def test_five_crop(device):
script_five_crop = torch.jit.script(F.five_crop)
img_tensor, pil_img = _create_data(32, 34, device=device)
cropped_pil_images = F.five_crop(pil_img, [10, 11])
cropped_tensors = F.five_crop(img_tensor, [10, 11])
for i in range(5):
_assert_equal_tensor_to_pil(cropped_tensors[i], cropped_pil_images[i])
cropped_tensors = script_five_crop(img_tensor, [10, 11])
for i in range(5):
_assert_equal_tensor_to_pil(cropped_tensors[i], cropped_pil_images[i])
batch_tensors = _create_data_batch(16, 18, num_samples=4, device=device)
tuple_transformed_batches = F.five_crop(batch_tensors, [10, 11])
for i in range(len(batch_tensors)):
img_tensor = batch_tensors[i, ...]
tuple_transformed_imgs = F.five_crop(img_tensor, [10, 11])
assert len(tuple_transformed_imgs) == len(tuple_transformed_batches)
for j in range(len(tuple_transformed_imgs)):
true_transformed_img = tuple_transformed_imgs[j]
transformed_img = tuple_transformed_batches[j][i, ...]
assert_equal(true_transformed_img, transformed_img)
# scriptable function test
s_tuple_transformed_batches = script_five_crop(batch_tensors, [10, 11])
for transformed_batch, s_transformed_batch in zip(tuple_transformed_batches, s_tuple_transformed_batches):
assert_equal(transformed_batch, s_transformed_batch)
def test_five_crop(self):
script_five_crop = torch.jit.script(F.five_crop)
img_tensor, pil_img = self._create_data(32, 34, device=self.device)
cropped_pil_images = F.five_crop(pil_img, [10, 11])
cropped_tensors = F.five_crop(img_tensor, [10, 11])
for i in range(5):
self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])
cropped_tensors = script_five_crop(img_tensor, [10, 11])
for i in range(5):
self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])
batch_tensors = self._create_data_batch(16, 18, num_samples=4, device=self.device)
tuple_transformed_batches = F.five_crop(batch_tensors, [10, 11])
for i in range(len(batch_tensors)):
img_tensor = batch_tensors[i, ...]
tuple_transformed_imgs = F.five_crop(img_tensor, [10, 11])
self.assertEqual(len(tuple_transformed_imgs), len(tuple_transformed_batches))
for j in range(len(tuple_transformed_imgs)):
true_transformed_img = tuple_transformed_imgs[j]
transformed_img = tuple_transformed_batches[j][i, ...]
self.assertTrue(true_transformed_img.equal(transformed_img))
# scriptable function test
s_tuple_transformed_batches = script_five_crop(batch_tensors, [10, 11])
for transformed_batch, s_transformed_batch in zip(tuple_transformed_batches, s_tuple_transformed_batches):
self.assertTrue(transformed_batch.equal(s_transformed_batch))
def transforms(x, y):
x = list(tf.five_crop(x, 512))
y = list(tf.five_crop(y, 512))
for i in range(len(x)):
if random.random() > 0.5:
x[i] = tf.hflip(x[i])
y[i] = tf.hflip(y[i])
if random.random() > 0.5:
x[i] = tf.vflip(x[i])
y[i] = tf.vflip(y[i])
if random.random() > 0.5:
angle = random.randint(-30, 30)
x[i] = tf.rotate(x[i], angle)
y[i] = tf.rotate(y[i], angle)
x = torch.stack([tf.to_tensor(crop) for crop in x])
y = torch.stack([tf.to_tensor(crop) for crop in y])
# Replace all nonzero values with 1
y[y != 0] = 1
return x, y
def __call__(self, sample):
image = sample['image']
sample['image'] = F.five_crop(image, self.size)
result = []
list_image = F.five_crop(image, self.size)
for item in zip(list_image):
_sample = copy.deepcopy(sample)
_sample['image'] = item[0]
result.append(_sample)
return result
def test_five_crop(self):
script_five_crop = torch.jit.script(F.five_crop)
img_tensor, pil_img = self._create_data(32, 34, device=self.device)
cropped_pil_images = F.five_crop(pil_img, [10, 11])
cropped_tensors = F.five_crop(img_tensor, [10, 11])
for i in range(5):
self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])
cropped_tensors = script_five_crop(img_tensor, [10, 11])
for i in range(5):
self.compareTensorToPIL(cropped_tensors[i], cropped_pil_images[i])
def test_five_crop(self):
img_tensor = torch.randint(0, 255, (1, 32, 32), dtype=torch.uint8)
img_tensor_clone = img_tensor.clone()
cropped_tensor = F_t.five_crop(img_tensor, [10, 10])
cropped_pil_image = F.five_crop(transforms.ToPILImage()(img_tensor),
[10, 10])
self.assertTrue(
torch.equal(cropped_tensor[0],
(transforms.ToTensor()(cropped_pil_image[0]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[1],
(transforms.ToTensor()(cropped_pil_image[2]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[2],
(transforms.ToTensor()(cropped_pil_image[1]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[3],
(transforms.ToTensor()(cropped_pil_image[3]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[4],
(transforms.ToTensor()(cropped_pil_image[4]) * 255).to(
torch.uint8)))
self.assertTrue(torch.equal(img_tensor, img_tensor_clone))
def test_five_crop(self):
script_five_crop = torch.jit.script(F_t.five_crop)
img_tensor = torch.randint(0, 255, (1, 32, 32), dtype=torch.uint8)
img_tensor_clone = img_tensor.clone()
cropped_tensor = F_t.five_crop(img_tensor, [10, 10])
cropped_pil_image = F.five_crop(transforms.ToPILImage()(img_tensor),
[10, 10])
self.assertTrue(
torch.equal(cropped_tensor[0],
(transforms.ToTensor()(cropped_pil_image[0]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[1],
(transforms.ToTensor()(cropped_pil_image[2]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[2],
(transforms.ToTensor()(cropped_pil_image[1]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[3],
(transforms.ToTensor()(cropped_pil_image[3]) * 255).to(
torch.uint8)))
self.assertTrue(
torch.equal(cropped_tensor[4],
(transforms.ToTensor()(cropped_pil_image[4]) * 255).to(
torch.uint8)))
self.assertTrue(torch.equal(img_tensor, img_tensor_clone))
# scriptable function test
cropped_script = script_five_crop(img_tensor, [10, 10])
for cropped_script_img, cropped_tensor_img in zip(
cropped_script, cropped_tensor):
self.assertTrue(torch.equal(cropped_script_img,
cropped_tensor_img))
def transforms(img,
seg,
size=(360, 640),
augment=True,
hflip_prob=0.5,
five_crop_prob=0.5,
five_crop_scale=0.6,
rotate_prob=0.5,
max_rotate=30.0,
tensor_output=True,
normalize_mean=torch.Tensor([0.3518, 0.3932, 0.4011]),
normalize_std=torch.Tensor([0.2363, 0.2494, 0.2611]),
_ignore_index=255):
''' BDD transforms pipeline '''
if augment and random.random() < hflip_prob:
img = tfunc.hflip(img)
seg = tfunc.hflip(seg)
if augment and random.random() < five_crop_prob:
i = random.randint(0, 4)
resize = lambda x, scale: tuple(int(i * scale) for i in x)[::-1]
img = tfunc.five_crop(img, resize(img.size, five_crop_scale))[i]
seg = tfunc.five_crop(seg, resize(seg.size, five_crop_scale))[i]
if augment and random.random() < rotate_prob:
angle = random.randrange(-max_rotate, max_rotate)
# mask to track rotation and ignore newly added pixels
mask = Image.new('1', seg.size, (1, ))
img = tfunc.rotate(img, angle)
seg = tfunc.rotate(seg, angle)
mask = tfunc.rotate(mask, angle)
white = Image.new('L', seg.size, (_ignore_index))
seg = Image.composite(seg, white, mask=mask)
img = tfunc.resize(img, size, interpolation=Image.NEAREST)
seg = tfunc.resize(seg, size, interpolation=Image.NEAREST)
if tensor_output:
img = tfunc.to_tensor(img)
img = (img - normalize_mean[:, None, None]) / normalize_std[:, None,
None]
seg = torch.LongTensor(np.array(seg))
return img, seg
def __call__(self, sample):
A, B = sample['A'], sample['B']
sample['A'] = F.five_crop(A, self.size)
sample['B'] = F.five_crop(B, self.size)
result = []
list_A = F.five_crop(A, self.size)
list_B = F.five_crop(B, self.size)
for item in zip(list_A, list_B):
_sample = copy.deepcopy(sample)
_sample['A'] = item[0]
_sample['B'] = item[1]
result.append(_sample)
# item[0].show()
# item[1].show()
return result
def apply_random_transform(img: PIL.Image, disp, lbl):
# All available transforms:
# https://pytorch.org/docs/stable/torchvision/transforms.html
if rnd.random() > 0.2:
img = TF.hflip(img)
disp = TF.hflip(disp)
lbl = TF.hflip(lbl)
angle = rnd.random() * 30 - 15
translate = (0, 0)
scale = rnd.random() * 0.5 + 1
shear = rnd.random() * 10 - 5
# Ensures we scale at least enough to avoid black background from rotation.
scale = max(scale, rotation_scaling(img.size, angle))
img = TF.affine(img, angle, translate, scale, shear, resample=PIL.Image.BICUBIC)
disp = TF.affine(disp, angle, translate, scale, shear, resample=PIL.Image.NEAREST)
# Resampling must be nearest, anything else does not make sense for labels.
lbl = TF.affine(lbl, angle, translate, scale, shear, resample=PIL.Image.NEAREST)
img = TF.adjust_brightness(img, 0.75 + rnd.random() * 0.5)
img = TF.center_crop(img, output_size=(512, 640))
disp = TF.center_crop(disp, output_size=(512, 640))
lbl = TF.center_crop(lbl, output_size=(512, 640))
img_cropped = TF.five_crop(img, size=(512, 640))
disp_cropped = TF.five_crop(disp, size=(512, 640))
lbl_cropped = TF.five_crop(lbl, size=(512, 640))
imgs = [img]
imgs.extend(img_cropped)
disps = [disp]
disps.extend(disp_cropped)
lbls = [lbl]
lbls.extend(lbl_cropped)
idx = rnd.choice(range(len(imgs)))
return imgs[idx], disps[idx], lbls[idx]
def __call__(self, img):
resizes = [F.resize(img, size) for size in self.resizes]
all_crops = []
for crop in resizes:
all_crops += F.five_crop(crop, self.size)
# Also including a crop that includes the whole image, even if aspect ratio is not respected.
all_crops.append(F.resize(img, self.size))
for i, crop in enumerate(all_crops):
if random() > 0.5:
all_crops[i] = F.hflip(crop)
return tuple(all_crops)
def __call__(self, img, target=None, mask=None):
if target is not None and mask is None:
return F.five_crop(img, self.size), F.five_crop(target, self.size)
if target is not None and mask is not None:
return F.five_crop(img, self.size), F.five_crop(
target, self.size), F.five_crop(mask, self.size)
return F.five_crop(img, self.size)
def crop(img, target_size=512):
cutouts = TF.five_crop(img, target_size)
return [TF.to_tensor(img) for img in cutouts]
def __call__(self, img):
return F.five_crop(img, self.size)
def __call__(self, img, target=None):
if target is not None:
return F.five_crop(img, self.size), F.five_crop(target, self.size)
return F.five_crop(img, self.size)
def __call__(self, img):
# randomly return one of the five crops
return F.five_crop(img, self.size)[random.randint(0, 4)]
def __call__(self, img, mask):
return F.five_crop(img, self.size), F.five_crop(mask, self.size)
