def test_nan_to_int(shape, label_idx):
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.NanToInt(5)])
sample["labels"][label_idx] = np.nan
sample = transform(sample)
assert sample["labels"][label_idx] == 5
def test_compose(shape):
sample = create_input(shape)
transform = cpt.Compose(
[tvt.RandomHorizontalFlip(),
tvt.ToTensor(),
cpt.RandomGaussianBlur()])
sample = transform(sample)
assert sample["image"] is not None
def test_remap_label(shape, label_idx, start_label, end_label):
sample = create_input(shape)
transform = cpt.Compose(
[tvt.ToTensor(),
cpt.RemapLabel(start_label, end_label)])
sample["labels"][label_idx] = start_label
sample = transform(sample)
assert sample["labels"][label_idx] == end_label
def test_add_noise(shape):
"""Test this to guard against an implementation change."""
seed = 456
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.AddGaussianNoise(p=1)])
# run the custom noise
np.random.seed(seed)
image = tvt.functional.to_tensor(sample["image"]) * 1
np.random.uniform()
snr_level = np.random.uniform(
low=transform.transforms[1].snr_range[0],
high=transform.transforms[1].snr_range[1],
)
signal_level = np.mean(image.numpy())
image = image + (signal_level / snr_level) * torch.tensor(
np.random.normal(size=tuple(image.shape)),
dtype=image.dtype,
)
# transform blur
np.random.seed(seed)
sample = transform(sample)
assert torch.allclose(sample["image"], image)
# retest for 0 probability
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.AddGaussianNoise(p=-0.1)])
# run the custom blur
image = tvt.functional.to_tensor(sample["image"]) * 1
# transform blur
sample = transform(sample)
assert torch.allclose(sample["image"], image)
def test_rand_gauss_blur(shape):
"""Test this to guard against an implementation change."""
seed = 123
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.RandomGaussianBlur(p=1)])
# run the custom blur
np.random.seed(seed)
image = tvt.functional.to_tensor(sample["image"]) * 1
sigma = np.random.uniform(transform.transforms[1].sigma_range[0],
transform.transforms[1].sigma_range[1])
image = torch.tensor(
gaussian_filter(image.numpy(), sigma),
dtype=image.dtype,
)
# transform blur
transform = cpt.Compose([
tvt.ToTensor(),
cpt.RandomGaussianBlur(p=1, sigma_range=(sigma, sigma))
])
sample = transform(sample)
assert torch.allclose(sample["image"], image)
# retest for 0 probability
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.RandomGaussianBlur(p=-0.1)])
# run the custom blur
image = tvt.functional.to_tensor(sample["image"]) * 1
# transform blur
sample = transform(sample)
assert torch.allclose(sample["image"], image)
def test_tensor_to_rgb(shape):
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.TensorToRGB()])
image = tvt.functional.to_tensor(sample["image"]) * 1
expands = list()
for i in range(image.ndim):
if i == 0:
expands.append(3)
else:
expands.append(-1)
image = image.expand(*expands)
sample = transform(sample)
assert torch.allclose(sample["image"], image)
def test_histnorm(shape):
"""Test this to guard against an implementation change."""
sample = create_input(shape)
transform = cpt.Compose([tvt.ToTensor(), cpt.HistogramNormalize()])
image = np.transpose(
torch.tensor(np.array(sample["image"]), dtype=torch.float).numpy(),
(2, 0, 1))
# get image histogram
image_histogram, bins = np.histogram(image.flatten(),
transform.transforms[1].number_bins,
density=True)
cdf = image_histogram.cumsum() # cumulative distribution function
cdf = 255 * cdf / cdf[-1] # normalize
# use linear interpolation of cdf to find new pixel values
image_equalized = np.interp(image.flatten(), bins[:-1], cdf)
image_equalized.reshape(image.shape)
image = torch.tensor(image_equalized.reshape(image.shape)).to(torch.float)
sample = transform(sample)
assert torch.allclose(sample["image"], image)