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
unique_answers = set()
vocabulary = set()
for _, question, _, answer in clevr_util.questions_iter(
directory=directory, parts=parts, mode='train'):
self.question_size = max(self.question_size, len(question))
vocabulary.update(question)
assert len(answer) == 1
unique_answers.add(answer[0])
vocabularies = dict(language=sorted(vocabulary))
self.answers = sorted(unique_answers)
self.answers.append('[UNKNOWN]')
self.num_answers = len(self.answers)
super(CLEVR, self).__init__(world_size=world_size,
vectors=dict(question=self.question_size),
vocabularies=vocabularies)
self.clevr = {
mode: clevr_util.clevr(directory=directory, parts=parts, mode=mode)
for mode in ('train', 'validation', 'test')
}
def __init__(self, directory, parts=None, pixel_noise_stddev=0.0):
parts = parts.split(',')
values = dict(world='world',
world_model='model',
question='alternatives(language)',
question_length='alternatives(int)',
question_model='alternatives(model)',
answer='alternatives(language)',
answer_length='alternatives(int)',
alternatives='int')
world_size = tuple(
next(
clevr_util.images_iter(directory=directory,
parts=parts,
mode='train')).shape[:2])
self.question_size = 0
self.answer_size = 0
vocabulary = set()
for _, question, _, answer in clevr_util.questions_iter(
directory=directory, parts=parts, mode='train'):
self.question_size = max(self.question_size, len(question))
self.answer_size = max(self.answer_size, len(answer))
vocabulary.update(question)
vocabulary.update(answer)
vocabularies = dict(language=sorted(vocabulary))
super(CLEVR, self).__init__(values=values,
world_size=world_size,
pixel_noise_stddev=pixel_noise_stddev,
vectors=dict(question=self.question_size,
answer=self.answer_size),
vocabularies=vocabularies)
self.clevr = {
mode: clevr_util.clevr(directory=directory, parts=parts, mode=mode)
for mode in ('train', 'validation', 'test')
}
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
vocabulary = set()
for _, question, _, answer in clevr_util.questions_iter(directory=directory, parts=parts, mode='train'):
self.question_size = max(self.question_size, len(question))
self.answer_size = max(self.answer_size, len(answer))
vocabulary.update(question)
vocabulary.update(answer)
vocabularies = dict(language=sorted(vocabulary))
super(CLEVR, self).__init__(world_size=world_size, vectors=dict(question=self.question_size, answer=self.answer_size), vocabularies=vocabularies)
self.clevr = {mode: clevr_util.clevr(directory=directory, parts=parts, mode=mode) for mode in ('train', 'validation', 'test')}
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 __init__(self, directory, parts=None, pixel_noise_stddev=None):
values = dict(world='world',
world_model='model',
question='alternatives(language)',
question_length='alternatives(int)',
question_model='alternatives(model)',
answer='alternatives(int)',
alternatives='int')
if parts is not None:
parts = parts.split(',')
world_size = tuple(
next(
clevr_util.images_iter(directory=directory,
parts=parts,
mode='train')).shape[:2])
self.question_size = 0
unique_answers = set()
vocabulary = set()
for _, question, _, answer in clevr_util.questions_iter(
directory=directory, parts=parts, mode='train'):
self.question_size = max(self.question_size, len(question))
vocabulary.update(question)
assert len(answer) == 1
unique_answers.add(answer[0])
vocabularies = dict(language=sorted(vocabulary))
self.answers = sorted(unique_answers)
self.answers.append('[UNKNOWN]')
self.num_answers = len(self.answers)
super(ClevrDataset,
self).__init__(values=values,
world_size=world_size,
pixel_noise_stddev=pixel_noise_stddev,
vectors=dict(question=self.question_size),
vocabularies=vocabularies)
self.clevr = {
mode: clevr_util.clevr(directory=directory, parts=parts, mode=mode)
for mode in ('train', 'validation', 'test')
}