def test_glm_multinomial_coeffs():
trainF = h2o.import_file(pyunit_utils.locate("smalldata/iris/iris_train.csv"))
y = "species"
x = [0,1,2,3]
bin_LS = glm(family='multinomial', seed=12345)
bin_LS.train(x=x, y=y, training_frame=trainF)
print(bin_LS.summary())
coefficient_table_original = bin_LS._model_json["output"]["coefficients_table"]
coefficient_table = bin_LS._model_json["output"]["coefficients_table_multinomials_with_class_names"]
coeffNamesOld = coefficient_table_original.col_header
coeffNames = coefficient_table.col_header
validCoefficientNames = [u"names", u"coefs_class_Iris-setosa", u"coefs_class_Iris-versicolor",
u"coefs_class_Iris-virginica", u"std_coefs_class_Iris-setosa",
u"std_coefs_class_Iris-versicolor", u"std_coefs_class_Iris-virginica"]
oldCoefficientNames = [u"names", u"coefs_class_0", u"coefs_class_1",
u"coefs_class_2", u"std_coefs_class_0",
u"std_coefs_class_1", u"std_coefs_class_2"]
print(coefficient_table)
print(coefficient_table_original)
# compare coefficient names
assert len(set(coeffNames).intersection(validCoefficientNames))==len(coeffNames),\
"Expected coefficient names: {0}. Actual coefficient names: {1}".format(validCoefficientNames, coeffNames)
assert len(set(coeffNamesOld).intersection(oldCoefficientNames))==len(coeffNames), \
"Expected original coefficient names: {0}. Actual original coefficient names: " \
"{1}".format(oldCoefficientNames, coeffNamesOld)
# compare table contents to make sure they contain the same values
pyunit_utils.assert_H2OTwoDimTable_equal_upto(coefficient_table_original, coefficient_table, [u'coefs_class_0'],
tolerance=1e-10)
def testFrameTransform():
train = h2o.import_file(path=pyunit_utils.locate("smalldata/prostate/prostate_complete.csv.zip"))
y = 'CAPSULE'
x = ['AGE','VOL','DCAPS']
train[10,2] = None
train[20,7] = None
train[y] = train[y].asfactor()
# build model choosing skip
model1 = H2OANOVAGLMEstimator(family='binomial', lambda_=0, missing_values_handling="skip")
model1.train(x=x, y=y, training_frame=train)
# build model deleting the two rows with missing values
train.drop([10, 20], axis=0)
model2 = H2OANOVAGLMEstimator(family='binomial', lambda_=0, missing_values_handling="skip")
model2.train(x=x, y=y, training_frame=train)
# the two models should be the same, compare the model summaries
summary1 = model1._model_json['output']['model_summary']
summary2 = model2._model_json['output']['model_summary']
pyunit_utils.assert_H2OTwoDimTable_equal_upto(summary1, summary2, summary1.col_header)
def partial_plot_test():
# Import data set that contains NAs
data = h2o.import_file(pyunit_utils.locate('smalldata/prostate/prostate_cat_NA.csv'))
x = data.names
y = 'CAPSULE'
x.remove(y)
weights = h2o.H2OFrame([3.0]*data.nrow)
tweight2 = [1.0]*data.nrow
random.seed(12345)
for ind in range(len(tweight2)):
tweight2[ind] = random.randint(0,5)
weights2 = h2o.H2OFrame(tweight2)
data = data.cbind(weights)
data = data.cbind(weights2)
data.set_name(data.ncol-2, "constWeight")
data.set_name(data.ncol-1, "variWeight")
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=50, learn_rate=0.05, seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
# pdp without weight or NA
pdpOrig = gbm_model.partial_plot(data=data,cols=['AGE', 'RACE'],server=True, plot=True)
# pdp with constant weight and NA
pdpcWNA = gbm_model.partial_plot(data=data, cols=['AGE', 'RACE'], server=True, plot=True,
weight_column="constWeight", include_na=True)
# compare results
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpOrig[0], pdpcWNA[0], pdpOrig[0].col_header, tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpOrig[1], pdpcWNA[1], pdpOrig[1].col_header, tolerance=1e-10)
# pdp with changing weight NA
pdpvWNA = gbm_model.partial_plot(data=data, cols=['AGE', 'RACE'], server=True, plot=True,
weight_column="variWeight", include_na=True)
ageList = pyunit_utils.extract_col_value_H2OTwoDimTable(pdpvWNA[0], "age")
raceList = pyunit_utils.extract_col_value_H2OTwoDimTable(pdpvWNA[1], "race")
raceList.remove(raceList[2])
raceList.append(data[21,"RACE"]) # replace with NA word
ageList[len(ageList)-1] = float('nan') # replace nan with proper form for python
compare_weightedStats(gbm_model, 'smalldata/prostate/prostate_cat_NA.csv', raceList, "RACE", tweight2, pdpvWNA[1], tol=1e-10)
compare_weightedStats(gbm_model, 'smalldata/prostate/prostate_cat_NA.csv', ageList, "AGE", tweight2, pdpvWNA[0], tol=1e-10)
def partial_plot_test_with_user_splits():
data = h2o.import_file(pyunit_utils.locate('smalldata/prostate/prostate_cat_NA.csv'))
x = data.names
y = 'CAPSULE'
x.remove(y)
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=50, learn_rate=0.05, seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
user_splits = dict()
user_splits['AGE'] = [43.0, 44.89473684210526, 46.78947368421053, 48.68421052631579, 50.578947368421055,
52.473684210526315, 54.368421052631575, 56.26315789473684, 58.1578947368421,
60.05263157894737, 61.94736842105263, 63.84210526315789, 65.73684210526315,
67.63157894736842, 69.52631578947368, 71.42105263157895, 73.3157894736842,
75.21052631578948, 77.10526315789474]
user_splits['RACE'] = ["Black"]
# pdp without weight or NA
file, filename = tempfile.mkstemp(suffix=".png")
pdpOrig = gbm_model.partial_plot(data=data,cols=['AGE', 'RACE', 'DCAPS'],server=True, plot=True, save_to_file=filename)
assert os.path.getsize(filename) > 0
os.unlink(filename)
if os.path.isfile(filename):
os.remove(filename)
pdpUserSplit = gbm_model.partial_plot(data=data,cols=['AGE', 'RACE', 'DCAPS'],server=True, plot=True,
user_splits=user_splits)
# compare results
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpUserSplit[0], pdpOrig[0], pdpUserSplit[0].col_header, tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpUserSplit[1], pdpOrig[1], pdpUserSplit[1].col_header, tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpOrig[2], pdpUserSplit[2], pdpUserSplit[2].col_header, tolerance=1e-10)
def partial_plot_test_with_user_splits():
data = h2o.import_file(pyunit_utils.locate('smalldata/prostate/prostate_cat_NA.csv'))
x = data.names
y = 'CAPSULE'
x.remove(y)
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=50, learn_rate=0.05, seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
user_splits = dict()
user_splits['AGE'] = [43.0, 44.89473684210526, 46.78947368421053, 48.68421052631579, 50.578947368421055,
52.473684210526315, 54.368421052631575, 56.26315789473684, 58.1578947368421,
60.05263157894737, 61.94736842105263, 63.84210526315789, 65.73684210526315,
67.63157894736842, 69.52631578947368, 71.42105263157895, 73.3157894736842,
75.21052631578948, 77.10526315789474]
user_splits['RACE'] = ["Black"]
# pdp without weight or NA
file, filename = tempfile.mkstemp(suffix=".png")
pdpOrig = gbm_model.partial_plot(data=data,cols=['AGE', 'RACE', 'DCAPS'],server=True, plot=True, save_to_file=filename)
assert os.path.getsize(filename) > 0
os.unlink(filename)
pdpUserSplit = gbm_model.partial_plot(data=data,cols=['AGE', 'RACE', 'DCAPS'],server=True, plot=True,
user_splits=user_splits)
# compare results
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpUserSplit[0], pdpOrig[0], pdpUserSplit[0].col_header, tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpUserSplit[1], pdpOrig[1], pdpUserSplit[1].col_header, tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpOrig[2], pdpUserSplit[2], pdpUserSplit[2].col_header, tolerance=1e-10)
def partial_plot_test():
# Import data set that contains NAs
data = h2o.import_file(pyunit_utils.locate("smalldata/airlines/AirlinesTrainWgt.csv"), na_strings=["NA"])
test = h2o.import_file(pyunit_utils.locate("smalldata/airlines/AirlinesTrainWgt.csv"), na_strings=["NA"])
x = data.names
y = "IsDepDelayed"
data[y] = data[y]
x.remove(y)
x.remove("Weight")
x.remove("IsDepDelayed_REC")
WC = "Weight"
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=80, learn_rate=0.1, seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
# pdp with weight and no NA
pdpw = gbm_model.partial_plot(data=test, cols=["Input_miss", "Distance"], server=True, plot=False,
weight_column=WC)
# pdp with weight and NA
pdpwNA = gbm_model.partial_plot(data=test, cols=["Input_miss", "Distance"], server=True, plot=False,
weight_column=WC, include_na = True)
input_miss_list = pyunit_utils.extract_col_value_H2OTwoDimTable(pdpwNA[0], "input_miss")
assert math.isnan(input_miss_list[-1]), "Expected last element to be nan but is not."
distance_list = pyunit_utils.extract_col_value_H2OTwoDimTable(pdpwNA[1], "distance")
assert math.isnan(distance_list[-1]), "Expected last element to be nan but is not."
# compare pdpw with pdpwNA, they should equal upto NA since the pdpw does not have NAs.
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpw[0], pdpwNA[0], pdpw[0].col_header, tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpw[1], pdpwNA[1], pdpw[1].col_header, tolerance=1e-10)
# compare pdpwNA with theoretical results
pyunit_utils.compare_weightedStats(gbm_model, test, input_miss_list, "Input_miss",
test[WC].as_data_frame(use_pandas=False, header=False), pdpwNA[0], tol=1e-10)
pyunit_utils.compare_weightedStats(gbm_model, test, distance_list, "Distance",
test[WC].as_data_frame(use_pandas=False, header=False), pdpwNA[1], tol=1e-10)
def partial_plot_test_with_user_splits():
data = h2o.import_file(
pyunit_utils.locate('smalldata/prostate/prostate_cat_NA.csv'))
x = data.names
y = 'CAPSULE'
x.remove(y)
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=50,
learn_rate=0.05,
seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
file, filename = tempfile.mkstemp(suffix=".png")
user_splits = dict()
user_splits['AGE'] = [
43.0, 44.89473684210526, 46.78947368421053, 48.68421052631579,
50.578947368421055, 52.473684210526315, 54.368421052631575,
56.26315789473684, 58.1578947368421, 60.05263157894737,
61.94736842105263, 63.84210526315789, 65.73684210526315,
67.63157894736842, 69.52631578947368, 71.42105263157895,
73.3157894736842, 75.21052631578948, 77.10526315789474
]
user_splits['RACE'] = ["Black", "White"]
pdpUserSplit2D = gbm_model.partial_plot(data=data,
server=True,
plot=True,
user_splits=user_splits,
col_pairs_2dpdp=[['AGE', 'PSA'],
['AGE', 'RACE']],
save_to_file=filename)
pdpUserSplit1D2D = gbm_model.partial_plot(data=data,
cols=['AGE', 'RACE', 'DCAPS'],
server=True,
plot=True,
user_splits=user_splits,
col_pairs_2dpdp=[['AGE', 'PSA'],
['AGE',
'RACE']],
save_to_file=filename)
pdpUserSplit1D = gbm_model.partial_plot(data=data,
cols=['AGE', 'RACE', 'DCAPS'],
server=True,
plot=True,
user_splits=user_splits,
save_to_file=filename)
# compare results 1D pdp
for i in range(3):
pyunit_utils.assert_H2OTwoDimTable_equal_upto(
pdpUserSplit1D[i],
pdpUserSplit1D2D[i],
pdpUserSplit1D[i].col_header,
tolerance=1e-10)
# compare results 2D pdp
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpUserSplit2D[0],
pdpUserSplit1D2D[3],
pdpUserSplit2D[0].col_header,
tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpUserSplit2D[1],
pdpUserSplit1D2D[4],
pdpUserSplit2D[1].col_header,
tolerance=1e-10)
def partial_plot_test():
# Import data set that contains NAs
data = h2o.import_file(
pyunit_utils.locate("smalldata/airlines/AirlinesTrainWgt.csv"),
na_strings=["NA"])
test = h2o.import_file(
pyunit_utils.locate("smalldata/airlines/AirlinesTrainWgt.csv"),
na_strings=["NA"])
x = data.names
y = "IsDepDelayed"
data[y] = data[y]
x.remove(y)
x.remove("Weight")
x.remove("IsDepDelayed_REC")
WC = "Weight"
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=80,
learn_rate=0.1,
seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
# pdp with weight and no NA
pdpw = gbm_model.partial_plot(data=test,
cols=["Input_miss", "Distance"],
server=True,
plot=False,
weight_column=WC)
# pdp with weight and NA
pdpwNA = gbm_model.partial_plot(data=test,
cols=["Input_miss", "Distance"],
server=True,
plot=False,
weight_column=WC,
include_na=True)
input_miss_list = pyunit_utils.extract_col_value_H2OTwoDimTable(
pdpwNA[0], "input_miss")
assert math.isnan(
input_miss_list[-1]), "Expected last element to be nan but is not."
distance_list = pyunit_utils.extract_col_value_H2OTwoDimTable(
pdpwNA[1], "distance")
assert math.isnan(
distance_list[-1]), "Expected last element to be nan but is not."
# compare pdpw with pdpwNA, they should equal upto NA since the pdpw does not have NAs.
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpw[0],
pdpwNA[0],
pdpw[0].col_header,
tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpw[1],
pdpwNA[1],
pdpw[1].col_header,
tolerance=1e-10)
# compare pdpwNA with theoretical results
pyunit_utils.compare_weightedStats(gbm_model,
test,
input_miss_list,
"Input_miss",
test[WC].as_data_frame(use_pandas=False,
header=False),
pdpwNA[0],
tol=1e-10)
pyunit_utils.compare_weightedStats(gbm_model,
test,
distance_list,
"Distance",
test[WC].as_data_frame(use_pandas=False,
header=False),
pdpwNA[1],
tol=1e-10)
def partial_plot_test():
# Import data set that contains NAs
data = h2o.import_file(
pyunit_utils.locate('smalldata/prostate/prostate_cat_NA.csv'))
x = data.names
y = 'CAPSULE'
x.remove(y)
weights = h2o.H2OFrame([3.0] * data.nrow)
tweight2 = [1.0] * data.nrow
random.seed(12345)
for ind in range(len(tweight2)):
tweight2[ind] = random.randint(0, 5)
weights2 = h2o.H2OFrame(tweight2)
data = data.cbind(weights)
data = data.cbind(weights2)
data.set_name(data.ncol - 2, "constWeight")
data.set_name(data.ncol - 1, "variWeight")
# Build a GBM model predicting for response CAPSULE
gbm_model = H2OGradientBoostingEstimator(ntrees=50,
learn_rate=0.05,
seed=12345)
gbm_model.train(x=x, y=y, training_frame=data)
# pdp without weight or NA
pdpOrig = gbm_model.partial_plot(data=data,
cols=['AGE', 'RACE'],
server=True,
plot=True)
# pdp with constant weight and NA
pdpcWNA = gbm_model.partial_plot(data=data,
cols=['AGE', 'RACE'],
server=True,
plot=True,
weight_column="constWeight",
include_na=True)
# compare results
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpOrig[0],
pdpcWNA[0],
pdpOrig[0].col_header,
tolerance=1e-10)
pyunit_utils.assert_H2OTwoDimTable_equal_upto(pdpOrig[1],
pdpcWNA[1],
pdpOrig[1].col_header,
tolerance=1e-10)
# pdp with changing weight NA
pdpvWNA = gbm_model.partial_plot(data=data,
cols=['AGE', 'RACE'],
server=True,
plot=True,
weight_column="variWeight",
include_na=True)
ageList = pyunit_utils.extract_col_value_H2OTwoDimTable(pdpvWNA[0], "age")
raceList = pyunit_utils.extract_col_value_H2OTwoDimTable(
pdpvWNA[1], "race")
raceList.remove(raceList[2])
raceList.append(data[21, "RACE"]) # replace with NA word
ageList[len(ageList) - 1] = float(
'nan') # replace nan with proper form for python
compare_weightedStats(gbm_model,
'smalldata/prostate/prostate_cat_NA.csv',
raceList,
"RACE",
tweight2,
pdpvWNA[1],
tol=1e-10)
compare_weightedStats(gbm_model,
'smalldata/prostate/prostate_cat_NA.csv',
ageList,
"AGE",
tweight2,
pdpvWNA[0],
tol=1e-10)