def test_nested_invert(self):
tf = T.Compose([
T.Compose([
T.ToPILImage(),
T.RandomHorizontalFlip(0),
T.RandomHorizontalFlip(1),
]),
])
img_pil = tf(self.img_tensor)
self.assertIsInstance(img_pil, Image.Image)
self.assertIsInstance(tf.invert(img_pil), torch.Tensor)
def test_replay(self):
tf = T.Compose([
T.RandomCrop(self.crop_size),
T.RandomVerticalFlip(),
T.RandomHorizontalFlip(),
T.ToTensor(),
])
img_tf1 = tf.replay(self.img_pil)
img_tf2 = tf.replay(self.img_pil)
self.assertTrue(torch.allclose(img_tf1, img_tf2))
def test_invert(self):
tf = T.RandomOrder([
invertransforms.crop_pad.CenterCrop(self.crop_size),
T.RandomHorizontalFlip(1),
T.RandomVerticalFlip(1),
])
img_inv = tf(self.img_pil)
img_pil = tf.invert(img_inv)
self.assertEqual(img_pil.size, self.img_pil.size)
center = (self.w // 2 - 10, self.h // 2 + 10)
self.assertEqual(img_pil.getpixel(center),
self.img_pil.getpixel(center))
def test_track(self):
tf = T.Compose([
T.ToPILImage(),
T.RandomVerticalFlip(p=0.5),
# the crop will include the center pixels
T.RandomCrop(size=tuple(
int(0.8 * self.img_size[i]) for i in range(2))),
T.RandomHorizontalFlip(p=0.5),
T.ToTensor(),
])
imgs_tf = [tf.track(self.img_tensor) for _ in range(10)]
for i, img_tf in enumerate(imgs_tf):
n = min(self.img_size) // 10
center_pixels = (0, ) + tuple(
slice(self.img_size[i] // 2 - n, self.img_size[i] // 2 + n)
for i in range(2))
self.assertTrue(
torch.allclose(tf[i](img_tf)[center_pixels],
self.img_tensor[center_pixels]))
def test_invert_before_apply(self):
with self.assertRaises(InvertibleError):
T.RandomVerticalFlip().inverse()
with self.assertRaises(InvertibleError):
T.RandomHorizontalFlip().inverse()