def __init__(self, directory, pixel_noise_stddev=0.0):
values = dict(world1='world',
world2='world',
world3='world',
world_model1='model',
world_model2='model',
world_model3='model',
description='language',
description_length='int',
description_model='model',
agreement='float')
world_size = tuple(
next(nlvr_util.images_iter(directory=directory,
mode='train'))[1][0].shape[:2])
self.description_size = 0
vocabulary = set()
for _, _, description, _ in nlvr_util.descriptions_iter(
directory=directory, mode='train'):
self.description_size = max(self.description_size,
len(description))
vocabulary.update(description)
vocabularies = dict(language=sorted(vocabulary))
super(NLVR,
self).__init__(values=values,
world_size=world_size,
pixel_noise_stddev=pixel_noise_stddev,
vectors=dict(description=self.description_size),
vocabularies=vocabularies)
self.nlvr = {
mode: nlvr_util.nlvr(directory=directory, mode=mode)
for mode in ('train', 'validation', 'test')
}
def __init__(self, directory):
world_size = tuple(next(nlvr_util.images_iter(directory=directory, mode='train'))[1][0].shape[:2])
self.description_size = 0
vocabulary = set()
for _, _, description, _ in nlvr_util.descriptions_iter(directory=directory, mode='train'):
self.description_size = max(self.description_size, len(description))
vocabulary.update(description)
vocabularies = dict(language=sorted(vocabulary))
super(NLVR, self).__init__(world_size=world_size, vectors=dict(description=self.description_size), vocabularies=vocabularies)
self.nlvr = {mode: nlvr_util.nlvr(directory=directory, mode=mode) for mode in ('train', 'validation', 'test')}
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')
}