def test_ols_equivalence(self):
# Simple ols problem
# y = a * x + b + r
# Where r ~ N(0, 1)
n = 40
a = 0.27
b = 1.2
x = np.arange(0, n)
r = np.random.normal(0, 1, n)
y = (a * x + b + r).reshape(n, 1)
features = x.reshape(n, 1)
features = np.concatenate([features, np.ones_like(features)], axis=1)
df = self.spark.createDataFrame(
[(Vectors.dense(y[i]), Matrices.dense(1, 2, features[i]))
for i in range(n)], ["measurement", "measurementModel"])
lkf = LinearKalmanFilter()\
.setInitialStateMean(Vectors.dense(0.0, 0.0))\
.setMeasurementModelCol("measurementModel")\
.setMeasurementCol("measurement")\
.setInitialStateCovariance(Matrices.dense(2, 2, (np.eye(2)*10).reshape(4, 1)))\
.setProcessModel(Matrices.dense(2, 2, np.eye(2).reshape(4, 1)))\
.setProcessNoise(Matrices.dense(2, 2, np.zeros(4)))\
.setMeasurementNoise(Matrices.dense(1, 1, [10E-5]))
model = lkf.transform(df)
state = model.filter(
"stateIndex = {}".format(n)).collect()[0].state.mean.values
# Check equivalence with least squares solution with numpy
expected, _, _, _ = np.linalg.lstsq(features, y, rcond=None)
np.testing.assert_array_almost_equal(state,
expected.reshape(2),
decimal=5)
def test_multiple_model_adaptive_filter(self):
n = 100
a = 0.27
b = 1.2
x = np.concatenate([np.arange(0, n), np.arange(0, n)])
r = np.random.normal(0, 1, n * 2)
y = (a * x + b + r).reshape(n * 2, 1)
features = x.reshape(n * 2, 1)
features = np.concatenate([features, np.ones_like(features)], axis=1)
state_keys = ["1"] * n + ["2"] * n
df = self.spark.createDataFrame(
[(state_keys[i], Vectors.dense(
y[i]), Matrices.dense(1, 2, features[i]))
for i in range(n * 2)],
["state_key", "measurement", "measurementModel"])
mmaeFilter = LinearKalmanFilter(2, 1)\
.setStateKeyCol("state_key")\
.setMeasurementModelCol("measurementModel")\
.setMeasurementCol("measurement")\
.setInitialCovariance(Matrices.dense(2, 2, (np.eye(2)*10).reshape(4, 1)))\
.setProcessModel(Matrices.dense(2, 2, np.eye(2).reshape(4, 1)))\
.setProcessNoise(Matrices.dense(2, 2, np.zeros(4)))\
.setMeasurementNoise(Matrices.dense(1, 1, [1.0]))\
.setSlidingLikelihoodWindow(5)\
.setEnableMultipleModelAdaptiveEstimation()
model = mmaeFilter.transform(df)
state = model.filter(
"stateIndex = {}".format(n)).collect()[0].state.values
expected, _, _, _ = np.linalg.lstsq(features, y, rcond=None)
np.testing.assert_array_almost_equal(state,
expected.reshape(2),
decimal=0)
def test_matrix_multiply(self):
mat1 = Matrices.dense(2, 2, [1.0, 2.0, 3.0, 4.0])
mat2 = Matrices.dense(2, 2, [4.0, 5.0, 6.0, 7.0])
result = self.spark.createDataFrame([(mat1, mat2)], ["mat1", "mat2"])\
.withColumn("result", multiplyMatrix("mat1", "mat2"))\
.select("result").head().result
np.testing.assert_array_almost_equal(
result.toArray(), np.dot(mat1.toArray(), mat2.toArray()))
def test_matrix_projection(self):
mat = Matrices.dense(2, 2, [1.0, 2.0, 3.0, 4.0])
proj = Matrices.dense(1, 2, [4.0, 5.0])
result = self.spark.createDataFrame([(mat, proj)], ["mat", "proj"]) \
.withColumn("result", projectMatrix("mat", "proj")) \
.select("result").head().result
expected = np.dot(np.dot(proj.toArray(), mat.toArray()),
proj.toArray().transpose())
np.testing.assert_array_almost_equal(result.toArray(), expected)
def test_multinomial_logistic_regression_with_bound(self):
data_path = "data/mllib/sample_multiclass_classification_data.txt"
df = self.spark.read.format("libsvm").load(data_path)
lor = LogisticRegression(
regParam=0.01,
lowerBoundsOnCoefficients=Matrices.dense(3, 4, range(12)),
upperBoundsOnIntercepts=Vectors.dense(0.0, 0.0, 0.0),
)
model = lor.fit(df)
expected = [
[4.593, 4.5516, 9.0099, 12.2904],
[1.0, 8.1093, 7.0, 10.0],
[3.041, 5.0, 8.0, 11.0],
]
for i in range(0, len(expected)):
self.assertTrue(
np.allclose(model.coefficientMatrix.toArray()[i],
expected[i],
atol=1e-4))
self.assertTrue(
np.allclose(model.interceptVector.toArray(),
[-0.9057, -1.1392, -0.0033],
atol=1e-4))
def test_matrix_array(self):
mat = Matrices.dense(2, 2, [1.0, 2.0, 3.0, 4.0])
arr = self.spark.createDataFrame([(mat,)], ["mat"]) \
.withColumn("arr", matrixToArray("mat")) \
.select("arr").head().arr
np.testing.assert_array_almost_equal(
np.array(arr.values),
mat.toArray().reshape(4, order="F"))
def test_matrix_vector_multiply(self):
mat = Matrices.dense(2, 2, [1.0, 2.0, 3.0, 4.0])
vec = Vectors.dense(1.0, 2.0)
result = self.spark.createDataFrame([(mat, vec)], ["mat", "vec"]) \
.withColumn("result", multiplyMatrixVector("mat", "vec")) \
.select("result").head().result
np.testing.assert_array_almost_equal(
result.toArray(), np.dot(mat.toArray(), vec.toArray()))
def do_spark():
#sparky = SparkSession.builder.getOrCreate()
#df = sparky.sql('''select 'spark' as hello ''')
v = Vectors.dense(2.0, 0.0, 3.0, 4.0, 5.0)
#dm1 = Matrices.dense(2, 3, [1, 2, 3, 4, 5, 6])
#dm2 = Matrices.dense(2, 3, [7, 8, 9, 10, 11, 12])
#mat = sc.io.mmread('494_bus.mtx')
#Something bigger with matrices
n = 50
s1 = np.random.normal(0, 1, n * n)
s2 = np.random.normal(0, 1, n * n)
dm3 = Matrices.dense(n, n, s1)
dm4 = Matrices.dense(n, n, s2)
dm3s = dm3.toSparse()
dm4s = dm4.toSparse()
return "{}".format(v.norm(2))
def test_binomial_logistic_regression_with_bound(self):
df = self.spark.createDataFrame(
[(1.0, 1.0, Vectors.dense(0.0, 5.0)),
(0.0, 2.0, Vectors.dense(1.0, 2.0)),
(1.0, 3.0, Vectors.dense(2.0, 1.0)),
(0.0, 4.0, Vectors.dense(3.0, 3.0)), ], ["label", "weight", "features"])
lor = LogisticRegression(regParam=0.01, weightCol="weight",
lowerBoundsOnCoefficients=Matrices.dense(1, 2, [-1.0, -1.0]),
upperBoundsOnIntercepts=Vectors.dense(0.0))
model = lor.fit(df)
self.assertTrue(
np.allclose(model.coefficients.toArray(), [-0.2944, -0.0484], atol=1E-4))
self.assertTrue(np.isclose(model.intercept, 0.0, atol=1E-4))
def test_persistance(self):
filter = LinearKalmanFilter() \
.setStateSize(2) \
.setInitialStateMean(Vectors.dense([0.0, 0.0])) \
.setInitialStateCovariance(Matrices.dense(2, 2, [1.0, 0.0, 0.0, 0.0]))\
path = tempfile.mkdtemp()
model_path = os.path.join(path, "lkf")
filter.save(model_path)
loaded = LinearKalmanFilter.load(model_path)
assert (loaded.getInitialStateMean() == filter.getInitialStateMean())
assert (loaded.getInitialStateCovariance() ==
filter.getInitialStateCovariance())
assert (loaded.getStateSize() == filter.getStateSize())
def test_batch_save_and_resume(self):
n = 100
ts = np.arange(0, n)
zs = np.random.normal(0, 1, n) + ts
split_point = n // 2
initial = zs[:split_point]
remaining = zs[split_point:]
filter = LinearKalmanFilter()\
.setMeasurementCol("measurement")\
.setInitialStateMean(
Vectors.dense([0.0, 0.0]))\
.setInitialStateCovariance(
Matrices.dense(2, 2, [1, 0, 0, 1]))\
.setProcessModel(
Matrices.dense(2, 2, [1, 0, 1, 1]))\
.setProcessNoise(
Matrices.dense(2, 2, [0.01, 0.0, 0.0, 0.01]))\
.setMeasurementNoise(
Matrices.dense(1, 1, [1]))\
.setMeasurementModel(
Matrices.dense(1, 2, [1, 0]))
initial_filter = filter.setInitialStateCovariance(
Matrices.dense(2, 2, [1000.0, 0.0, 0.0, 1000.0]))
def create_df(m):
return self.spark.createDataFrame([(Vectors.dense(m[i]), )
for i in range(len(m))],
["measurement"])
initial_measurements = create_df(initial)
complete_measurements = create_df(zs)
initial_state = initial_filter.transform(initial_measurements)\
.filter("stateIndex == {}".format(len(initial)))\
.select("stateKey", "state")
complete_state = initial_filter.transform(complete_measurements) \
.filter("stateIndex == {}".format(len(zs)))\
.select("stateKey", "state")
restarted_filter = filter\
.setInitialStateDistributionCol("state")
remaining_measurements = create_df(remaining)\
.crossJoin(initial_state)
restarted_state = restarted_filter.transform(remaining_measurements)\
.filter("stateIndex == {}".format(n - split_point))\
.select("stateKey", "state")
assert (restarted_state.collect() == complete_state.collect())
def test_multinomial_logistic_regression_with_bound(self):
data_path = "data/mllib/sample_multiclass_classification_data.txt"
df = self.spark.read.format("libsvm").load(data_path)
lor = LogisticRegression(regParam=0.01,
lowerBoundsOnCoefficients=Matrices.dense(3, 4, range(12)),
upperBoundsOnIntercepts=Vectors.dense(0.0, 0.0, 0.0))
model = lor.fit(df)
expected = [[4.593, 4.5516, 9.0099, 12.2904],
[1.0, 8.1093, 7.0, 10.0],
[3.041, 5.0, 8.0, 11.0]]
for i in range(0, len(expected)):
self.assertTrue(
np.allclose(model.coefficientMatrix.toArray()[i], expected[i], atol=1E-4))
self.assertTrue(
np.allclose(model.interceptVector.toArray(), [-0.9057, -1.1392, -0.0033], atol=1E-4))
mps = int(sys.argv[2])
spark = SparkSession.builder.appName("RateSourceLKF").getOrCreate()
spark.sparkContext.setLogLevel("WARN")
noise_param = 1
input_df = spark.readStream.format("rate").option("rowsPerSecond", mps).load()\
.withColumn("mod", F.col("value") % num_states)\
.withColumn("stateKey", F.col("mod").cast("String"))\
.withColumn("trend", (F.col("value")/num_states).cast("Integer") + F.randn() * noise_param)
lkf = LinearKalmanFilter(2, 1)\
.setStateKeyCol("stateKey")\
.setMeasurementCol("measurement")\
.setInitialCovariance(Matrices.dense(2, 2, [10000.0, 0.0, 0.0, 10000.0]))\
.setProcessModel(Matrices.dense(2, 2, [1.0, 0.0, 1.0, 1.0]))\
.setProcessNoise(Matrices.dense(2, 2, [0.0001, 0.0, 0.0, 0.0001]))\
.setMeasurementNoise(Matrices.dense(1, 1, [noise_param]))\
.setMeasurementModel(Matrices.dense(1, 2, [1.0, 0.0]))
assembler = VectorAssembler(inputCols=["trend"], outputCol="measurement")
measurements = assembler.transform(input_df)
query = lkf.transform(measurements)\
.writeStream\
.queryName("RateSourceLKF")\
.outputMode("append")\
.format("console")\
.start()
num_states = int(sys.argv[1])
mps = int(sys.argv[2])
spark = SparkSession.builder.appName("LKFRateSourceOLS").getOrCreate()
spark.sparkContext.setLogLevel("WARN")
# OLS problem, states to be estimated are a, b and c
# z = a*x + b * y + c + w, where w ~ N(0, 1)
a = 0.5
b = 0.2
c = 1.2
noise_param = 1
label_udf = F.udf(lambda x, y, w: Vectors.dense([x * a + y * b + c + w]),
VectorUDT())
features_udf = F.udf(lambda x, y: Matrices.dense(1, 3, [x, y, 1]),
MatrixUDT())
features = spark.readStream.format("rate").option("rowsPerSecond", mps).load()\
.withColumn("mod", F.col("value") % num_states)\
.withColumn("stateKey", F.col("mod").cast("String"))\
.withColumn("x", (F.col("value")/num_states).cast("Integer").cast("Double"))\
.withColumn("y", F.sqrt("x"))\
.withColumn("w", F.randn(0) * noise_param)\
.withColumn("label", label_udf("x", "y", "w"))\
.withColumn("features", features_udf("x", "y"))
lkf = LinearKalmanFilter()\
.setStateKeyCol("stateKey")\
.setMeasurementCol("label")\
.setMeasurementModelCol("features") \
