def run_experiment(argin):
num_iters = argin["num_iters"]
num_chains = argin["num_chains"]
num_rows = argin["num_rows"]
num_cols = argin["num_cols"]
num_views = argin["num_views"]
num_clusters = argin["num_clusters"]
prop_missing = argin["prop_missing"]
separation = argin["separation"]
ct_kernel = argin["ct_kernel"]
multinomial_categories = argin["multinomial_categories"]
seed = argin["seed"]
random.seed(seed)
# TODO: use dha.csv
ofilename = "reasonably_calibrated_ofile.csv"
table_name = "reasonably_calibrated"
argin["distargs"] = [{"K": multinomial_categories}] * num_cols
argin["cctypes"] = ["multinomial"] * num_cols
argin["separation"] = [argin["separation"]] * num_views
T_array, structure = eu.gen_data(ofilename, argin, save_csv=True)
filename, indices, col_names = eu.gen_missing_data_csv(ofilename, prop_missing, [])
# create a client
client = Client()
# caluclate empirical frequency of each point
frequencies = []
for col in range(num_cols):
frequencies.append(numpy.zeros(multinomial_categories))
T_int = numpy.array(T_array, dtype=int)
n_indices = len(indices[0])
for i in range(n_indices):
r = indices[0][i]
c = indices[1][i]
x = T_int[r, c]
frequencies[c][x] += 1.0
frequencies = [f / numpy.sum(f) for f in frequencies]
# set up a dict fro the different config data
result = dict()
# do analyses
for config in ["cc", "crp", "nb"]:
config_string = eu.config_map[config]
table = table_name + "-" + config
# drop old btable, create a new one with the new data and init models
client("DROP BTABLE %s;" % table, yes=True)
client("CREATE BTABLE %s FROM %s;" % (table, filename))
client("INITIALIZE %i MODELS FOR %s %s;" % (num_chains, table, config_string))
if ct_kernel == 1:
client("ANALYZE %s FOR %i ITERATIONS WITH MH KERNEL WAIT;" % (table, num_iters))
else:
client("ANALYZE %s FOR %i ITERATIONS WAIT;" % (table, num_iters))
# imput each index in indices and calculate the squared error
results_config = []
for col in range(num_cols):
results_config.append(numpy.zeros(multinomial_categories))
for col in range(num_cols):
col_name = col_names[col]
out = client(
"INFER %s FROM %s WITH CONFIDENCE .95 WITH 1 SAMPLES;" % (col_name, table),
pretty=False,
pandas_output=False,
)
for i in range(n_indices):
r = indices[0][i]
c = indices[1][i]
if c == col:
x = out[0]["data"][r][1]
results_config[c][int(x)] += 1.0
results_config = [f / sum(f) for f in results_config]
result[config] = results_config
retval = dict()
retval["actual_frequencies"] = frequencies
retval["inferred_P_cc"] = result["cc"]
retval["inferred_P_crp"] = result["crp"]
retval["inferred_P_nb"] = result["nb"]
retval["config"] = argin
return retval
def run_experiment(argin):
num_iters = argin["num_iters"]
num_chains = argin["num_chains"]
num_rows = argin["num_rows"]
num_cols = argin["num_cols"]
num_views = argin["num_views"]
num_clusters = argin["num_clusters"]
prop_missing = argin["prop_missing"]
impute_samples = argin["impute_samples"]
separation = argin["separation"]
ct_kernel = argin["ct_kernel"]
seed = argin["seed"]
if seed > 0 :
random.seed(seed)
filename = "exp_fills_in_ofile.csv"
table_name = 'exp_fills_in'
argin['cctypes'] = ['continuous']*num_cols
argin['separation'] = [argin['separation']]*num_views
eu.gen_data(filename, argin, save_csv=True)
# generate a new csv
all_filenames = []
all_indices = []
for p in prop_missing:
data_filename, indices, col_names, extra = eu.gen_missing_data_csv(filename,
p, [], True)
all_indices.append(indices)
all_filenames.append(data_filename)
# get the starting table so we can calculate errors
T_array = extra['array_filled']
num_rows, num_cols = T_array.shape
# create a client
client = Client()
# set up a dict fro the different config data
result = dict()
result['cc'] = numpy.zeros(len(prop_missing))
result['crp'] = numpy.zeros(len(prop_missing))
result['nb'] = numpy.zeros(len(prop_missing))
# do analyses
for p in range(len(prop_missing)):
this_indices = all_indices[p]
this_filename = all_filenames[p]
for config in ['cc', 'crp', 'nb']:
config_string = eu.config_map[config]
table = table_name + '-' + config
# drop old btable, create a new one with the new data and init models
client('DROP BTABLE %s;' % table, yes=True)
client('CREATE BTABLE %s FROM %s;' % (table, this_filename))
client('INITIALIZE %i MODELS FOR %s %s;' % (num_chains, table, config_string))
if ct_kernel == 1:
client('ANALYZE %s FOR %i ITERATIONS WITH MH KENEL WAIT;' % (table, num_iters) )
else:
client('ANALYZE %s FOR %i ITERATIONS WAIT;' % (table, num_iters) )
MSE = 0.0
count = 0.0
# imput each index in indices and calculate the squared error
for col in range(0,num_cols):
col_name = col_names[col]
# confidence is set to zero so that a value is always returned
out = client('INFER %s from %s WITH CONFIDENCE %f WITH %i SAMPLES;' % (col_name, table, 0, impute_samples), pretty=False, pandas_output=False )
data = out[0]['data']
# calcaulte MSE
for row, tcol in zip(this_indices[0], this_indices[1]):
if tcol == col:
MSE += ( T_array[row,col] - data[row][1] )**2.0
count += 1.0
result[config][p] = MSE/count
print "error = %f" % result[config][p]
retval = dict()
retval['MSE_naive_bayes_indexer'] = result['nb']
retval['MSE_crp_mixture_indexer'] = result['crp']
retval['MSE_crosscat_indexer'] = result['cc']
retval['prop_missing'] = prop_missing
retval['config'] = argin
return retval
def run_experiment(argin):
num_iters = argin["num_iters"]
num_chains = argin["num_chains"]
num_rows = argin["num_rows"]
num_cols = argin["num_cols"]
num_views = argin["num_views"]
num_clusters = argin["num_clusters"]
separation = argin["separation"]
seed = argin["seed"]
ct_kernel = argin["ct_kernel"]
if seed > 0:
random.seed(seed)
argin['cctypes'] = ['continuous'] * num_cols
argin['separation'] = [argin['separation']] * num_views
# have to generate synthetic data
filename = "exp_estimate_joint_ofile.csv"
table_name = 'exp_estimate_joint'
# generate starting data
T_o, structure = eu.gen_data(filename, argin, save_csv=True)
# generate a new csv with bottom row removed (held-out data)
data_filename = 'exp_estimate_joint.csv'
T_h = eu.gen_held_out_data(filename, data_filename, 1)
# get the column names
with open(filename, 'r') as f:
csv_header = f.readline()
col_names = csv_header.split(',')
col_names[-1] = col_names[-1].strip()
# set up a dict fro the different config data
result = dict()
true_held_out_p = []
for col in range(num_cols):
x = T_o[-1, col]
logp = eu.get_true_logp(numpy.array([x]), col, structure)
true_held_out_p.append(numpy.exp(logp))
# start a client
client = Client()
# do analyses
for config in ['cc', 'crp', 'nb']:
config_string = eu.config_map[config]
table = table_name + '-' + config
# drop old btable, create a new one with the new data and init models
client('DROP BTABLE %s;' % table, yes=True)
client('CREATE BTABLE %s FROM %s;' % (table, data_filename))
client('INITIALIZE %i MODELS FOR %s %s;' %
(num_chains, table, config_string))
these_ps = numpy.zeros(num_iters)
these_ps_errors = numpy.zeros(num_iters)
for i in range(num_iters):
if ct_kernel == 1:
client('ANALYZE %s FOR 1 ITERATIONS WITH MH KERNEL WAIT;' %
table)
else:
client('ANALYZE %s FOR 1 ITERATIONS WAIT;' % table)
# imput each index in indices and calculate the squared error
mean_p = []
mean_p_error = []
for col in range(0, num_cols):
col_name = col_names[col]
x = T_o[-1, col]
out = client('SELECT PROBABILITY OF %s=%f from %s;' %
(col_name, x, table),
pretty=False,
pandas_output=False)
p = out[0]['data'][0][1]
mean_p.append(p)
mean_p_error.append((true_held_out_p[col] - p)**2.0)
these_ps[i] = numpy.mean(mean_p)
these_ps_errors[i] = numpy.mean(mean_p_error)
key_str_p = 'mean_held_out_p_' + config
key_str_error = 'mean_error_' + config
result[key_str_p] = these_ps
result[key_str_error] = these_ps_errors
retval = dict()
retval['MSE_naive_bayes_indexer'] = result['mean_error_nb']
retval['MSE_crp_mixture_indexer'] = result['mean_error_crp']
retval['MSE_crosscat_indexer'] = result['mean_error_cc']
retval['MEAN_P_naive_bayes_indexer'] = result['mean_held_out_p_nb']
retval['MEAN_P_crp_mixture_indexer'] = result['mean_held_out_p_crp']
retval['MEAN_P_crosscat_indexer'] = result['mean_held_out_p_cc']
retval['config'] = argin
return retval
def run_experiment(argin):
num_iters = argin["num_iters"]
num_chains = argin["num_chains"]
num_rows = argin["num_rows"]
num_cols = argin["num_cols"]
num_views = argin["num_views"]
num_clusters = argin["num_clusters"]
prop_missing = argin["prop_missing"]
impute_samples = argin["impute_samples"]
separation = argin["separation"]
ct_kernel = argin["ct_kernel"]
seed = argin["seed"]
if seed > 0:
random.seed(seed)
filename = "exp_fills_in_ofile.csv"
table_name = 'exp_fills_in'
argin['cctypes'] = ['continuous'] * num_cols
argin['separation'] = [argin['separation']] * num_views
eu.gen_data(filename, argin, save_csv=True)
# generate a new csv
all_filenames = []
all_indices = []
for p in prop_missing:
data_filename, indices, col_names, extra = eu.gen_missing_data_csv(
filename, p, [], True)
all_indices.append(indices)
all_filenames.append(data_filename)
# get the starting table so we can calculate errors
T_array = extra['array_filled']
num_rows, num_cols = T_array.shape
# create a client
client = Client()
# set up a dict fro the different config data
result = dict()
result['cc'] = numpy.zeros(len(prop_missing))
result['crp'] = numpy.zeros(len(prop_missing))
result['nb'] = numpy.zeros(len(prop_missing))
# do analyses
for p in range(len(prop_missing)):
this_indices = all_indices[p]
this_filename = all_filenames[p]
for config in ['cc', 'crp', 'nb']:
config_string = eu.config_map[config]
table = table_name + '-' + config
# drop old btable, create a new one with the new data and init models
client('DROP BTABLE %s;' % table, yes=True)
client('CREATE BTABLE %s FROM %s;' % (table, this_filename))
client('INITIALIZE %i MODELS FOR %s %s;' %
(num_chains, table, config_string))
if ct_kernel == 1:
client('ANALYZE %s FOR %i ITERATIONS WITH MH KENEL WAIT;' %
(table, num_iters))
else:
client('ANALYZE %s FOR %i ITERATIONS WAIT;' %
(table, num_iters))
MSE = 0.0
count = 0.0
# imput each index in indices and calculate the squared error
for col in range(0, num_cols):
col_name = col_names[col]
# confidence is set to zero so that a value is always returned
out = client(
'INFER %s from %s WITH CONFIDENCE %f WITH %i SAMPLES;' %
(col_name, table, 0, impute_samples),
pretty=False,
pandas_output=False)
data = out[0]['data']
# calcaulte MSE
for row, tcol in zip(this_indices[0], this_indices[1]):
if tcol == col:
MSE += (T_array[row, col] - data[row][1])**2.0
count += 1.0
result[config][p] = MSE / count
print "error = %f" % result[config][p]
retval = dict()
retval['MSE_naive_bayes_indexer'] = result['nb']
retval['MSE_crp_mixture_indexer'] = result['crp']
retval['MSE_crosscat_indexer'] = result['cc']
retval['prop_missing'] = prop_missing
retval['config'] = argin
return retval
def run_experiment(argin):
num_iters = argin["num_iters"]
num_chains = argin["num_chains"]
num_rows = argin["num_rows"]
num_cols = argin["num_cols"]
num_views = argin["num_views"]
num_clusters = argin["num_clusters"]
separation = argin["separation"]
seed = argin["seed"]
ct_kernel = argin["ct_kernel"]
if seed > 0:
random.seed(seed)
argin['cctypes'] = ['continuous']*num_cols
argin['separation'] = [argin['separation']]*num_views
# have to generate synthetic data
filename = "exp_estimate_joint_ofile.csv"
table_name = 'exp_estimate_joint'
# generate starting data
T_o, structure = eu.gen_data(filename, argin, save_csv=True)
# generate a new csv with bottom row removed (held-out data)
data_filename = 'exp_estimate_joint.csv'
T_h = eu.gen_held_out_data(filename, data_filename, 1)
# get the column names
with open(filename, 'r') as f:
csv_header = f.readline()
col_names = csv_header.split(',')
col_names[-1] = col_names[-1].strip()
# set up a dict fro the different config data
result = dict()
true_held_out_p = []
for col in range(num_cols):
x = T_o[-1,col]
logp = eu.get_true_logp(numpy.array([x]), col, structure)
true_held_out_p.append(numpy.exp(logp))
# start a client
client = Client()
# do analyses
for config in ['cc', 'crp', 'nb']:
config_string = eu.config_map[config]
table = table_name + '-' + config
# drop old btable, create a new one with the new data and init models
client('DROP BTABLE %s;' % table, yes=True)
client('CREATE BTABLE %s FROM %s;' % (table, data_filename))
client('INITIALIZE %i MODELS FOR %s %s;' % (num_chains, table, config_string))
these_ps = numpy.zeros(num_iters)
these_ps_errors = numpy.zeros(num_iters)
for i in range(num_iters):
if ct_kernel == 1:
client('ANALYZE %s FOR 1 ITERATIONS WITH MH KERNEL WAIT;' % table )
else:
client('ANALYZE %s FOR 1 ITERATIONS WAIT;' % table )
# imput each index in indices and calculate the squared error
mean_p = []
mean_p_error = []
for col in range(0,num_cols):
col_name = col_names[col]
x = T_o[-1,col]
out = client('SELECT PROBABILITY OF %s=%f from %s;' % (col_name, x, table), pretty=False, pandas_output=False)
p = out[0]['data'][0][1]
mean_p.append(p)
mean_p_error.append( (true_held_out_p[col]-p)**2.0 )
these_ps[i] = numpy.mean(mean_p)
these_ps_errors[i] = numpy.mean(mean_p_error)
key_str_p = 'mean_held_out_p_' + config
key_str_error = 'mean_error_' + config
result[key_str_p] = these_ps
result[key_str_error] = these_ps_errors
retval = dict()
retval['MSE_naive_bayes_indexer'] = result['mean_error_nb']
retval['MSE_crp_mixture_indexer'] = result['mean_error_crp']
retval['MSE_crosscat_indexer'] = result['mean_error_cc']
retval['MEAN_P_naive_bayes_indexer'] = result['mean_held_out_p_nb']
retval['MEAN_P_crp_mixture_indexer'] = result['mean_held_out_p_crp']
retval['MEAN_P_crosscat_indexer'] = result['mean_held_out_p_cc']
retval['config'] = argin
return retval
def run_experiment(argin):
num_iters = argin["num_iters"]
num_chains = argin["num_chains"]
num_rows = argin["num_rows"]
num_cols = argin["num_cols"]
num_views = argin["num_views"]
num_clusters = argin["num_clusters"]
prop_missing = argin["prop_missing"]
separation = argin["separation"]
ct_kernel = argin["ct_kernel"]
multinomial_categories = argin["multinomial_categories"]
seed = argin["seed"]
random.seed(seed)
# TODO: use dha.csv
ofilename = "reasonably_calibrated_ofile.csv"
table_name = 'reasonably_calibrated'
argin['distargs'] = [{"K": multinomial_categories}] * num_cols
argin['cctypes'] = ['multinomial'] * num_cols
argin['separation'] = [argin['separation']] * num_views
T_array, structure = eu.gen_data(ofilename, argin, save_csv=True)
filename, indices, col_names = eu.gen_missing_data_csv(
ofilename, prop_missing, [])
# create a client
client = Client()
# caluclate empirical frequency of each point
frequencies = []
for col in range(num_cols):
frequencies.append(numpy.zeros(multinomial_categories))
T_int = numpy.array(T_array, dtype=int)
n_indices = len(indices[0])
for i in range(n_indices):
r = indices[0][i]
c = indices[1][i]
x = T_int[r, c]
frequencies[c][x] += 1.0
frequencies = [f / numpy.sum(f) for f in frequencies]
# set up a dict fro the different config data
result = dict()
# do analyses
for config in ['cc', 'crp', 'nb']:
config_string = eu.config_map[config]
table = table_name + '-' + config
# drop old btable, create a new one with the new data and init models
client('DROP BTABLE %s;' % table, yes=True)
client('CREATE BTABLE %s FROM %s;' % (table, filename))
client('INITIALIZE %i MODELS FOR %s %s;' %
(num_chains, table, config_string))
if ct_kernel == 1:
client('ANALYZE %s FOR %i ITERATIONS WITH MH KERNEL WAIT;' %
(table, num_iters))
else:
client('ANALYZE %s FOR %i ITERATIONS WAIT;' % (table, num_iters))
# imput each index in indices and calculate the squared error
results_config = []
for col in range(num_cols):
results_config.append(numpy.zeros(multinomial_categories))
for col in range(num_cols):
col_name = col_names[col]
out = client(
"INFER %s FROM %s WITH CONFIDENCE .95 WITH 1 SAMPLES;" %
(col_name, table),
pretty=False,
pandas_output=False)
for i in range(n_indices):
r = indices[0][i]
c = indices[1][i]
if c == col:
x = out[0]['data'][r][1]
results_config[c][int(x)] += 1.0
results_config = [f / sum(f) for f in results_config]
result[config] = results_config
retval = dict()
retval['actual_frequencies'] = frequencies
retval['inferred_P_cc'] = result['cc']
retval['inferred_P_crp'] = result['crp']
retval['inferred_P_nb'] = result['nb']
retval['config'] = argin
return retval