def nullify(self, argin):
"""replace user-specified missing value with NULL
<table> <value>
Example:
bayeslite> .nullify mytable NaN
bayeslite> .nullify mytable ''
"""
parser = utils.ArgumentParser(prog='.nullify')
parser.add_argument('table', type=str, help='Name of the table.')
parser.add_argument('value', type=str, help='Target string to nullify.')
try:
args = parser.parse_args(shlex.split(argin))
except utils.ArgparseError as e:
self.stdout.write('%s' % (e.message, ))
return
bdbcontrib.nullify(self._bdb, args.table, args.value)
def nullify(self, argin):
"""replace user-specified missing value with NULL
<table> <value>
Example:
bayeslite> .nullify mytable NaN
bayeslite> .nullify mytable ''
"""
parser = utils.ArgumentParser(prog='.nullify')
parser.add_argument('table', type=str,
help='Name of the table.')
parser.add_argument('value', type=str,
help='Target string to nullify.')
try:
args = parser.parse_args(shlex.split(argin))
except utils.ArgparseError as e:
self.stdout.write('%s' % (e.message,))
return
bdbcontrib.nullify(self._bdb, args.table, args.value)
def nullify(self, value):
"""Wraps bdbcontrib.nullify by passing bdb and name.
bdbcontrib_nullify_doc"""
bdbcontrib.nullify(self.bdb, self.name, value)
def test_composer_integration__ci_slow():
# But currently difficult to seperate these tests into smaller tests because
# of their sequential nature. We will still test all internal functions
# with different regimes of operation.
# SETUP
# -----
# Dataset.
bdb = bayeslite.bayesdb_open()
bayeslite.bayesdb_read_csv_file(bdb, 'satellites', PATH_SATELLITES_CSV,
header=True, create=True)
bdbcontrib.nullify(bdb, 'satellites', 'NaN')
# Composer.
composer = Composer(n_samples=5)
composer.register_foreign_predictor(
multiple_regression.MultipleRegression)
composer.register_foreign_predictor(keplers_law.KeplersLaw)
composer.register_foreign_predictor(random_forest.RandomForest)
# Use complex generator for interesting test cases.
bayeslite.bayesdb_register_metamodel(bdb, composer)
bdb.execute('''
CREATE GENERATOR t1 FOR satellites USING composer(
default (
Country_of_Operator CATEGORICAL, Operator_Owner CATEGORICAL,
Users CATEGORICAL, Purpose CATEGORICAL,
Class_of_orbit CATEGORICAL, Perigee_km NUMERICAL,
Apogee_km NUMERICAL, Eccentricity NUMERICAL,
Launch_Mass_kg NUMERICAL, Dry_Mass_kg NUMERICAL,
Power_watts NUMERICAL, Date_of_Launch NUMERICAL,
Contractor CATEGORICAL,
Country_of_Contractor CATEGORICAL, Launch_Site CATEGORICAL,
Launch_Vehicle CATEGORICAL,
Source_Used_for_Orbital_Data CATEGORICAL,
longitude_radians_of_geo NUMERICAL,
Inclination_radians NUMERICAL,
),
random_forest (
Type_of_Orbit CATEGORICAL
GIVEN Apogee_km, Perigee_km,
Eccentricity, Period_minutes, Launch_Mass_kg,
Power_watts, Anticipated_Lifetime, Class_of_orbit
),
keplers_law (
Period_minutes NUMERICAL
GIVEN Perigee_km, Apogee_km
),
multiple_regression (
Anticipated_Lifetime NUMERICAL
GIVEN Dry_Mass_kg, Power_watts, Launch_Mass_kg,
Contractor
),
DEPENDENT(Apogee_km, Perigee_km, Eccentricity),
DEPENDENT(Contractor, Country_of_Contractor),
INDEPENDENT(Country_of_Operator, Date_of_Launch)
);''')
# ----------------------
# TEST INITIALIZE MODELS
# ----------------------
bdb.execute('INITIALIZE 2 MODELS FOR t1')
# Check number of models.
df = bdbcontrib.describe_generator_models(bdb, 't1')
assert len(df) == 2
df = bdbcontrib.describe_generator_models(bdb, 't1_cc')
assert len(df) == 2
# -------------------
# TEST ANALYZE MODELS
# -------------------
bdb.execute('ANALYZE t1 FOR 2 ITERATIONS WAIT;')
# Check number of iterations of composer.
df = bdbcontrib.describe_generator_models(bdb, 't1')
for index, modelno, iterations in df.itertuples():
assert iterations == 2
# Check number of iterations of composer_cc.
df = bdbcontrib.describe_generator_models(bdb, 't1_cc')
for index, modelno, iterations in df.itertuples():
assert iterations == 2
# ----------------------------------
# TEST COLUMN DEPENDENCE PROBABILITY
# ----------------------------------
# Special 0/1 regimes.
# Local with a INDEPENDENT local should be 0.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Date_of_Launch
WITH Country_of_Operator FROM t1 LIMIT 1
''')
assert curs.next()[0] == 0
# Local with a DEPENDENT local should be 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Perigee_km WITH Eccentricity
FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Apogee_km WITH Eccentricity
FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1
# Foreign with a local parent should be 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Period_minutes WITH Apogee_km
FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Anticipated_Lifetime WITH Power_watts
FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
# Foreign with a foreign parent should be 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Type_of_Orbit WITH
Anticipated_Lifetime FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
# Foreign with a local non-parent DEPENDENT with local parent should be 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Period_minutes WITH
Eccentricity FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
# Foreign with foreign sharing common direct ancestor should be 1.
# Launch_Mass_kg is the common parent.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Anticipated_Lifetime WITH
Type_of_Orbit FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
# Foreign with a foreign sharing a common DEPENDENT ancestor should be 1.
# Eccentricity is a parent of Type_of_orbit, and is dependent
# with Period_minutes through DEPENDENT(Apogee_km, Perigee_km, Eccentricity)
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Period_minutes WITH
Type_of_Orbit FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
# Column with itself should be 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Anticipated_Lifetime WITH
Anticipated_Lifetime FROM t1 LIMIT 1
''')
assert curs.next()[0] == 1.
# Unknown [0,1] regimes.
# Foreign with a local of unknown relation with parents.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Anticipated_Lifetime WITH
longitude_radians_of_geo FROM t1 LIMIT 1
''')
assert 0 <= curs.next()[0] <= 1.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Period_minutes WITH
longitude_radians_of_geo FROM t1 LIMIT 1
''')
assert 0 <= curs.next()[0] <= 1.
# Foreign with a foreign of unknown ancestry relation.
curs = bdb.execute('''
ESTIMATE DEPENDENCE PROBABILITY OF Anticipated_Lifetime WITH
Period_minutes FROM t1 LIMIT 1
''')
assert 0 <= curs.next()[0] <= 1.
# ----------------------------------
# TEST SIMULATE
# ----------------------------------
# Crash tests for various code paths. Quality of simulations ignored.
# Joint local.
curs = bdb.execute('''
SIMULATE Power_watts, Launch_Mass_kg FROM t1 LIMIT 2;
''')
assert len(curs.fetchall()) == 2
# Forward simulate foreign.
curs = bdb.execute('''
SIMULATE Period_minutes FROM t1 GIVEN Apogee_km = 1000, Perigee_km = 980
LIMIT 2;
''')
assert len(curs.fetchall()) == 2
# Forward simulate foreign with missing parents.
curs = bdb.execute('''
SIMULATE Anticipated_Lifetime FROM t1 GIVEN Dry_Mass_kg = 2894,
Launch_Mass_kg = 1730 LIMIT 2;
''')
assert len(curs.fetchall()) == 2
# Joint simulate foreign with parents, and missing parents.
curs = bdb.execute('''
SIMULATE Type_of_Orbit, Eccentricity FROM t1 GIVEN Dry_Mass_kg = 2894,
Launch_Mass_kg = 1730 LIMIT 2;
''')
assert len(curs.fetchall()) == 2
# Joint simulate foreign with non-parents.
curs = bdb.execute('''
SIMULATE Period_minutes, Eccentricity FROM t1 GIVEN Apogee_km = 38000
LIMIT 2;
''')
assert len(curs.fetchall()) == 2
# Simulate joint local conditioned on two foreigns.
curs = bdb.execute('''
SIMULATE Country_of_Operator, Inclination_radians FROM t1
GIVEN Period_minutes = 1432, Anticipated_Lifetime = 5 LIMIT 2;
''')
assert len(curs.fetchall()) == 2
# Simulate joint foreign conditioned on third foreign.
curs = bdb.execute('''
SIMULATE Period_minutes, Anticipated_Lifetime FROM t1
GIVEN Type_of_Orbit = 'Deep Highly Eccentric' LIMIT 2
''')
assert len(curs.fetchall()) == 2
# Simulate foreign conditioned on itself.
curs = bdb.execute('''
SIMULATE Period_minutes, Apogee_km FROM t1
GIVEN Period_minutes = 102 LIMIT 2
''')
assert [s[0] for s in curs] == [102] * 2
# -----------------------------
# TEST COLUMN VALUE PROBABILITY
# -----------------------------
# Crash tests for various code path. Quality of logpdf ignored.
# Conditional local.
curs = bdb.execute('''
ESTIMATE PROBABILITY OF Power_watts = 800 GIVEN (Perigee_km = 980,
Launch_Mass_kg = 890) FROM t1 LIMIT 1;
''')
assert 0. <= curs.next()[0]
# Unconditional foreign
curs = bdb.execute('''
ESTIMATE PROBABILITY OF Period_minutes = 1020 FROM t1 LIMIT 1;
''')
assert 0. <= curs.next()[0]
# Conditional foreign on parent and non-parents.
curs = bdb.execute('''
ESTIMATE PROBABILITY OF Period_minutes = 1020 GIVEN
(Apogee_km = 38000, Eccentricity = 0.03) FROM t1 LIMIT 1;
''')
assert 0 <= curs.next()[0]
# Conditional foriegn on foreign.
curs = bdb.execute('''
ESTIMATE PROBABILITY OF Anticipated_Lifetime = 4.09 GIVEN
(Class_of_Orbit = 'LEO', Purpose='Astrophysics',
Period_minutes = 1436) FROM t1 LIMIT 1;
''')
assert 0. <= curs.next()[0]
# Categorical foreign should be less than 1.
curs = bdb.execute('''
ESTIMATE PROBABILITY OF Type_of_Orbit = 'Polar' FROM t1 LIMIT 1;
''')
assert curs.next()[0] <= 1.
# Query inconsistent with evidence should be 0.
curs = bdb.execute('''
ESTIMATE PROBABILITY OF "Type_of_Orbit" = 'Polar'
GIVEN ("Type_of_Orbit" = 'Deep Highly Eccentric') FROM t1 LIMIT 1;
''')
assert curs.next()[0] == 0.
# In theory, query consistent with evidence should be 1, but this is very
# hard to ensure due to stochastic sampling giving different estimates of
# P(Y), once in joint and once in marginal Monte Carlo estimation.
# -----------------------
# TEST MUTUAL INFORMATION
# -----------------------
# Two local columns.
curs = bdb.execute('''
ESTIMATE MUTUAL INFORMATION OF Country_of_Contractor WITH
longitude_radians_of_geo USING 5 SAMPLES FROM t1 LIMIT 1;
''')
# XXX Small sample sizes non-deterministically produce negative MI
assert -1 <= curs.next()[0]
# One local and one foreign column.
curs = bdb.execute('''
ESTIMATE MUTUAL INFORMATION OF Period_minutes WITH
longitude_radians_of_geo USING 5 SAMPLES FROM t1 LIMIT 1;
''')
# XXX This non-deterministically fails when sample sizes are small
# assert 0. <= curs.next()[0]
assert float("-inf") <= curs.next()[0]
# Two foreign columns.
curs = bdb.execute('''
ESTIMATE MUTUAL INFORMATION OF Period_minutes WITH
Anticipated_Lifetime USING 5 SAMPLES FROM t1 LIMIT 1;
''')
# XXX This non-deterministically fails when sample sizes are small
# assert 0. <= curs.next()[0]
assert float("-inf") <= curs.next()[0]
# -----------------------
# TEST PREDICT CONFIDENCE
# -----------------------
# Continuous local column.
curs = bdb.execute('''
INFER EXPLICIT PREDICT Dry_Mass_kg CONFIDENCE c FROM t1 LIMIT 1;
''')
assert curs.next()[1] >= 0.
# Discrete local column with no children.
curs = bdb.execute('''
INFER EXPLICIT PREDICT Purpose CONFIDENCE c FROM t1 LIMIT 1;
''')
assert 0 <= curs.next()[1] <= 1
# Discrete local column with children.
curs = bdb.execute('''
INFER EXPLICIT PREDICT Contractor CONFIDENCE c FROM t1 LIMIT 1;
''')
assert 0 <= curs.next()[1] <= 1
# Continuous foreign columns.
curs = bdb.execute('''
INFER EXPLICIT PREDICT Period_minutes CONFIDENCE c FROM t1 LIMIT 1;
''')
assert curs.next()[1] >= 0.
# Discrete foreign column.
curs = bdb.execute('''
INFER EXPLICIT PREDICT Type_of_Orbit CONFIDENCE c FROM t1 LIMIT 1;
''')
assert 0 <= curs.next()[1] <= 1
bdb.close()
def doit(out_dir, num_models, num_iters, checkpoint_freq, seed):
then = time.time()
timestamp = datetime.datetime.fromtimestamp(then).strftime('%Y-%m-%d')
user = subprocess.check_output(["whoami"]).strip()
host = subprocess.check_output(["hostname"]).strip()
filestamp = '-' + timestamp + '-' + user
def out_file_name(base, ext):
return out_dir + '/' + base + filestamp + ext
csv_file = os.path.join(os.path.dirname(__file__), 'satellites.csv')
bdb_file = out_file_name('satellites', '.bdb')
# so we can build bdb models
os.environ['BAYESDB_WIZARD_MODE']='1'
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
if os.path.exists(bdb_file):
print 'Error: File', bdb_file, 'already exists. Please remove it.'
sys.exit(1)
# create database mapped to filesystem
log('opening bdb on disk: %s' % bdb_file)
bdb = bayeslite.bayesdb_open(pathname=bdb_file, builtin_metamodels=False)
def execute(bql):
log("executing %s" % bql)
bdb.execute(bql)
# read csv into table
log('reading data from %s' % csv_file)
bayeslite.bayesdb_read_csv_file(bdb, 'satellites', csv_file,
header=True, create=True, ifnotexists=True)
# Add a "not applicable" orbit sub-type
log('adding "not applicable" orbit sub-type')
bdb.sql_execute('''UPDATE satellites
SET type_of_orbit = 'N/A'
WHERE (class_of_orbit = 'GEO' OR class_of_orbit = 'MEO')
AND type_of_orbit = 'NaN'
''')
# nullify "NaN"
log('nullifying NaN')
bdbcontrib.nullify(bdb, 'satellites', 'NaN')
# register crosscat metamodel
cc = ccme.MultiprocessingEngine(seed=seed)
ccmm = bayeslite.metamodels.crosscat.CrosscatMetamodel(cc)
bayeslite.bayesdb_register_metamodel(bdb, ccmm)
# create the crosscat generator using
execute('''
CREATE GENERATOR satellites_cc FOR satellites USING crosscat (
GUESS(*),
name IGNORE,
Country_of_Operator CATEGORICAL,
Operator_Owner CATEGORICAL,
Users CATEGORICAL,
Purpose CATEGORICAL,
Class_of_Orbit CATEGORICAL,
Type_of_Orbit CATEGORICAL,
Perigee_km NUMERICAL,
Apogee_km NUMERICAL,
Eccentricity NUMERICAL,
Period_minutes NUMERICAL,
Launch_Mass_kg NUMERICAL,
Dry_Mass_kg NUMERICAL,
Power_watts NUMERICAL,
Date_of_Launch NUMERICAL,
Anticipated_Lifetime NUMERICAL,
Contractor CATEGORICAL,
Country_of_Contractor CATEGORICAL,
Launch_Site CATEGORICAL,
Launch_Vehicle CATEGORICAL,
Source_Used_for_Orbital_Data CATEGORICAL,
longitude_radians_of_geo NUMERICAL,
Inclination_radians NUMERICAL
)
''')
execute('INITIALIZE %d MODELS FOR satellites_cc' % (num_models,))
cur_iter_ct = 0
def snapshot():
log('vacuuming')
bdb.sql_execute('vacuum')
cur_infix = '-%dm-%di' % (num_models, cur_iter_ct)
save_file_name = out_file_name('satellites', cur_infix + '.bdb')
meta_file_name = out_file_name('satellites', cur_infix + '-meta.txt')
log('recording snapshot ' + save_file_name)
os.system("cp %s %s" % (bdb_file, save_file_name))
report(save_file_name, meta_file_name)
def record_metadata(f, saved_file_name, sha_sum, total_time,
plot_file_name=None):
f.write("DB file " + saved_file_name + "\n")
f.write(sha_sum)
f.write("built from " + csv_file + "\n")
f.write("by %s@%s\n" % (user, host))
f.write("at seed %s\n" % seed)
f.write("in %3.2f seconds\n" % total_time)
f.write("with %s models analyzed for %s iterations\n"
% (num_models, num_iters))
f.write("by bayeslite %s, with crosscat %s and bdbcontrib %s\n"
% (bayeslite.__version__, crosscat.__version__, bdbcontrib.__version__))
if plot_file_name is not None:
f.write("diagnostics recorded to %s\n" % plot_file_name)
f.flush()
def report(saved_file_name, metadata_file, echo=False, plot_file_name=None):
sha256 = hashlib.sha256()
with open(saved_file_name, 'rb') as fd:
for chunk in iter(lambda: fd.read(65536), ''):
sha256.update(chunk)
sha_sum = sha256.hexdigest() + '\n'
total_time = time.time() - then
with open(metadata_file, 'w') as fd:
record_metadata(fd, saved_file_name,
sha_sum, total_time, plot_file_name)
fd.write('using script ')
fd.write('-' * 57)
fd.write('\n')
fd.flush()
os.system("cat %s >> %s" % (__file__, metadata_file))
if echo:
record_metadata(sys.stdout, saved_file_name,
sha_sum, total_time, plot_file_name)
def final_report():
# create a diagnostics plot
plot_file_name = out_file_name('satellites', '-logscores.pdf')
log('writing diagnostic plot to %s' % plot_file_name)
_fig = bdbcontrib.plot_crosscat_chain_diagnostics(bdb, 'logscore',
'satellites_cc')
plt.savefig(plot_file_name)
final_metadata_file = out_file_name('satellites', '-meta.txt')
report(bdb_file, final_metadata_file,
echo=True, plot_file_name=plot_file_name)
snapshot()
while cur_iter_ct < num_iters:
execute('ANALYZE satellites_cc FOR %d ITERATIONS WAIT' % checkpoint_freq)
cur_iter_ct += checkpoint_freq
snapshot()
final_report()
log('closing bdb %s' % bdb_file)
bdb.close()
os.system("cd %s && ln -s satellites%s.bdb satellites.bdb" % (out_dir, filestamp))
