def apply_to_mask(self, mask, selected_object, bbox, **params):
if selected_object is None:
if self.always_resize:
mask = F.resize(mask,
height=self.height,
width=self.width,
interpolation=cv2.INTER_NEAREST)
return mask
rmin, rmax, cmin, cmax = bbox
mask = mask[rmin:rmax + 1, cmin:cmax + 1]
if isinstance(selected_object, tuple):
layer_indx, mask_id = selected_object
obj_mask = mask[:, :, layer_indx] == mask_id
new_mask = np.zeros_like(mask)
new_mask[:, :, layer_indx][obj_mask] = mask_id
else:
obj_mask = mask == selected_object
new_mask = mask.copy()
new_mask[np.logical_not(obj_mask)] = 0
new_mask = F.resize(new_mask,
height=self.height,
width=self.width,
interpolation=cv2.INTER_NEAREST)
return new_mask
def apply(self, image, **params):
height, width, _ = image.shape
if width > 320 or height > 320:
scale = random.random() * 0.4 + 0.2
image = functional.resize(image, height=int(height * scale), width=int(width * scale), interpolation=self.interpolation)
return functional.resize(image, height=height, width=width, interpolation=self.interpolation)
else:
return image
def apply(self, img, selected_object, bbox, **params):
if selected_object is None:
if self.always_resize:
img = F.resize(img, height=self.height, width=self.width)
return img
rmin, rmax, cmin, cmax = bbox
img = img[rmin:rmax + 1, cmin:cmax + 1]
img = F.resize(img, height=self.height, width=self.width)
return img
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
self.img_width = img.shape[1]
self.img_height = img.shape[0]
return af.resize(img,
height=self.height,
width=self.width,
interpolation=interpolation)
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
new_height, new_width = _compute_new_shape(self.length, params["rows"],
params["cols"])
return F.resize(img,
height=new_height,
width=new_width,
interpolation=interpolation)
def predict(model, from_file_names, batch_size, to_path, img_transform):
loader = DataLoader(dataset=SaltDataset(from_file_names,
transform=img_transform,
mode='predict'),
shuffle=False,
batch_size=batch_size,
num_workers=args.workers,
pin_memory=torch.cuda.is_available())
with torch.no_grad():
for _, (inputs, paths) in enumerate(tqdm(loader, desc='Predict')):
inputs = utils.cuda(inputs)
outputs = model(inputs)
for i, _ in enumerate(paths):
factor = prepare_data.binary_factor
t_mask = (F.sigmoid(outputs[i, 0]).data.cpu().numpy() *
factor).astype(np.uint8)
final_mask = center_crop(t_mask, 202, 202)
final_mask = resize(final_mask,
original_height,
original_width,
interpolation=cv2.INTER_AREA)
to_path.mkdir(exist_ok=True, parents=True)
cv2.imwrite(str(to_path / (Path(paths[i]).stem + '.png')),
final_mask)
def albumentations(self, img):
img = albumentations.random_crop(img,
crop_height=64,
crop_width=64,
h_start=0,
w_start=0)
return albumentations.resize(img, height=512, width=512)
def fixed_size_rotate_crop(
img,
angle,
interpolation=cv2.INTER_LINEAR,
border_mode=cv2.BORDER_REFLECT_101,
value=None,
):
original_height, original_width = img.shape[:2]
rot_img = rotate_no_crop(img, angle, interpolation, border_mode, value)
new_height, new_width = rotatedRectWithMaxArea(original_height,
original_width, angle)
cropped_img = F.center_crop(rot_img, new_height, new_width)
resize_height, resize_width = get_resized_max_ratio(
original_height, original_width, new_height, new_width)
resized_img = F.resize(
cropped_img,
height=resize_height,
width=resize_width,
interpolation=interpolation,
)
return F.center_crop(resized_img, original_height, original_width)
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
a, b, *_ = img.shape
alpha = float(self.max_size) / max(a, b)
n_height, n_width = int(a * alpha), int(b * alpha)
return F.resize(img,
height=n_height,
width=n_width,
interpolation=interpolation)
def test_resize_nearest_interpolation(target):
img = np.array([[1, 1, 1, 1], [2, 2, 2, 2], [3, 3, 3, 3], [4, 4, 4, 4]], dtype=np.uint8)
expected = np.array([[1, 1], [3, 3]], dtype=np.uint8)
img, expected = convert_2d_to_target_format([img, expected], target=target)
resized_img = F.resize(img, 2, 2, interpolation=cv2.INTER_NEAREST)
height, width = resized_img.shape[:2]
assert height == 2
assert width == 2
assert np.array_equal(resized_img, expected)
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
height, width = img.shape[:2]
new_height, new_width = self._compute_new_hw(height, width)
return F.resize(img,
height=new_height,
width=new_width,
interpolation=interpolation)
def test_resize_different_height_and_width(target):
img = np.ones((100, 100), dtype=np.uint8)
img = convert_2d_to_target_format([img], target=target)
resized_img = F.resize(img, height=20, width=30)
height, width = resized_img.shape[:2]
assert height == 20
assert width == 30
if target == 'image':
num_channels = resized_img.shape[2]
assert num_channels == 3
def apply(self,
img,
crop_height=0,
crop_width=0,
h_start=0,
w_start=0,
interpolation=cv2.INTER_LINEAR,
**params):
img = img[h_start:h_start + crop_height, w_start:w_start + crop_width]
return F.resize(img, self.height, self.width, interpolation)
def apply(self,
img,
new_height=0,
new_width=0,
interpolation=cv2.INTER_LINEAR,
**params):
return F.resize(img,
height=new_height,
width=new_width,
interpolation=interpolation)
def test_resize_default_interpolation_float(target):
img = np.array([[0.1, 0.1, 0.1, 0.1], [0.2, 0.2, 0.2, 0.2],
[0.3, 0.3, 0.3, 0.3], [0.4, 0.4, 0.4, 0.4]],
dtype=np.float32)
expected = np.array([[0.15, 0.15], [0.35, 0.35]], dtype=np.float32)
img, expected = convert_2d_to_target_format([img, expected], target=target)
resized_img = F.resize(img, 2, 2)
height, width = resized_img.shape[:2]
assert height == 2
assert width == 2
assert_array_almost_equal_nulp(resized_img, expected)
def test_resize_nearest_interpolation_float(target):
img = np.array([[0.1, 0.1, 0.1, 0.1], [0.2, 0.2, 0.2, 0.2],
[0.3, 0.3, 0.3, 0.3], [0.4, 0.4, 0.4, 0.4]],
dtype=np.float32)
expected = np.array([[0.1, 0.1], [0.3, 0.3]], dtype=np.float32)
img, expected = convert_2d_to_target_format([img, expected], target=target)
resized_img = F.resize(img, 2, 2, interpolation=cv2.INTER_NEAREST)
height, width = resized_img.shape[:2]
assert height == 2
assert width == 2
assert np.array_equal(resized_img, expected)
def random_resize_and_stretch(self, img, new_width_range, stretch_limit = 0):
new_width_range = T.to_tuple(new_width_range)
stretch_limit = T.to_tuple(stretch_limit, bias=1)
new_sz = int(random.uniform(new_width_range[0], new_width_range[1]))
stretch = random.uniform(stretch_limit[0], stretch_limit[1])
img_max_sz = img.shape[1] #max(img.shape[0]*stretch, img.shape[1]) #img.shape[1] # GVNC - now it is resizing to max
new_width = int(img.shape[1]*new_sz/img_max_sz)
new_width = ((new_width+31)//32)*32
new_height = int(img.shape[0]*stretch*new_sz/img_max_sz)
new_height = ((new_height+31)//32)*32
return albu_f.resize(img, height=new_height, width=new_width, interpolation=cv2.INTER_LINEAR)
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
if self.output_size == -1:
return img
input_shape = img.shape
# We want X/Y = x/y and we have size = x*y so :
ratio = input_shape[1] / input_shape[0]
new_height = int(np.sqrt(self.output_size / ratio))
new_width = int(self.output_size / new_height)
return F.resize(
img, height=new_height, width=new_width, interpolation=interpolation
)
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
height, width, _ = img.shape
target_width = width * self.height // height
if self.min_width:
target_width = max(target_width, self.min_width)
if self.max_width:
target_width = min(target_width, self.max_width)
return F.resize(img,
height=self.height,
width=target_width,
interpolation=interpolation)
def apply(self,
img,
height_scale=1.0,
width_scale=1.0,
h_start=0,
w_start=0,
interpolation=cv2.INTER_LINEAR,
pad_value=0,
pad_loc_seed=None,
**params):
img_height, img_width = img.shape[:2]
crop_height, crop_width = int(img_height * height_scale), int(
img_width * width_scale)
crop = self.random_crop(img, crop_height, crop_width, h_start, w_start,
pad_value, pad_loc_seed)
return F.resize(crop, img_height, img_width, interpolation)
def __call__(self,
force_apply=True,
**data) -> Tuple[torch.Tensor, torch.Tensor]:
image, mask = data["image"], data["mask"]
new_size = np.round(
np.array(image.shape[:2]) / self.downsampling_factor).astype(
np.int64)
image = F.resize(
image,
height=new_size[0],
width=new_size[1],
interpolation=self.interpolation,
)
# create output dictionary
data["image"], data["mask"] = image, mask
return data
def apply(self, img, **params):
h, w = img.shape[:2]
nh = int(h / self.rf) * self.rf
nw = int(w / self.rf) * self.rf
interpolation = cv2.INTER_LINEAR #cv2.INTER_NEAREST
return AF.resize(img, nh, nw, interpolation)
def albumentations(self, img):
return albumentations.resize(img, height=512, width=512)
def __call__(self, image: np.ndarray) -> np.ndarray:
image = F.resize(image, height=self.height, width=self.width)
color_features = self.color_spaces(image)
features = self.extract_global_features(color_features)
return features
def random_stretch_image(img, **params):
# return functional.resize(img, (img.height, int(img.width * random.uniform(0.95, 1.05))))
height, width, _ = img.shape
return AF.resize(img, height,
int(width * random.uniform(0.95, 1.05)))
def rescale(img, scale: float, interpolation: int = cv2.INTER_AREA):
h, w = img.shape[:2]
h2, w2 = int(np.round(h * scale)), int(np.round(w * scale))
return AF.resize(img, h2, w2, interpolation)
def apply(self, img, interpolation=cv2.INTER_LINEAR, **params):
# print('apply, ', params)
return F.resize(img,
height=params['height'],
width=params['width'],
interpolation=interpolation)
def resize_from_any(image: np.ndarray) -> np.ndarray:
return resize(image, 101, 101)
def crop_resize_from_256(image: np.ndarray) -> np.ndarray:
cropped = center_crop(image, 202, 202)
return resize(cropped, 101, 101)
def preprocessing(image: np.array) -> torch.Tensor:
height, width = 640, 640
image = resize(image, height, width)
image = torch_tools.imagenet_normalize(image)
return to_tensor(image).squeeze().unsqueeze(0)
