def __init__(self, directory, parts=None):
world_size = tuple(next(clevr_util.images_iter(directory=directory, parts=parts, mode='train')).shape[:2])
self.question_size = 0
self.answer_size = 0
words = set()
for _, question, _, answer in clevr_util.questions_iter(directory=directory, parts=parts, mode='train'):
question = util.string2tokens(string=question)
answer = util.string2tokens(string=answer)
self.question_size = max(self.question_size, len(question))
self.answer_size = max(self.answer_size, len(answer))
words.update(question)
words.update(answer)
words = sorted(words)
super(CLEVRDataset, self).__init__(world_size=world_size, vectors=dict(question=self.question_size, answer=self.answer_size), words=words)
self.clevr = {mode: clevr_util.clevr(directory=directory, parts=parts, mode=mode) for mode in ('train', 'validation', 'test')}
def realize(self, captions):
try:
ace = subprocess.Popen([self.ace_path, '-g', self.erg_path, '-1Te', '-r', 'root_strict'], stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
except Exception as e:
import sys
from datetime import datetime
print(datetime.now().strftime('%H:%M:%S'))
print(e.strerror)
print(sys.exc_info()[0])
raise
dmrs_list = list()
mrs_list = list()
for caption in captions:
dmrs = self.clause_dmrs(caption)
dmrs.apply_paraphrases(self.post_processing.values())
dmrs.remove_underspecifications()
dmrs_list.append(dmrs)
mrs_list.append(dmrs.get_mrs() + '\n')
stdout_data, stderr_data = ace.communicate(''.join(mrs_list).encode())
stderr_data = stderr_data.decode('utf-8').splitlines()
stdout_data = stdout_data.decode('utf-8').splitlines()
assert all(self.regex.match(line) for line in stderr_data), '\n\n' + '\n'.join('{}\n{}\n{}\n'.format(line, dmrs.dumps_xml().decode(), mrs) for line, dmrs, mrs in zip(stderr_data, dmrs_list, mrs_list) if not self.regex.match(line)) + '\nFailures: {}\n'.format(len(captions) - int(stderr_data[-2][16:stderr_data[-2].index(' ', 16)])) # self.proposition_dmrs(caption).dumps_xml()
caption_strings = [line for line in stdout_data if line]
assert len(caption_strings) == len(captions)
for n, caption in enumerate(caption_strings):
captions[n] = util.string2tokens(string=caption)
return captions
def realize(self, captions):
try:
ace = subprocess.Popen([
self.ace_path, '-g', self.erg_path, '-1e', '-r', 'root_strict'
],
stdin=subprocess.PIPE,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE)
except Exception as e:
import sys
from datetime import datetime
print(datetime.now().strftime('%H:%M:%S'))
print(e.strerror)
print(sys.exc_info()[0])
raise
dmrs_list = list()
mrs_list = list()
for caption in captions:
dmrs = self.caption_dmrs(caption=caption)
dmrs = dmrs.apply_paraphrases(self.post_processing.values())
dmrs.remove_underspecifications()
dmrs_list.append(dmrs)
mrs_list.append(dmrs.get_mrs() + '\n')
stdout_data, stderr_data = ace.communicate(
input=''.join(mrs_list).encode())
stderr_data = stderr_data.decode('utf-8').splitlines()
stdout_data = stdout_data.decode('utf-8').splitlines()
failures = 0
n = 0
unexpected = False
for line in stderr_data:
if n == len(captions):
assert self.final_regex.match(line), line
continue
if self.successful_regex.match(line):
if unexpected:
print(dmrs_list[n].dumps_xml().decode())
print(mrs_list[n])
unexpected = False
n += 1
elif self.unsuccessful_regex.match(line):
print(dmrs_list[n].dumps_xml().decode())
print(mrs_list[n])
failures += 1
n += 1
else:
print('Unexpected: ' + line)
unexpected = True
if failures > 0:
print('Failures: {}'.format(failures))
exit(0)
caption_strings = [line for line in stdout_data if line]
assert len(caption_strings) == len(
captions), stdout_data + '\n' + stderr_data
for n, caption in enumerate(caption_strings):
captions[n] = util.string2tokens(string=caption)
return captions
def __init__(self, directory):
world_size = tuple(
next(nlvr_util.images_iter(directory=directory,
mode='train'))[1][0].shape[:2])
self.description_size = 0
words = set()
for _, _, description, _ in nlvr_util.descriptions_iter(
directory=directory, mode='train'):
description = util.string2tokens(string=description)
self.description_size = max(self.description_size,
len(description))
words.update(description)
words = sorted(words)
super(NLVRDataset,
self).__init__(world_size=world_size,
vectors=dict(description=self.description_size),
words=words)
self.nlvr = {
mode: nlvr_util.nlvr(directory=directory, mode=mode)
for mode in ('train', 'validation', 'test')
}
def descriptions_iter(directory, mode):
mode = 'dev' if mode == 'validation' else mode
path = os.path.join(directory, mode, mode + '.json')
with open(path, 'r') as filehandle:
for line in filehandle:
line = line.strip()
description_dict = json.loads(s=line)
identifier = description_dict['identifier']
assert identifier[-2:] in ('-0', '-1', '-2', '-3')
world_model1, world_model2, world_model3 = description_dict[
'structured_rep']
description = description_dict['sentence'].lower()
if description[-1] != '.':
description += '.'
description = util.string2tokens(string=description)
agreement = description_dict['label']
assert agreement in ('true', 'false')
agreement = (agreement == 'true')
assert len(
description_dict['evals']) == (1 if mode == 'train' else 5)
assert len(description_dict) == 5
yield identifier, (world_model1, world_model2,
world_model3), description, agreement
def questions_iter(directory, mode, parts=None):
split = 'val' if mode == 'validation' else mode
if parts is not None:
split += parts[mode]
path = os.path.join(directory, 'questions',
'CLEVR_{}_questions.json'.format(split))
with open(path, 'r') as filehandle:
chars = filehandle.read(2)
assert chars == '{"'
chars = filehandle.read(1)
while chars != 'q':
while filehandle.read(1) != '"':
pass
chars = filehandle.read(3)
assert chars == ': {'
while filehandle.read(1) != '}':
pass
chars = filehandle.read(3)
assert chars == ', "'
chars = filehandle.read(1)
chars = filehandle.read(11)
assert chars == 'uestions": '
image_index = 0
for n, question_dict in enumerate(json_list_generator(fp=filehandle)):
if image_index != question_dict['image_index']:
image_index += 1
assert image_index == question_dict['image_index']
question = question_dict['question'].lower()
if question[-1] != '?':
question += '?'
question = util.string2tokens(string=question)
if mode == 'test':
question_model = dict()
answer = '[UNKNOWN]'
else:
family = question_dict['question_family_index']
program = question_dict['program']
question_model = dict(family=family, program=program)
answer = question_dict['answer'].lower()
if answer in numbers:
answer = numbers[answer]
answer = util.string2tokens(string=answer)
assert question_dict['question_index'] == n
assert question_dict['split'] == split
assert question_dict[
'image_filename'] == 'CLEVR_{}_{:0>6}.png'.format(
split, image_index)
assert len(question_dict) == 8 or (mode == 'test'
and len(question_dict) == 5)
yield image_index, question, question_model, answer
chars = filehandle.read(1)
while chars == ',':
chars = filehandle.read(2)
assert chars == ' "'
while filehandle.read(1) != '"':
pass
chars = filehandle.read(3)
assert chars == ': {'
while filehandle.read(1) != '}':
pass
chars = filehandle.read(1)
assert chars == '}'
chars = filehandle.read()
assert not chars
