def test_ra(self):
dataset = CIFAR10(root=osp.join('../..', 'data', 'CIFAR10'),
train=False,
download=True,
transform=None)
idx = torch.randint(low=0, high=len(dataset), size=()).item() # 908
img, label = dataset[idx]
print(f'Image index: {idx}.')
print(f'Label: {label}.')
magnitude = 1.0
ra = RandAugment(n_augm_apply=1,
magnitude=magnitude,
augm_pool=RAND_AUGMENT_DEFAULT_POOL[2:3],
magnitude_policy='constant')
img_ra = ra(img)
img_tens = ToTensor()(img)
img_ra_tens = ToTensor()(img_ra)
print(img_tens[0:5, 0, 0])
print(img_ra_tens[0:5, 0, 0])
print('Img : ', img_tens.shape)
print('Img RA : ', img_ra_tens.shape)
if magnitude == 0:
self.assertTrue(img_tens.eq(img_ra_tens).all())
else:
self.assertTrue(img_tens.ne(img_ra_tens).any())
plt.figure()
plt.imshow(img)
plt.figure()
plt.imshow(img_ra)
plt.show(block=False)
def __init__(self, pil_transform: ImageTransform, p: float = 1.0):
"""
Class that convert PIL image to tensor internally for apply tensor transforms.
Tensors images will have the shape (width, height, 3).
:param pil_transform: The tensor transform to wrap.
:param p: The probability to apply the transform. (default: 1.0)
"""
super().__init__(
transform=pil_transform,
pre_convert=ToTensor(),
post_convert=ToPIL(mode=None),
p=p,
)
def __init__(self,
pil_transform: ImageTransform,
mode: Optional[str] = 'RGB',
p: float = 1.0):
"""
Class that convert tensor to PIL image internally for apply PIL transforms.
Tensors images must have the shape (width, height, 3).
:param pil_transform: The PIL transform to wrap.
:param mode: The PIL image mode of the image. (default: 'RGB')
:param p: The probability to apply the transform. (default: 1.0)
"""
super().__init__(
transform=pil_transform,
pre_convert=ToPIL(mode=mode),
post_convert=ToTensor(),
p=p,
)
def test_to_tensor(self):
to_tens_mlu = ToTensor()
to_tens_tvi = torchvision.transforms.ToTensor()
# Black image (width, height, 3)
data = Image.new('RGB', (32, 64), color='black')
other_mlu = to_tens_mlu(data)
other_tvi = to_tens_tvi(data)
self.assertEqual(
other_mlu.shape, other_tvi.shape,
f'Mismatch shapes for conversion {to_tens_mlu.__class__.__name__}(x) == {to_tens_tvi.__class__.__name__}(x)'
)
self.assertTrue(
other_mlu.eq(other_tvi).all(),
f'Mismatch values for conversion {to_tens_mlu.__class__.__name__}(x) == {to_tens_tvi.__class__.__name__}(x)'
)
def test_pil_conversions(self):
to_ten = ToTensor()
to_num = ToNumpy()
to_lis = ToList()
to_pil = ToPIL()
# Black image (width, height)
data = Image.new('RGB', (32, 64), color='black')
to_base = to_pil
for to in [to_ten, to_num, to_lis, to_pil]:
other = to_base(to(data))
self.assertEqual(
data.size, other.size,
f'Mismatch shapes for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
self.assertTrue(
data == other,
f'Mismatch values for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
def test_numpy_conversions(self):
# Black image (width, height, channel)
data = np.zeros((32, 64, 3))
to_ten = ToTensor()
to_num = ToNumpy()
to_lis = ToList()
to_pil = ToPIL()
to_base = to_num
for to in [to_ten, to_num, to_lis, to_pil]:
other = to_base(to(data))
self.assertEqual(
data.shape, other.shape,
f'Mismatch shapes for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
self.assertTrue((data == other).all(
), f'Mismatch values for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
def test_tensor_conversions(self):
# Black image (channel, height, width)
data = torch.zeros(3, 64, 32)
to_ten = ToTensor(device=data.device)
to_num = ToNumpy()
to_lis = ToList()
to_pil = ToPIL()
to_base = to_ten
for to in [to_ten, to_num, to_lis, to_pil]:
other = to_base(to(data))
self.assertEqual(
data.shape, other.shape,
f'Mismatch shapes for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
self.assertTrue(
data.eq(other).all(),
f'Mismatch values for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
def test_list_conversions(self):
# Black image (width, height, channel)
data = [[[0 for _ in range(3)] for _ in range(64)] for _ in range(32)]
to_ten = ToTensor()
to_num = ToNumpy()
to_lis = ToList()
to_pil = ToPIL()
to_base = to_lis
for to in [to_ten, to_num, to_lis, to_pil]:
other = to_base(to(data))
data_shape = _get_list_shape(data)
other_shape = _get_list_shape(other)
self.assertEqual(
data_shape, other_shape,
f'Mismatch shapes for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)
self.assertTrue(
data == other,
f'Mismatch values for conversion {to_base.__class__.__name__}({to.__class__.__name__}(x)) == x'
)