def load(batch_size, batch_index):
"""
return summarized data for a table, e.g. a list of points summary.
ignore numeric values with $, %, ", etc., i.e., only take the ones which can be parsed by locale.atof
"""
batch_files = training_files[batch_size * batch_index:batch_size *
(batch_index + 1)]
batch_files_nst = list(
map(lambda batch_file: batch_file.rstrip('.json') + '_nst.csv',
batch_files))
batch_files_wordlist = list(
map(lambda batch_file: batch_file.rstrip('.json') + '_wordlist.csv',
batch_files))
all_nst = [
list(
map(
to_int,
np.genfromtxt(os.path.join(training_data_dir, batch_file_nst),
delimiter=',')[0]))
for batch_file_nst in batch_files_nst
]
targets = np.array([
np.genfromtxt(os.path.join(training_data_dir, batch_file_wordlist),
delimiter=',')
for batch_file_wordlist in batch_files_wordlist
])
results = []
for i in range(len(all_nst)):
result = []
table = Table(
json.load(open(os.path.join(training_data_dir, batch_files[i]))))
attributes = table.get_attributes()
column_num = len(attributes)
target = targets[i]
nst = all_nst[i]
target_transformed = [
index_of(list(map(lambda num: to_int(num), row)), 1)
if idx < column_num else -1
for idx, row in enumerate(target.transpose())
]
for j in range(column_num):
if j >= 10:
break
if nst[j] == nst_encoding([
True, False, False
]) or nst[j] == nst_encoding([True, True, False]):
attribute = attributes[j]
if all(
list(
map(
lambda n: is_numeric(n) or n.upper() in
['', 'NA', 'N/A'], attribute))):
result.append(summary(target_transformed[j], attribute))
results.append(result)
return results
def load_data_12k_with_raw(batch_size, batch_index=0):
# load training data from file, to be implemented
# put size number of data into one array
# start from batch_index batch
# return three arrays: raw, input, target
batch_files = training_files_12k[batch_size * batch_index:batch_size *
(batch_index + 1)]
batch_files_ner = list(
map(lambda batch_file: batch_file.rstrip('.json') + '_ner.csv',
batch_files))
batch_files_wordlist = list(
map(lambda batch_file: batch_file.rstrip('.json') + '_wordlist.csv',
batch_files))
raws = numpy.array([
json.load(
open(os.path.join(training_data_12k_dir, batch_file),
encoding='utf-8')) for batch_file in batch_files
])
inputs = numpy.array([
numpy.genfromtxt(os.path.join(training_data_12k_dir, batch_file_ner),
delimiter=',') for batch_file_ner in batch_files_ner
])
targets = numpy.array([
numpy.genfromtxt(os.path.join(training_data_12k_dir,
batch_file_wordlist),
delimiter=',')
for batch_file_wordlist in batch_files_wordlist
])
inputs_transformed = []
targets_transformed = []
# Use One Hot Encoding
for i in range(len(inputs)):
table = Table(
json.load(open(os.path.join(training_data_12k_dir,
batch_files[i]))))
column_num = len(table.get_header())
input = inputs[i]
target = targets[i]
assert len(input) == len(tag_to_index)
inputs_transformed.append(
numpy.array([
int(
round(
sum(
numpy.array([(2**i) * num
for (i, num) in enumerate(row)]))))
if idx < column_num else -1
for idx, row in enumerate(input.transpose())
]).transpose())
targets_transformed.append(
numpy.array([
index_of(list(map(lambda num: int(round(num)), row)), 1)
if idx < column_num else -1
for idx, row in enumerate(target.transpose())
]).transpose())
return numpy.array(raws), numpy.array(inputs_transformed), numpy.array(
targets_transformed)
def load_nst_majo(batch_size, batch_index):
batch_files = training_files[batch_size * batch_index:batch_size * (batch_index + 1)]
batch_files_nst = list(map(lambda batch_file: batch_file.rstrip('.json') + '_nst.csv', batch_files))
batch_files_wordlist = list(map(lambda batch_file: batch_file.rstrip('.json') + '_wordlist.csv', batch_files))
inputs_major = [
list(map(to_int, numpy.genfromtxt(os.path.join(training_data_dir, batch_file_nst), delimiter=',')[0])) for
batch_file_nst in batch_files_nst]
inputs_max = [
list(map(to_int, numpy.genfromtxt(os.path.join(training_data_dir, batch_file_nst), delimiter=',')[1])) for
batch_file_nst in batch_files_nst]
inputs_overall = [
list(map(to_int, numpy.genfromtxt(os.path.join(training_data_dir, batch_file_nst), delimiter=',')[2])) for
batch_file_nst in batch_files_nst]
targets = numpy.array(
[numpy.genfromtxt(os.path.join(training_data_dir, batch_file_wordlist), delimiter=',') for batch_file_wordlist
in batch_files_wordlist])
targets_transformed = []
for i in range(len(targets)):
table = Table(json.load(open(os.path.join(training_data_dir, batch_files[i]))))
column_num = len(table.get_header())
target = targets[i]
targets_transformed.append(
numpy.array([index_of(list(map(lambda num: int(round(num)), row)), 1) if idx < column_num else -1 for
idx, row in enumerate(target.transpose())]).transpose())
return numpy.array([[x[0] * 4 + x[2] * 1 for x in zip(inputs[0], inputs[1], inputs[2])] for inputs in
zip(inputs_major, inputs_max, inputs_overall)]), numpy.array(targets_transformed)
def load_sample_random_label(sample_index, batch_size, batch_index):
# load testing data of sample with random labels
# put size number of data into one array
# start from batch_index batch
result = []
batch_files = testing_files_random_label[sample_index][
batch_size * batch_index:batch_size * (batch_index + 1)]
for batch_file in batch_files:
table = Table(
json.load(
open(os.path.join(testing_data_random_label_dir, batch_file))))
column_num = len(table.get_header())
batch_file_ner = batch_file.rstrip('.json') + '_ner.csv'
batch_file_wordlist = batch_file.rstrip('.json') + '_wordlist.csv'
batch_file_activate = batch_file.rstrip('.json') + '_activate.json'
input = numpy.genfromtxt(os.path.join(testing_data_random_label_dir,
batch_file_ner),
delimiter=',').transpose()
target = numpy.genfromtxt(os.path.join(testing_data_random_label_dir,
batch_file_wordlist),
delimiter=',').transpose()
activate = json.load(
open(
os.path.join(activate_data_random_label_dir,
batch_file_activate)))
input_transformed = [
int(
round(
sum(
numpy.array([(2**i) * num
for (i, num) in enumerate(row)]))))
if idx < column_num else -1 for idx, row in enumerate(input)
]
target_transformed = [
index_of(list(map(lambda num: int(round(num)), row)), 1)
if idx < column_num else -1 for idx, row in enumerate(target)
]
activate_transformed = [
num if idx < column_num else -1 for idx, num in enumerate(activate)
]
result.append(
[input_transformed, target_transformed, activate_transformed])
return result
def load_data(batch_size, batch_index=0):
# load training data from file, to be implemented
# put size number of data into one array
# start from batch_index batch
# return two arrays: input, target
batch_files = training_files[batch_size * batch_index:batch_size * (batch_index + 1)]
batch_files_ner = list(map(lambda batch_file: batch_file.rstrip('.json') + '_ner.csv', batch_files))
batch_files_nst = list(map(lambda batch_file: batch_file.rstrip('.json') + '_nst.csv', batch_files))
batch_files_date = list(map(lambda batch_file: batch_file.rstrip('.json') + '_date.csv', batch_files))
batch_files_wordlist = list(map(lambda batch_file: batch_file.rstrip('.json') + '_wordlist.csv', batch_files))
ner_inputs = numpy.array(
[numpy.genfromtxt(os.path.join(training_data_dir, batch_file_ner), delimiter=',') for batch_file_ner in
batch_files_ner])
nst_inputs = numpy.array(
[list(map(to_int, numpy.genfromtxt(os.path.join(training_data_dir, batch_file_nst), delimiter=',')[0])) for
batch_file_nst in batch_files_nst])
date_inputs = numpy.array(
[list(map(to_int, numpy.genfromtxt(os.path.join(training_data_dir, batch_file_date), delimiter=','))) for
batch_file_date in batch_files_date])
targets = numpy.array(
[numpy.genfromtxt(os.path.join(training_data_dir, batch_file_wordlist), delimiter=',') for batch_file_wordlist
in batch_files_wordlist])
inputs_transformed = []
targets_transformed = []
assert len(ner_inputs) == len(nst_inputs)
assert len(ner_inputs) == len(date_inputs)
for i in range(len(ner_inputs)):
# print(batch_files[i])
table = Table(json.load(open(os.path.join(training_data_dir, batch_files[i]))))
column_num = len(table.get_header())
attributes = table.get_attributes()
ner_input = ner_inputs[i]
nst_input = nst_inputs[i]
date_input = date_inputs[i]
target = targets[i]
assert len(ner_input) == len(tag_to_index)
# Encode 3 class NER (4:location, 5:person, 6:organization)
new_input_transformed = numpy.array([int(round(sum([(2 ** (i + 3)) * num for (i, num) in enumerate(ner_row)])))
if idx < column_num else -1 for idx, ner_row in
enumerate(ner_input.transpose())]).transpose()
# print('ner', new_input_transformed)
# Add encoded NST and date (1:text, 2:symbol, 3:number, 7:date)
new_input_transformed = new_input_transformed + numpy.array(nst_input) + numpy.array(date_input) * (2 ** 6)
# print('nst', numpy.array(nst_input))
# print('date', numpy.array(date_input) * (2 ** 6))
# Check is_numeric, is_float and is_ordered (8:is_numeric, 9:is_float, 10:is_ordered)
is_numeric_input = [-1] * 10
is_float_input = [-1] * 10
is_ordered_input = [-1] * 10
for idx in range(min(column_num, 10)):
is_numeric_input[idx] = 0
is_float_input[idx] = 0
is_ordered_input[idx] = 0
if nst_input[idx] == nst_encoding([True, False, False]) or \
nst_input[idx] == nst_encoding([True, True, False]):
attribute = attributes[idx]
if all(list(map(lambda n: is_numeric(n) or n.upper() in ['', 'NA', 'N/A'], attribute))):
is_numeric_input[idx] = 1
is_float_input[idx] = int('.' in ''.join(attribute))
locale.setlocale(locale.LC_ALL, 'en_US.UTF-8')
values = numpy.array(list(map(locale.atof, filter(is_numeric, attribute))))
# 0: random, 1: desc, 2: asc
is_ordered_input[idx] = 2 if numpy.all(numpy.diff(values) > 0) else \
1 if numpy.all(numpy.diff(values) < 0) else 0
new_input_transformed = new_input_transformed + \
numpy.array(is_numeric_input) * (2 ** 7) + \
numpy.array(is_float_input) * (2 ** 8) + \
numpy.array(is_ordered_input) * (2 ** 9)
# print('is_numeric', numpy.array(is_numeric_input) * (2 ** 7))
# print('is_float', numpy.array(is_float_input) * (2 ** 8))
# print('is_ordered', numpy.array(is_ordered_input) * (2 ** 9))
# Change all negative values to -1 (empty column)
new_input_transformed = numpy.array([x if x >= 0 else -1 for x in new_input_transformed])
# print('overall', new_input_transformed)
inputs_transformed.append(new_input_transformed)
targets_transformed.append(
numpy.array([index_of(list(map(lambda num: int(round(num)), row)), 1) if idx < column_num else -1 for
idx, row in enumerate(target.transpose())]).transpose())
return numpy.array(inputs_transformed), numpy.array(targets_transformed)
