the value of the dependent variable given a vector x"""
"""containing values for the independent variables."""
_linear_predictor_typecheck(x, _coeff)
return dot(_coeff, x) + _intercept
_linear_predictor_typecheck(x, self._coeff)
return dot(self._coeff, x) + self._intercept
# Map a pickled Python RDD of numpy double vectors to a Java RDD of
# _serialized_double_vectors
@ -145,7 +222,11 @@ def _regression_train_wrapper(sc, train_func, klass, data, initial_weights):
return klass(_deserialize_double_vector(ans[0]), ans[1]);
class LinearRegressionModel(LinearRegressionModelBase):
"""A linear regression model derived from a least-squares fit."""
"""A linear regression model derived from a least-squares fit.
>>> data = array([0.0, 0.0, 1.0, 1.0, 3.0, 2.0, 2.0, 3.0]).reshape(4,2)
>>> lrm = LinearRegressionModel.train(sc, sc.parallelize(data), initial_weights=array([1.0]))
"""
@classmethod
def train(cls, sc, data, iterations=100, step=1.0,
mini_batch_fraction=1.0, initial_weights=None):
@ -156,8 +237,12 @@ class LinearRegressionModel(LinearRegressionModelBase):
LinearRegressionModel, data, initial_weights)
class LassoModel(LinearRegressionModelBase):
"""A linear regression model derived from a least-squares fit with an """
"""l_1 penalty term."""
"""A linear regression model derived from a least-squares fit with an
l_1 penalty term.
>>> data = array([0.0, 0.0, 1.0, 1.0, 3.0, 2.0, 2.0, 3.0]).reshape(4,2)
>>> lrm = LassoModel.train(sc, sc.parallelize(data), initial_weights=array([1.0]))
"""
@classmethod
def train(cls, sc, data, iterations=100, step=1.0, reg_param=1.0,
mini_batch_fraction=1.0, initial_weights=None):
@ -168,8 +253,12 @@ class LassoModel(LinearRegressionModelBase):
LassoModel, data, initial_weights)
class RidgeRegressionModel(LinearRegressionModelBase):
"""A linear regression model derived from a least-squares fit with an """
"""l_2 penalty term."""
"""A linear regression model derived from a least-squares fit with an
l_2 penalty term.
>>> data = array([0.0, 0.0, 1.0, 1.0, 3.0, 2.0, 2.0, 3.0]).reshape(4,2)
>>> lrm = RidgeRegressionModel.train(sc, sc.parallelize(data), initial_weights=array([1.0]))
"""
@classmethod
def train(cls, sc, data, iterations=100, step=1.0, reg_param=1.0,
mini_batch_fraction=1.0, initial_weights=None):
@ -180,7 +269,11 @@ class RidgeRegressionModel(LinearRegressionModelBase):
RidgeRegressionModel, data, initial_weights)
class LogisticRegressionModel(LinearModel):
"""A linear binary classification model derived from logistic regression."""
"""A linear binary classification model derived from logistic regression.
>>> data = array([0.0, 0.0, 1.0, 1.0, 1.0, 2.0, 1.0, 3.0]).reshape(4,2)
>>> lrm = LogisticRegressionModel.train(sc, sc.parallelize(data))
"""
def predict(self, x):
_linear_predictor_typecheck(x, _coeff)
margin = dot(x, _coeff) + intercept
@ -197,7 +290,11 @@ class LogisticRegressionModel(LinearModel):
LogisticRegressionModel, data, initial_weights)
class SVMModel(LinearModel):
"""A support vector machine."""
"""A support vector machine.
>>> data = array([0.0, 0.0, 1.0, 1.0, 1.0, 2.0, 1.0, 3.0]).reshape(4,2)
>>> svm = SVMModel.train(sc, sc.parallelize(data))
"""
def predict(self, x):
_linear_predictor_typecheck(x, _coeff)
margin = dot(x, _coeff) + intercept
@ -212,15 +309,24 @@ class SVMModel(LinearModel):
SVMModel, data, initial_weights)
class KMeansModel(object):
"""A clustering model derived from the k-means method."""
"""A clustering model derived from the k-means method.
>>> data = array([0.0, 0.0, 1.0,1.0, 9.0,8.0, 8.0,9.0]).reshape(4,2)
>>> clusters = KMeansModel.train(sc, sc.parallelize(data), 2, maxIterations=10, runs=30, initialization_mode="random")
>>> clusters.predict(array([0.0, 0.0])) == clusters.predict(array([1.0, 1.0]))
True
>>> clusters.predict(array([8.0, 9.0])) == clusters.predict(array([9.0, 8.0]))
True
>>> clusters = KMeansModel.train(sc, sc.parallelize(data), 2)
"""
def __init__(self, centers_):
self.centers = centers_
def predict(self, x):
best = 0
best_distance = 1e75
for i in range(0, centers.shape[0]):
diff = x - centers[i]
for i in range(0, self.centers.shape[0]):
diff = x - self.centers[i]
distance = sqrt(dot(diff, diff))
if distance < best_distance:
best = i
@ -239,3 +345,17 @@ class KMeansModel(object):
raise RuntimeError("JVM call result had first element of type "
+ type(ans[0]) + " which is not bytearray");
return KMeansModel(_deserialize_double_matrix(ans[0]));
def _test():
import doctest
globs = globals().copy()
globs['sc'] = SparkContext('local[4]', 'PythonTest', batchSize=2)
(failure_count, test_count) = doctest.testmod(globs=globs,
optionflags=doctest.ELLIPSIS)
globs['sc'].stop()
print failure_count,"failures among",test_count,"tests"
if failure_count:
exit(-1)
if __name__ == "__main__":
_test()