def make_unmnist(config):
"""Creates the MNIST dataset."""
def to_float(x):
return tf.to_float(x) / 255.
transform = [to_float]
if config.canvas_size != 28:
transform.append(
functools.partial(preprocess.pad_and_shift,
output_size=config.canvas_size,
shift=None))
batch_size = config.batch_size
trainset = image.create('mnist',
subset='train',
batch_size=batch_size,
transforms=transform)
del trainset["label"]
validset = image.create('mnist',
subset='test',
batch_size=batch_size,
transforms=transform)
del validset["label"]
res = AttrDict(trainset=trainset, validset=validset)
return res
def make_svhn(config):
"""Creates the svhn dataset."""
def to_float(x):
return tf.to_float(x) / 255.
transform = [to_float]
if config.canvas_size != 32:
transform.append(
functools.partial(preprocess.pad_and_shift,
output_size=config.canvas_size,
shift=None))
#transform.append(
# functools.partial(preprocess.normalized_sobel_edges))
batch_size = config.batch_size
res = AttrDict(trainset=image.create('svhn',
subset='train',
batch_size=batch_size,
transforms=transform),
validset=image.create('svhn',
subset='test',
batch_size=batch_size,
transforms=transform))
return res
def make_dataset256(config):
"""Creates the dataset_256 dataset."""
# data is created online, so there is no point in having
# a separate dataset for validation
def to_float(x):
return tf.to_float(x) / 255.
transform = [to_float]
transform.append(
functools.partial(preprocess.pad_and_shift,
output_size=config.canvas_size,
shift=None))
batch_size = config.batch_size
res = AttrDict(trainset=image.create('dataset256',
subset='train',
batch_size=batch_size,
transforms=transform),
validset=image.create('dataset256',
subset='test',
batch_size=batch_size,
transforms=transform))
return res
def make_clevr_veggies(config):
"""Creates the CLEVR Veggies dataset."""
def to_float(x):
return tf.to_float(x) / 255.
transform = [to_float]
batch_size = config.batch_size
res = AttrDict(trainset=image.create('clevr_veggies',
subset='train',
batch_size=batch_size,
transforms=transform),
validset=image.create('clevr_veggies',
subset='val',
batch_size=batch_size,
transforms=transform))
return res
