def pca_arrests():
print("Importing USArrests.csv data...")
arrestsH2O = h2o.upload_file(
pyunit_utils.locate("smalldata/pca_test/USArrests.csv"))
print(
"Testing to see whether the trained PCA are essentially the same using different implementation..."
)
eigenvector_standard = None
for impl in [
"MTJ_EVD_DENSEMATRIX", "MTJ_EVD_SYMMMATRIX", "MTJ_SVD_DENSEMATRIX",
"JAMA"
]:
print("Run PCA with implementation: " + impl)
model = H2OPCA(k=4, pca_impl=impl, seed=1234)
model.train(x=list(range(4)), training_frame=arrestsH2O)
eigenvectors = model._model_json["output"]["eigenvectors"]
if eigenvector_standard is not None:
# Compare to see if they are fundamentally the same
pyunit_utils.assert_H2OTwoDimTable_equal(
eigenvector_standard,
eigenvectors,
model._model_json["output"]["names"],
tolerance=1e-6,
check_sign=True,
check_all=False)
else:
eigenvector_standard = eigenvectors
def glrm_iris():
print("Importing iris.csv data...")
irisH2O = h2o.upload_file(pyunit_utils.locate("smalldata/iris/iris.csv"))
irisH2O.describe()
print("@@@@@@ Building PCA with GramSVD...\n")
glrmPCA = H2OPCA(k=5, transform="STANDARDIZE", pca_method="GLRM", use_all_factor_levels=True, seed=21)
glrmPCA.train(x=irisH2O.names, training_frame=irisH2O)
glrm_h2o = H2OGeneralizedLowRankEstimator(k=5, loss="Quadratic",transform="STANDARDIZE", recover_svd=True, seed=21)
glrm_h2o.train(x=irisH2O.names, training_frame=irisH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and GLRM...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(glrmPCA._model_json["output"]["importance"],
glrm_h2o._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1e-6)
print("@@@@@@ Comparing eigenvectors between GramSVD and GLRM...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(glrmPCA._model_json["output"]["eigenvectors"],
glrm_h2o._model_json["output"]["eigenvectors"],
glrm_h2o._model_json["output"]["names"], tolerance=1e-6,check_sign=True)
# check to make sure maximum proportional variance <= 1
assert glrmPCA._model_json["output"]["importance"].cell_values[1][1] <= 1, \
"Expected value <= 1.0 but received {0}".format(glrmPCA._model_json["output"]["importance"].cell_values[1][1])
def pca_pubdev_4167_OOM():
"""
This pyunit is written to make sure PCA works with customer data. It is mainly used by customer to verify
PCA operations and not to be used as a regular test since I do not want to expose customer data.
"""
h2o.remove_all()
transform_types = ["NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"] # make sure we check all tranforms
transformN = transform_types[randint(0, len(transform_types)-1)]
print("transform used on dataset is {0}.\n".format(transformN))
training_data = h2o.import_file(path=pyunit_utils.locate("/Users/wendycwong/gitBackup/SDatasets/pubdev_4167_Avkash/m120K.tar")) # Nidhi: import may not work
gramSVDPCA = H2OPCA(k=training_data.ncols, transform=transformN)
gramSVDPCA.train(x=list(range(0, training_data.ncols)), training_frame=training_data)
powerSVDPCA = H2OPCA(k=training_data.ncols, transform=transformN, pca_method="Power")
powerSVDPCA.train(x=list(range(0, training_data.ncols)), training_frame=training_data)
# compare singular values and stuff between power and GramSVD methods
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVDPCA._model_json["output"]["importance"],
powerSVDPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1e-5, check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVDPCA._model_json["output"]["eigenvectors"],
powerSVDPCA._model_json["output"]["eigenvectors"],
powerSVDPCA._model_json["output"]["names"], tolerance=1e-1,
check_sign=True)
def glrm_iris():
print("Importing iris.csv data...")
irisH2O = h2o.upload_file(pyunit_utils.locate("smalldata/iris/iris.csv"))
irisH2O.describe()
print("@@@@@@ Building PCA with GramSVD...\n")
glrmPCA = H2OPCA(k=5, transform="STANDARDIZE", pca_method="GLRM", use_all_factor_levels=True, seed=21)
glrmPCA.train(x=irisH2O.names, training_frame=irisH2O)
glrm_h2o = H2OGeneralizedLowRankEstimator(k=5, loss="Quadratic",transform="STANDARDIZE", recover_svd=True, seed=21)
glrm_h2o.train(x=irisH2O.names, training_frame=irisH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and GLRM...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(glrmPCA._model_json["output"]["importance"],
glrm_h2o._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1e-6)
print("@@@@@@ Comparing eigenvectors between GramSVD and GLRM...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(glrmPCA._model_json["output"]["eigenvectors"],
glrm_h2o._model_json["output"]["eigenvectors"],
glrm_h2o._model_json["output"]["names"], tolerance=1e-6,check_sign=True)
# check to make sure maximum proportional variance <= 1
assert glrmPCA._model_json["output"]["importance"].cell_values[1][1] <= 1, \
"Expected value <= 1.0 but received {0}".format(glrmPCA._model_json["output"]["importance"].cell_values[1][1])
def pca_wideDataset_rotterdam_glrm():
tol = 2e-5
h2o.remove_all()
print("Importing Rotterdam.csv data...")
rotterdamH2O = h2o.upload_file(
pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(rotterdamH2O.names) - y)
# special test with GLRM. Need use_all_levels to be true
print("------ Testing GLRM PCA --------")
gramSVD = H2OPCA(k=8,
impute_missing=True,
transform="DEMEAN",
seed=12345,
use_all_factor_levels=True)
gramSVD.train(x=x, training_frame=rotterdamH2O)
glrmPCA = H2OGeneralizedLowRankEstimator(k=8,
transform="DEMEAN",
seed=12345,
init="Random",
recover_svd=True,
regularization_x="None",
regularization_y="None",
max_iterations=11)
glrmPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print(
"@@@@@@ Comparing eigenvectors and eigenvalues between GramSVD and GLRM...\n"
)
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
glrmPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1,
check_all=False)
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["names"],
tolerance=tol,
check_sign=True,
check_all=False)
def pca_wideDataset_rotterdam():
h2o.remove_all()
print("Importing Rotterdam.csv data...")
rotterdamH2O = h2o.upload_file(
pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(rotterdamH2O.names) - y)
transform_types = ["NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"]
transformN = transform_types[randint(0, len(transform_types) - 1)]
print("transform used on dataset is {0}.\n".format(transformN))
gramSVD = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
seed=12345)
gramSVD.train(x=x, training_frame=rotterdamH2O)
powerPCA = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
pca_method="Power",
seed=12345) # power
powerPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
powerPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1e-6,
check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["names"],
tolerance=1e-6,
check_sign=True,
check_all=False)
def pca_pubdev_4167_OOM():
"""
This pyunit is written to make sure PCA works with customer data. It is mainly used by customer to verify
PCA operations and not to be used as a regular test since I do not want to expose customer data.
"""
h2o.remove_all()
transform_types = [
"NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"
] # make sure we check all tranforms
transformN = transform_types[randint(0, len(transform_types) - 1)]
print("transform used on dataset is {0}.\n".format(transformN))
training_data = h2o.import_file(path=pyunit_utils.locate(
"/Users/wendycwong/gitBackup/SDatasets/pubdev_4167_Avkash/m120K.tar")
) # Nidhi: import may not work
gramSVDPCA = H2OPCA(k=training_data.ncols, transform=transformN)
gramSVDPCA.train(x=list(range(0, training_data.ncols)),
training_frame=training_data)
powerSVDPCA = H2OPCA(k=training_data.ncols,
transform=transformN,
pca_method="Power")
powerSVDPCA.train(x=list(range(0, training_data.ncols)),
training_frame=training_data)
# compare singular values and stuff between power and GramSVD methods
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVDPCA._model_json["output"]["importance"],
powerSVDPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1e-5,
check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVDPCA._model_json["output"]["eigenvectors"],
powerSVDPCA._model_json["output"]["eigenvectors"],
powerSVDPCA._model_json["output"]["names"],
tolerance=1e-1,
check_sign=True)
def pca_max_k():
data = h2o.upload_file(pyunit_utils.locate("smalldata/pca_test/SDSS_quasar.txt.zip"))
x = list(set(data.names))
pcaGramSVD = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="GramSVD", impute_missing=True, max_iterations=100)
pcaGramSVD.train(x, training_frame=data)
pcaPower = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Power", impute_missing=True, max_iterations=100,
seed=12345)
pcaPower.train(x, training_frame=data)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(pcaGramSVD._model_json["output"]["importance"],
pcaPower._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1)
correctEigNum = pcaPower.full_parameters["k"]["actual_value"]
gramSVDNum = len(pcaGramSVD._model_json["output"]["importance"].cell_values[0]) - 1
powerNum = len(pcaPower._model_json["output"]["importance"].cell_values[0]) - 1
assert correctEigNum == gramSVDNum, "PCA GramSVD FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + gramSVDNum + "."
assert correctEigNum == powerNum, "PCA Power FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + powerNum + "."
pcaRandomized = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Randomized",
impute_missing=True, max_iterations=100, seed=12345)
pcaRandomized.train(x, training_frame=data)
# eigenvalues between the PCA and Randomize should be close, I hope...
print("@@@@@@ Comparing eigenvalues between Randomized and Power PCA...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(pcaRandomized._model_json["output"]["importance"],
pcaPower._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"])
pcaGLRM = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="GLRM", use_all_factor_levels=True,
max_iterations=100, seed=12345)
pcaGLRM.train(x, training_frame=data)
correctEigNum = pcaGLRM.full_parameters["k"]["actual_value"]
glrmNum = len(pcaGLRM._model_json["output"]["importance"].cell_values[0]) - 1
assert correctEigNum == glrmNum, "PCA GLRM FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + glrmNum + "."
def pca_wideDataset_rotterdam_pcapower():
tol = 2e-5
h2o.remove_all()
print("Importing Rotterdam.csv data...")
rotterdamH2O = h2o.upload_file(
pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(rotterdamH2O.names) - y)
print("------ Testing Power PCA --------")
gramSVD = H2OPCA(k=8,
impute_missing=True,
transform="STANDARDIZE",
seed=12345)
gramSVD.train(x=x, training_frame=rotterdamH2O)
powerPCA = H2OPCA(k=8,
impute_missing=True,
transform="STANDARDIZE",
pca_method="Power",
seed=12345) # power
powerPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
powerPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1e-6,
check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["names"],
tolerance=tol,
check_sign=True,
check_all=False)
def pca_arrests():
print("Importing USArrests.csv data...")
arrestsH2O = h2o.upload_file(pyunit_utils.locate("smalldata/pca_test/USArrests.csv"))
print("Testing to see whether the trained PCA are essentially the same using different implementation...")
eigenvector_standard = None
for impl in ["MTJ_EVD_DENSEMATRIX", "MTJ_EVD_SYMMMATRIX", "MTJ_SVD_DENSEMATRIX", "JAMA"]:
print("Run PCA with implementation: " + impl)
model = H2OPCA(k = 4, pca_impl=impl, seed=1234)
model.train(x=list(range(4)), training_frame=arrestsH2O)
eigenvectors = model._model_json["output"]["eigenvectors"]
if eigenvector_standard is not None:
# Compare to see if they are fundamentally the same
pyunit_utils.assert_H2OTwoDimTable_equal(
eigenvector_standard,
eigenvectors,
model._model_json["output"]["names"],
tolerance=1e-6,
check_sign=True,
check_all=False)
else:
eigenvector_standard = eigenvectors
def pca_wideDataset_rotterdam():
h2o.remove_all()
print("Importing Rotterdam.csv data...")
rotterdamH2O = h2o.upload_file(pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(rotterdamH2O.names)-y)
transform_types = ["NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"]
transformN = transform_types[randint(0, len(transform_types)-1)]
print("transform used on dataset is {0}.\n".format(transformN))
buildModel = [False, False, False]
buildModel[randint(0, len(buildModel)-1)] = True
expNum = 0
if (buildModel[expNum]):
# special test with GLRM. Need use_all_levels to be true
print("------ Testing GLRM PCA --------")
gramSVD = H2OPCA(k=8, impute_missing=True, transform=transformN, seed=12345, use_all_factor_levels=True)
gramSVD.train(x=x, training_frame=rotterdamH2O)
glrmPCA = H2OGeneralizedLowRankEstimator(k=8, transform=transformN, seed=12345, init="Random",
max_iterations=10, recover_svd=True, regularization_x="None",
regularization_y="None")
glrmPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvectors between GramSVD and GLRM...\n")
print("@@@@@@ Comparing eigenvalues between GramSVD and GLRM...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["importance"],
glrmPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1, check_all=False)
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["names"], tolerance=1e-6,
check_sign=True, check_all=False)
h2o.remove(gramSVD)
h2o.remove(glrmPCA)
expNum=expNum+1
if (buildModel[expNum]):
print("------ Testing Power PCA --------")
gramSVD = H2OPCA(k=8, impute_missing=True, transform=transformN, seed=12345)
gramSVD.train(x=x, training_frame=rotterdamH2O)
powerPCA = H2OPCA(k=8, impute_missing=True, transform=transformN, pca_method="Power", seed=12345) # power
powerPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["importance"],
powerPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1e-6, check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["names"], tolerance=1e-6, check_sign=True,
check_all=False)
expNum=expNum+1
if (buildModel[expNum]):
print("------ Testing Randomized PCA --------")
gramSVD = H2OPCA(k=8, impute_missing=True, transform=transformN, seed=12345)
gramSVD.train(x=x, training_frame=rotterdamH2O)
randomizedPCA = H2OPCA(k=8, impute_missing=True, transform=transformN, pca_method="Randomized", seed=12345,
max_iterations=5) # power
randomizedPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Randomized...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["importance"],
randomizedPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1e-1, check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["names"], tolerance=1e-6,
check_sign=True, check_all=False)
h2o.remove_all()
def pca_scoring_history_importance():
"""
This test aims to check and make sure PCA returns the scoring history and importance which are
reported missing for certain PCA mode. Apart from changing the PCA mode, I throw in the transform
type to test as well randomly.
"""
transform_types = ["NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"]
transformN = transform_types[randint(0, len(transform_types) - 1)]
print("Importing australia.csv data...\n")
australia = h2o.upload_file(
pyunit_utils.locate("smalldata/extdata/australia.csv"))
col_indices = list(range(0, australia.ncol))
print("transform is {0}.\n".format(transformN))
# checking out PCA with GramSVD
print("@@@@@@ Building PCA with GramSVD...\n")
gramSVD = H2OPCA(k=3, transform=transformN)
gramSVD.train(x=col_indices, training_frame=australia)
# check PCA with PCA set to Randomized
print("@@@@@@ Building PCA with Randomized...\n")
randomizedPCA = H2OPCA(k=3,
transform=transformN,
pca_method="Randomized",
compute_metrics=True,
use_all_factor_levels=True)
randomizedPCA.train(x=col_indices, training_frame=australia)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Randomized...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
randomizedPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1e-3)
print("@@@@@@ Comparing eigenvectors between GramSVD and Randomized...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["names"],
tolerance=5e-2,
check_sign=True)
# check PCA with PCA set to Power
print("@@@@@@ Building PCA with Power...\n")
powerPCA = H2OPCA(k=3,
transform=transformN,
pca_method="Power",
compute_metrics=True,
use_all_factor_levels=True)
powerPCA.train(x=col_indices, training_frame=australia)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
powerPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
])
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["names"],
tolerance=1e-5,
check_sign=True)
# check PCA with PCA set to GLRM
print("@@@@@@ Building PCA with GLRM...\n")
glrmPCA = H2OPCA(k=3,
transform=transformN,
pca_method="GLRM",
compute_metrics=True,
use_all_factor_levels=True)
glrmPCA.train(x=col_indices, training_frame=australia)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and GLRM...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
glrmPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=2e-2)
print("@@@@@@ Comparing eigenvectors between GramSVD and GLRM...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["names"],
tolerance=2e-1,
check_sign=True)
# make sure we find the scoring history and it is not empty for all the PCA modes
# just check and make sure the cell_values exceed 0
assert len(gramSVD._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to GramSVD is empty."
assert len(powerPCA._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to using is empty."
assert len(randomizedPCA._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to Randomized is " \
"empty."
assert len(glrmPCA._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to GLRM is empty."
def pca_wideDataset_rotterdam():
h2o.remove_all()
print("Importing Rotterdam.csv data...")
rotterdamH2O = h2o.upload_file(
pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(rotterdamH2O.names) - y)
transform_types = ["NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"]
transformN = transform_types[randint(0, len(transform_types) - 1)]
print("transform used on dataset is {0}.\n".format(transformN))
buildModel = [False, False, False]
buildModel[randint(0, len(buildModel) - 1)] = True
expNum = 0
if (buildModel[expNum]):
# special test with GLRM. Need use_all_levels to be true
print("------ Testing GLRM PCA --------")
gramSVD = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
seed=12345,
use_all_factor_levels=True)
gramSVD.train(x=x, training_frame=rotterdamH2O)
glrmPCA = H2OGeneralizedLowRankEstimator(k=8,
transform=transformN,
seed=12345,
init="Random",
max_iterations=10,
recover_svd=True,
regularization_x="None",
regularization_y="None")
glrmPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvectors between GramSVD and GLRM...\n")
print("@@@@@@ Comparing eigenvalues between GramSVD and GLRM...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
glrmPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1,
check_all=False)
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["names"],
tolerance=1e-6,
check_sign=True,
check_all=False)
h2o.remove(gramSVD)
h2o.remove(glrmPCA)
expNum = expNum + 1
if (buildModel[expNum]):
print("------ Testing Power PCA --------")
gramSVD = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
seed=12345)
gramSVD.train(x=x, training_frame=rotterdamH2O)
powerPCA = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
pca_method="Power",
seed=12345) # power
powerPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
powerPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1e-6,
check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["names"],
tolerance=1e-6,
check_sign=True,
check_all=False)
expNum = expNum + 1
if (buildModel[expNum]):
print("------ Testing Randomized PCA --------")
gramSVD = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
seed=12345)
gramSVD.train(x=x, training_frame=rotterdamH2O)
randomizedPCA = H2OPCA(k=8,
impute_missing=True,
transform=transformN,
pca_method="Randomized",
seed=12345,
max_iterations=5) # power
randomizedPCA.train(x=x, training_frame=rotterdamH2O)
# compare singular values and stuff with GramSVD
print(
"@@@@@@ Comparing eigenvalues between GramSVD and Randomized...\n"
)
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["importance"],
randomizedPCA._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1e-1,
check_all=False)
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(
gramSVD._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["names"],
tolerance=1e-6,
check_sign=True,
check_all=False)
h2o.remove_all()
def pca_max_k():
data = h2o.upload_file(
pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(data.names) - y)
pcaGramSVD = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="GramSVD",
impute_missing=True,
max_iterations=100)
pcaGramSVD.train(x, training_frame=data)
pcaPower = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="Power",
impute_missing=True,
max_iterations=100,
seed=12345)
pcaPower.train(x, training_frame=data)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
pcaGramSVD._model_json["output"]["importance"],
pcaPower._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
],
tolerance=1)
correctEigNum = pcaPower.full_parameters["k"]["actual_value"]
gramSVDNum = len(
pcaGramSVD._model_json["output"]["importance"].cell_values[0]) - 1
powerNum = len(
pcaPower._model_json["output"]["importance"].cell_values[0]) - 1
assert correctEigNum == gramSVDNum, "PCA GramSVD FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + gramSVDNum + "."
assert correctEigNum == powerNum, "PCA Power FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + powerNum + "."
# Randomized and GLRM does not have wide dataset implementation. Check with smaller datasets
data = h2o.upload_file(
pyunit_utils.locate("smalldata/prostate/prostate_cat.csv"))
x = list(set(data.names))
pcaRandomized = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="Randomized",
impute_missing=True,
max_iterations=100,
seed=12345)
pcaRandomized.train(x, training_frame=data)
# should still work with rank deficient dataset
pcaRandomizedF = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="Randomized",
use_all_factor_levels=True,
impute_missing=True,
max_iterations=100,
seed=12345)
pcaRandomizedF.train(x, training_frame=data)
pcaPower = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="Power",
impute_missing=True,
max_iterations=100,
seed=12345)
pcaPower.train(x, training_frame=data)
# should still work with rank deficient dataset
pcaPowerF = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="Power",
use_all_factor_levels=True,
impute_missing=True,
max_iterations=100,
seed=12345)
pcaPowerF.train(x, training_frame=data)
# eigenvalues between the PCA and Randomize should be close, I hope...
print(
"@@@@@@ Comparing eigenvalues between Randomized and Power PCA...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(
pcaRandomized._model_json["output"]["importance"],
pcaPower._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
])
# eigenvalues between the PCA and Randomize should be close with rank deficient dataset, I hope...
print(
"@@@@@@ Comparing eigenvalues between Randomized and Power PCA with rank deficient dataset...\n"
)
pyunit_utils.assert_H2OTwoDimTable_equal(
pcaRandomizedF._model_json["output"]["importance"],
pcaPowerF._model_json["output"]["importance"], [
"Standard deviation", "Cumulative Proportion",
"Cumulative Proportion"
])
pcaGLRM = H2OPCA(k=-1,
transform="STANDARDIZE",
pca_method="GLRM",
use_all_factor_levels=True,
max_iterations=100,
seed=12345)
pcaGLRM.train(x, training_frame=data)
correctEigNum = pcaGLRM.full_parameters["k"]["actual_value"]
glrmNum = len(
pcaGLRM._model_json["output"]["importance"].cell_values[0]) - 1
assert correctEigNum == glrmNum, "PCA GLRM FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + glrmNum + "."
def pca_max_k():
data = h2o.upload_file(pyunit_utils.locate("bigdata/laptop/jira/rotterdam.csv.zip"))
y = set(["relapse"])
x = list(set(data.names) - y)
buildModel = [False, False, False, False]
buildModel[randint(0, len(buildModel)-1)] = True
# test 1
if buildModel[0]:
pcaGramSVD = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="GramSVD", impute_missing=True, max_iterations=100)
pcaGramSVD.train(x, training_frame=data)
pcaPower = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Power", impute_missing=True, max_iterations=100,
seed=12345)
pcaPower.train(x, training_frame=data)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(pcaGramSVD._model_json["output"]["importance"],
pcaPower._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1)
correctEigNum = pcaPower.full_parameters["k"]["actual_value"]
gramSVDNum = len(pcaGramSVD._model_json["output"]["importance"].cell_values[0]) - 1
powerNum = len(pcaPower._model_json["output"]["importance"].cell_values[0]) - 1
assert correctEigNum == gramSVDNum, "PCA GramSVD FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + gramSVDNum + "."
assert correctEigNum == powerNum, "PCA Power FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + powerNum + "."
# Randomized and GLRM does not have wide dataset implementation. Check with smaller datasets
# test 2
data = h2o.upload_file(pyunit_utils.locate("smalldata/prostate/prostate_cat.csv"))
x = list(set(data.names))
if buildModel[1]:
pcaRandomized = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Randomized",
impute_missing=True, max_iterations=100, seed=12345)
pcaRandomized.train(x, training_frame=data)
pcaPower = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Power",
impute_missing=True, max_iterations=100, seed=12345)
pcaPower.train(x, training_frame=data)
# eigenvalues between the PCA and Randomize should be close, I hope...
print("@@@@@@ Comparing eigenvalues between Randomized and Power PCA...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(pcaRandomized._model_json["output"]["importance"],
pcaPower._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"])
# test 3
if buildModel[2]:
# should still work with rank deficient dataset
pcaRandomizedF = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Randomized", use_all_factor_levels=True,
impute_missing=True, max_iterations=100, seed=12345)
pcaRandomizedF.train(x, training_frame=data)
# should still work with rank deficient dataset
pcaPowerF = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="Power", use_all_factor_levels=True,
impute_missing=True, max_iterations=100, seed=12345)
pcaPowerF.train(x, training_frame=data)
# eigenvalues between the PCA and Randomize should be close with rank deficient dataset, I hope...
print("@@@@@@ Comparing eigenvalues between Randomized and Power PCA with rank deficient dataset...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(pcaRandomizedF._model_json["output"]["importance"],
pcaPowerF._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"])
# test 4
if buildModel[3]:
pcaGLRM = H2OPCA(k=-1, transform="STANDARDIZE", pca_method="GLRM", use_all_factor_levels=True,
max_iterations=100, seed=12345)
pcaGLRM.train(x, training_frame=data)
correctEigNum = pcaGLRM.full_parameters["k"]["actual_value"]
glrmNum = len(pcaGLRM._model_json["output"]["importance"].cell_values[0]) - 1
assert correctEigNum == glrmNum, "PCA GLRM FAIL: expected number of eigenvalues: " + correctEigNum + \
", actual: " + glrmNum + "."
def pca_scoring_history_importance():
"""
This test aims to check and make sure PCA returns the scoring history and importance which are
reported missing for certain PCA mode. Apart from changing the PCA mode, I throw in the transform
type to test as well randomly.
"""
transform_types = ["NONE", "STANDARDIZE", "NORMALIZE", "DEMEAN", "DESCALE"]
transformN = transform_types[randint(0, len(transform_types)-1)]
print("Importing australia.csv data...\n")
australia = h2o.upload_file(pyunit_utils.locate("smalldata/extdata/australia.csv"))
col_indices = list(range(0, australia.ncol))
print("transform is {0}.\n".format(transformN))
# checking out PCA with GramSVD
print("@@@@@@ Building PCA with GramSVD...\n")
gramSVD = H2OPCA(k=3, transform=transformN)
gramSVD.train(x=col_indices, training_frame=australia)
# check PCA with PCA set to Randomized
print("@@@@@@ Building PCA with Randomized...\n")
randomizedPCA = H2OPCA(k=3, transform=transformN, pca_method="Randomized", compute_metrics=True,
use_all_factor_levels=True)
randomizedPCA.train(x=col_indices, training_frame=australia)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Randomized...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["importance"],
randomizedPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=1e-3)
print("@@@@@@ Comparing eigenvectors between GramSVD and Randomized...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["eigenvectors"],
randomizedPCA._model_json["output"]["names"], tolerance=5e-2,
check_sign=True)
# check PCA with PCA set to Power
print("@@@@@@ Building PCA with Power...\n")
powerPCA = H2OPCA(k=3, transform=transformN, pca_method="Power", compute_metrics=True, use_all_factor_levels=True)
powerPCA.train(x=col_indices, training_frame=australia)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and Power...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["importance"],
powerPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"])
print("@@@@@@ Comparing eigenvectors between GramSVD and Power...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["eigenvectors"],
powerPCA._model_json["output"]["names"], tolerance=1e-5, check_sign=True)
# check PCA with PCA set to GLRM
print("@@@@@@ Building PCA with GLRM...\n")
glrmPCA = H2OPCA(k=3, transform=transformN, pca_method="GLRM", compute_metrics=True, use_all_factor_levels=True)
glrmPCA.train(x=col_indices, training_frame=australia)
# compare singular values and stuff with GramSVD
print("@@@@@@ Comparing eigenvalues between GramSVD and GLRM...\n")
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["importance"],
glrmPCA._model_json["output"]["importance"],
["Standard deviation", "Cumulative Proportion", "Cumulative Proportion"],
tolerance=2e-2)
print("@@@@@@ Comparing eigenvectors between GramSVD and GLRM...\n")
# compare singular vectors
pyunit_utils.assert_H2OTwoDimTable_equal(gramSVD._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["eigenvectors"],
glrmPCA._model_json["output"]["names"], tolerance=2e-1,check_sign=True)
# make sure we find the scoring history and it is not empty for all the PCA modes
# just check and make sure the cell_values exceed 0
assert len(gramSVD._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to GramSVD is empty."
assert len(powerPCA._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to using is empty."
assert len(randomizedPCA._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to Randomized is " \
"empty."
assert len(glrmPCA._model_json["output"]["scoring_history"].cell_values) > 0, "PCA Scoring history setting " \
"pca_method to GLRM is empty."