def test_camvid_dataset_with_transforms(self):
from inferno.io.box.camvid import get_camvid_loaders
from inferno.utils.io_utils import print_tensor
train_loader, validate_loader, test_loader = get_camvid_loaders(
self.get_camvid_root())
train_dataset = train_loader.dataset
image, label = train_dataset[0]
# Make sure the shapes checkout
self.assertSequenceEqual(image.size(), (3, 360, 480))
self.assertSequenceEqual(label.size(), (360, 480))
self.assertEqual(image.type(), 'torch.FloatTensor')
self.assertEqual(label.type(), 'torch.LongTensor')
# Print tensors to make sure they look legit
if not exists(self.PLOT_DIRECTORY):
os.mkdir(self.PLOT_DIRECTORY)
else:
assert isdir(self.PLOT_DIRECTORY)
print_tensor(image.numpy()[None, ...],
prefix='IMG--',
directory=self.PLOT_DIRECTORY)
print_tensor(label.numpy()[None, None, ...],
prefix='LAB--',
directory=self.PLOT_DIRECTORY)
print("[+] Inspect images at {}".format(self.PLOT_DIRECTORY))
def test_master_affinities_multi_order(self):
data_config = self.DATA_CONFIG.format('affinities_multi_order')
from neurofire.datasets.cremi.loaders.master import CREMIDatasets
from inferno.utils.io_utils import print_tensor
dataset = CREMIDatasets.from_config(data_config)
# Get from dataset
batch = dataset[0]
# Validate
self.assertEqual(len(batch), 2)
self.assertEqual(list(batch[0].size()), [1, 5, 512, 512])
self.assertEqual(list(batch[1].size())[1:], [5, 512, 512])
self.assertIn(batch[1].size(0),
[2 * 4, 3 * 4, (2 * 4) + 1, (3 * 4) + 1])
# Print to file
if os.path.exists(self.PLOT_DIRECTORY):
assert os.path.isdir(self.PLOT_DIRECTORY)
else:
os.mkdir(self.PLOT_DIRECTORY)
print_tensor(tensor=batch[0].numpy()[None, ...],
prefix='MOAFFRAW',
directory=self.PLOT_DIRECTORY)
print_tensor(tensor=batch[1].numpy()[None, ...],
prefix='MOAFFTAR',
directory=self.PLOT_DIRECTORY)
print("Plots printed to {}.".format(self.PLOT_DIRECTORY))
def _test_camvid_dataset_with_transforms(self):
from inferno.io.box.camvid import CamVid
from inferno.io.transform.base import Compose
from inferno.io.transform.image import PILImage2NumPyArray, RandomSizedCrop, Scale
from inferno.utils.io_utils import print_tensor
camvid = CamVid(self.get_camvid_root(),
image_transform=Compose(),
label_transform=Compose(),
joint_transform=Compose())
camvid.image_transform.add(PILImage2NumPyArray())
camvid.label_transform.add(PILImage2NumPyArray())
image, label = camvid[0]
self.assertSequenceEqual(image.shape, (3, 360, 480))
self.assertSequenceEqual(label.shape, (360, 480))
# Add crop trafo
camvid.joint_transform.add(RandomSizedCrop(ratio_between=(0.7, 1.0),
preserve_aspect_ratio=True))
# We need 2 scale transforms, one with order 3 (image) and the other with order 0 (label)
camvid.joint_transform.add(Scale(output_image_shape=(360, 480),
interpolation_order=3, apply_to=[0]))
camvid.joint_transform.add(Scale(output_image_shape=(360, 480),
interpolation_order=0, apply_to=[1]))
image, label = camvid[0]
self.assertSequenceEqual(image.shape, (3, 360, 480))
self.assertSequenceEqual(label.shape, (360, 480))
self.assertLessEqual(len(np.unique(label)), 12)
# Print tensors to make sure they look legit
if not exists(self.PLOT_DIRECTORY):
os.mkdir(self.PLOT_DIRECTORY)
else:
assert isdir(self.PLOT_DIRECTORY)
print_tensor(image[None, ...], prefix='IMG--', directory=self.PLOT_DIRECTORY)
print_tensor(label[None, None, ...], prefix='LAB--', directory=self.PLOT_DIRECTORY)
print("[+] Inspect images at {}".format(self.PLOT_DIRECTORY))
def test_cityscapes_dataset_with_transforms_unzipped(self):
from inferno.io.box.cityscapes import get_cityscapes_loaders
from inferno.utils.io_utils import print_tensor
train_loader, validate_loader = get_cityscapes_loaders(
join(self.get_cityscapes_root(), 'extracted'),
include_coarse_dataset=self.INCLUDE_COARSE,
read_from_zip_archive=False)
train_dataset = train_loader.dataset
tic = time.time()
image, label = train_dataset[0]
toc = time.time()
print("[+] Loaded sample in {} seconds.".format(toc - tic))
# Make sure the shapes checkout
self.assertSequenceEqual(image.size(), (3, 1024, 2048))
self.assertSequenceEqual(label.size(), (1024, 2048))
self.assertEqual(image.type(), 'torch.FloatTensor')
self.assertEqual(label.type(), 'torch.LongTensor')
# Print tensors to make sure they look legit
if not exists(self.PLOT_DIRECTORY):
os.mkdir(self.PLOT_DIRECTORY)
else:
assert isdir(self.PLOT_DIRECTORY)
print_tensor(image.numpy()[None, ...],
prefix='IMG--',
directory=self.PLOT_DIRECTORY)
for class_id in np.unique(label.numpy()):
print_tensor((label.numpy()[None, None,
...] == class_id).astype('float32'),
prefix='LAB-{}--'.format(class_id),
directory=self.PLOT_DIRECTORY)
print_tensor(label.numpy()[None, None, ...],
prefix='LAB--',
directory=self.PLOT_DIRECTORY)
print("[+] Inspect images at {}".format(self.PLOT_DIRECTORY))
def _test_master_membranes_or_affinities(self, data_config):
from neurofire.datasets.cremi.loaders.master import CREMIDatasets
from inferno.utils.io_utils import print_tensor
dataset = CREMIDatasets.from_config(data_config)
# Get from dataset
batch = dataset[0]
# Validate
self.assertEqual(len(batch), 2)
for _batch in batch:
self.assertEqual(list(_batch.size()), [1, 3, 512, 512])
# Print to file
if os.path.exists(self.PLOT_DIRECTORY):
assert os.path.isdir(self.PLOT_DIRECTORY)
else:
os.mkdir(self.PLOT_DIRECTORY)
print_tensor(tensor=batch[0].numpy()[None, ...],
prefix='RAW',
directory=self.PLOT_DIRECTORY)
print_tensor(tensor=batch[1].numpy()[None, ...],
prefix='TAR',
directory=self.PLOT_DIRECTORY)
print("Plots printed to {}.".format(self.PLOT_DIRECTORY))
def test_master(self):
from neurofire.datasets.isbi2012.loaders.master import ISBI2012Dataset
from inferno.utils.io_utils import print_tensor
dataset = ISBI2012Dataset.from_config(self.DATA_CONFIG)
# Get from dataset
batch = dataset[0]
# Validate
self.assertEqual(len(batch), 2)
for _batch in batch:
self.assertEqual(list(_batch.size()),
[1, 576, 576]) # batch axis added by loader
# Print to file
if os.path.exists(self.PLOT_DIRECTORY):
assert os.path.isdir(self.PLOT_DIRECTORY)
else:
os.mkdir(self.PLOT_DIRECTORY)
print_tensor(tensor=batch[0].numpy()[None, ...],
prefix='RAW',
directory=self.PLOT_DIRECTORY)
print_tensor(tensor=batch[1].numpy()[None, ...],
prefix='MEM',
directory=self.PLOT_DIRECTORY)
print("Plots printed to {}.".format(self.PLOT_DIRECTORY))