def __getitem__(self, idx: int) -> dict:
####
result = {}
image_path = self.images[idx]
image = utils.imread(image_path)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if np.shape(image)[1] != 512:
image = cv2.resize(image, (512, 512))
if self.masks is not None:
mask = utils.imread(self.masks[idx])
if np.shape(mask)[1] != 512:
mask = cv2.resize(mask, (512, 512))
# extract certain classes from mask (e.g. cars)
mask_kidney = mask[:, :, 0] == 255
mask_tumor = mask[:, :, 1] == 255
masks = [mask_kidney, mask_tumor]
mask = np.stack(masks, axis=-1).astype('float')
result["mask"] = mask
result["image"] = image
if self.transforms is not None:
result = self.transforms(**result)
result["filename"] = image_path.name
return result
def preprocess(self, image_path: Path):
try:
if self.extension in ("jpg", "JPG", "jpeg", "JPEG"):
image = np.array(
imread(uri=image_path,
grayscale=self.grayscale,
expand_dims=self.expand_dims,
exifrotate=not self.clear_exif))
else: # imread does not have exifrotate for non-jpeg type
image = np.array(
imread(uri=image_path,
grayscale=self.grayscale,
expand_dims=self.expand_dims))
except Exception as e:
print(f"Cannot read file {image_path}, exception: {e}")
return
if self.max_size is not None:
image = longest_max_size(image, self.max_size, self.interpolation)
target_path = self.out_dir / image_path.relative_to(self.in_dir)
target_path.parent.mkdir(parents=True, exist_ok=True)
image = image.clip(0, 255).round().astype(np.uint8)
imwrite(target_path, image)
def show(index: int,
images: List[Path],
masks: List[Path],
transforms=None) -> None:
image_path = images[index]
name = image_path.name
image = utils.imread(image_path)
mask = utils.imread(masks[index])
if transforms is not None:
temp = transforms(image=image, mask=mask)
image = temp["image"]
mask = temp["mask"]
show_examples(name, image, mask)
def preprocess(self, filepath: Path) -> None:
"""Process one file."""
try:
image = np.array(utils.imread(filepath))
except Exception as e:
logger.warning(f"Cannot read file {filepath}, exception: {e}")
return
filename, extention = filepath.stem, filepath.suffix
out_dir = (self.out_dir / filepath.relative_to(self.in_dir)).parent
out_dir.mkdir(parents=True, exist_ok=True)
patches = cut_with_overlap(
image=image.clip(0, 255).round().astype(np.uint8),
patch_height=self.patch_height,
patch_width=self.patch_width,
height_overlap=self.height_overlap,
width_overlap=self.width_overlap,
min_height=self.min_height,
min_width=self.min_width,
)
for index, patch in enumerate(patches):
out_path = out_dir / f"{filename}_{index}{extention}"
utils.imwrite(uri=out_path, im=patch)
def __getitem__(self, idx: int) -> dict:
image_path = self.images[idx]
name = image_path.name
image = utils.imread(image_path)
mask = None
if self.transforms is not None:
if self.masks is not None:
mask = imread(self.mask_path / f"{os.path.splitext(name)[0]}.png")
transformed = self.transforms(image=image, mask=mask)
image_tf = transformed["image"]
mask_tf = transformed["mask"]
result = {"image": image_tf, "mask": mask_tf}
else:
transformed = self.transforms(
image=image,
)
image_tf = transformed["image"]
result = {"image": image_tf}
else:
result = {"image": image, "mask": mask}
result["filename_img"] = image_path.name
result["filename_mask"] = f"{os.path.splitext(name)[0]}.png"
return result
def preprocess(self, image_path: Path):
"""@TODO: Docs. Contribution is welcome."""
try:
_, extension = os.path.splitext(image_path)
kwargs = {
"grayscale": self.grayscale,
"expand_dims": self.expand_dims,
}
if extension.lower() in {"jpg", "jpeg"}:
# imread does not have exifrotate for non-jpeg type
kwargs["exifrotate"] = not self.clear_exif
image = np.array(imread(uri=image_path, **kwargs))
except Exception as e:
print(f"Cannot read file {image_path}, exception: {e}")
return
if self.max_size is not None:
image = longest_max_size(image, self.max_size, self.interpolation)
target_path = self.out_dir / image_path.relative_to(self.in_dir)
target_path.parent.mkdir(parents=True, exist_ok=True)
image = image.clip(0, 255).round().astype(np.uint8)
imwrite(target_path, image)
def save_image(image, mask, path, orig_shape):
mask_path = f"{PurePath(path).parent}/{PurePath(path).stem}_mask.png"
colorize_mask(mask, palete).save(mask_path)
image = albu.Resize(512, 512)(image=image)['image']
mask = imread(mask_path)
out_im = image // 2 + mask // 2
io.imsave(path, albu.Resize(*orig_shape)(image=out_im)['image'])
return mask_path
def read_random_images(data,
transforms=None) -> List[Tuple[str, np.ndarray]]:
data_ = np.random.choice(data, size=4)
result = []
for d in data_:
image = imread(d['filepath'])
if transforms is not None:
image = transforms(image=image)['image']
result.append((d['class'], image))
return result
def test_imread():
"""Tests ``imread`` functionality."""
jpg_rgb_uri = (
"https://raw.githubusercontent.com/catalyst-team/catalyst-pics/master"
"/test_images/catalyst_icon.jpg")
jpg_grs_uri = (
"https://raw.githubusercontent.com/catalyst-team/catalyst-pics/master"
"/test_images/catalyst_icon_grayscale.jpg")
png_rgb_uri = (
"https://raw.githubusercontent.com/catalyst-team/catalyst-pics/master"
"/test_images/catalyst_icon.png")
png_grs_uri = (
"https://raw.githubusercontent.com/catalyst-team/catalyst-pics/master"
"/test_images/catalyst_icon_grayscale.png")
for uri in [jpg_rgb_uri, jpg_grs_uri, png_rgb_uri, png_grs_uri]:
img = utils.imread(uri)
assert img.shape == (400, 400, 3)
img = utils.imread(uri, grayscale=True)
assert img.shape == (400, 400, 1)
def show_random_transform(transforms):
# transforms = compose([resize_transforms(), hard_transforms(), post_transforms()])
length = len(ALL_IMAGES)
index = random.randint(0, length - 1)
IMG = ALL_IMAGES[index]
plt.figure(figsize=(20, 28))
plt.subplot(1, 2, 1)
plt.imshow(utils.imread(IMG))
plt.title(f"Image: {IMG}")
transformed = transforms(image=utils.imread(IMG))
transformed_image = transformed["image"]
TRANSFORMED = np.moveaxis((np.array(transformed_image) * 255).astype("uint8"), 0, 2)
plt.subplot(1, 2, 2)
plt.imshow(Image.fromarray(TRANSFORMED))
plt.title(f"Transformed")
Image.fromarray(TRANSFORMED).save("prova.png")
def show(index: int, images: List[Path], masks: List[Path], transforms=None) -> None:
image_path = images[index]
name = image_path.name
image = utils.imread(image_path)
mask = imread(train_mask_path / f"{os.path.splitext(name)[0]}_mask.png")
# mask = utils.imread(masks[index])
if transforms is not None:
temp = transforms(image=image, mask=mask)
image = temp["image"]
mask = temp["mask"]
show_examples(name, image, mask)
def preprocess(self, image_path: Path):
image = imread(image_path, rootpath=str(self.in_dir))
heigth, width = image.shape[:2]
mask = np.zeros((heigth, width, len(self.index2color)), dtype=np.uint8)
for index in range(len(self.index2color)):
mask[np.all((image == self.index2color[index]), axis=-1),
index] = 255
target_path = self.out_dir / f"{id_from_fname(image_path)}.tiff"
target_path.parent.mkdir(parents=True, exist_ok=True)
mimwrite_with_meta(target_path, np.dsplit(mask, mask.shape[2]),
{"compress": 9})
def compute_segmenation_weight(img_path, img_width, img_height, contrast=1):
"""
Function calculate weight matrix for content loss using fully
convolutional neural network
Parameters:
----------
img_path : str
Path to input picture
img_width : int
Width of image
img_height :int
Height of image
contrast: int
Contrast coefficient. Default value equal to one.
Bigger value means, that front object will have bigger values in weight matrix.
"""
default_image_size = 320
def pre_transforms(image_size=default_image_size):
return [albu.Resize(image_size, image_size, p=1)]
def post_transforms():
# we use ImageNet image normalization
# and convert it to torch.Tensor
return [albu.Normalize(), ToTensor()]
def compose(transforms_to_compose):
# combine all augmentations into one single pipeline
result = albu.Compose(
[item for sublist in transforms_to_compose for item in sublist])
return result
valid_transforms = compose([pre_transforms(), post_transforms()])
model = smp.FPN(encoder_name="resnext101_32x8d", classes=1)
model.load_state_dict(torch.load(f'./models/model.pth'))
model.eval()
image = {'image': utils.imread(img_path)}
x = valid_transforms(**image)['image']
y = model.predict(x.view(1, 3, 320, 320))[0, 0].sigmoid().numpy()
w = resize(y, (img_height, img_width), preserve_range=True)
w = w / w.sum() * img_width * img_height
wn = np.clip(contrast * (w - w.mean()) + w.mean(), 1e-1,
w.sum())[np.newaxis, np.newaxis, :, :]
wn = torch.from_numpy(wn).cuda().float()
del model, x, y, w
return wn
def prepare(self):
image_path = self.image
scale = 1 / self.downscale_factor
interpolation = self.interpolation
img_input_target = imread(image_path)
img_input = cv2.resize(
img_input_target,
None,
None,
fx=scale,
fy=scale,
interpolation=default_interpolations[interpolation])
return img_input, img_input_target
def __getitem__(self, idx: int) -> dict:
image_path = self.images[idx]
name = image_path.name
image = utils.imread(image_path)
result = {"image": image}
if self.masks is not None:
mask = imread(train_mask_path / f"{os.path.splitext(name)[0]}_mask.png")
result["mask"] = mask
if self.transforms is not None:
result = self.transforms(**result)
result["filename"] = image_path.name
return result
def __call__(self, element):
"""Reads a row from your annotations dict with filename and
transfer it to an image
Args:
element: elem in your dataset
Returns:
np.ndarray: Image
"""
image_name = str(element[self.input_key])
img = utils.imread(
image_name, rootpath=self.rootpath, grayscale=self.grayscale
)
output = {self.output_key: img}
return output
def __getitem__(self, idx: int) -> dict:
"""Main method"""
image_path = self.images[idx]
image = utils.imread(image_path)
result = {"image": image}
if self.masks is not None:
mask = gif_imread(self.masks[idx])
result["mask"] = mask
if self.transforms is not None:
result = self.transforms(result)
result["filename"] = image_path.name
return result
def __call__(self, row):
"""Reads a row from your annotations dict with filename and
transfer it to an image
Args:
row: elem in your dataset.
Returns:
np.ndarray: Image
"""
image_name = str(row[self.input_key])
img = imread(image_name,
rootpath=self.datapath,
grayscale=self.grayscale)
result = {self.output_key: img}
return result
def preprocess(self, image_path: Path):
try:
image = np.array(
imread(uri=image_path,
grayscale=self.grayscale,
expand_dims=self.expand_dims,
exifrotate=not self.clear_exif))
except Exception:
return
if self.max_size is not None:
image = longest_max_size(image, self.max_size, self.interpolation)
target_path = self.out_dir / image_path.relative_to(self.in_dir)
target_path.parent.mkdir(parents=True, exist_ok=True)
image = image.clip(0, 255).round().astype(np.uint8)
imwrite(target_path, image)
def main():
args = parse_arguments()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = DeepLab(num_classes=18, pretrained=False)
model.load_state_dict(torch.load(args.model_path, map_location=device)['model_state_dict'])
model.to(device)
model.eval()
image = imread(args.image_path)
valid_transformation = albu.Compose(
[albu.Resize(512, 512), albu.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)), ToTensor()]
)
im = valid_transformation(image=image)["image"].unsqueeze(0)
prediction = model(im.to(device))
prediction = prediction.squeeze(0).detach().cpu().numpy()
prediction = F.softmax(torch.from_numpy(prediction), dim=0).argmax(0).cpu().numpy()
save_image(image, prediction, args.out_path, image.shape[:2])
def __getitem__(self, idx: int) -> dict:
image_path = self.df['image'][idx]
mask_path = self.df['mask'][idx]
image = imread(image_path)
if mask_path is np.nan:
mask = self._create_empty_mask(image.shape[:2])
else:
mat_file = sio.loadmat(self.df['mask'][idx])
mask = self._create_mask(mat_file)
result = {"image": image}
if self.masks is not None:
result["mask"] = mask
if self.transforms is not None:
result = self.transforms(**result)
result['mask'] = torch.squeeze(result['mask']).permute(2, 0, 1)
return result
def __getitem__(self, idx: int) -> dict:
"""Fetch a data sample for a given index.
Args:
index: index of the element in the dataset
Returns:
Single element by index
"""
image_path = self.images[idx]
image = utils.imread(image_path)
result = {"image": image}
if self.masks is not None:
mask = gif_imread(self.masks[idx])
result["mask"] = mask
if self.transforms is not None:
result = self.transforms(result)
result["filename"] = image_path.name
return result
def colors_in_image(uri) -> Set:
image = imread(uri, rootpath=args.in_dir)
colors = np.unique(image.reshape(-1, image.shape[-1]), axis=0)
result = {tuple(row) for row in colors.tolist()} # np.array to hashable
return result
def __getitem__(self, idx: int) -> Dict[str, Any]:
annotation = self._annotations_dataset[idx]
image_name = annotation['image_name']
detections = annotation['detections']
image = utils.imread(image_name)
x_scale, y_scale = self.image_size / image.shape[1], self.image_size / image.shape[0]
image = cv2.resize(image, (self.image_size, self.image_size), cv2.INTER_LINEAR)
detections = [
{
'category_id': detection['category_id'],
'category_name': detection['category_name'],
'bbox': detection['bbox'].copy()
} for detection in detections
]
for detection in detections:
detection['bbox'][0::2] *= x_scale
detection['bbox'][1::2] *= y_scale
bboxes = []
labels = []
for detection in detections:
median_x = (detection['bbox'][0] + detection['bbox'][2]) // 2
median_y = (detection['bbox'][1] + detection['bbox'][3]) // 2
# CenterNet are VERY bad when center of detected objects not in the images
# Let's delete this bboxes
if not (0 <= median_x <= image.shape[1]) or not (0 <= median_y <= image.shape[0]):
continue
detection['bbox'][0::2] = np.clip(detection['bbox'][0::2], 0, image.shape[1])
detection['bbox'][1::2] = np.clip(detection['bbox'][1::2], 0, image.shape[0])
bboxes.append(detection['bbox'])
labels.append(detection['category_id'])
bboxes = np.array(bboxes)
labels = np.array(labels)
if self.transform is not None:
result = self.transform(
image=image,
bboxes=bboxes,
labels=labels,
)
else:
result = dict(
image=image,
bboxes=bboxes,
labels=labels,
)
image = result["image"].astype(np.uint8)
bboxes = result["bboxes"]
labels = result["labels"]
input_height, input_width = image.shape[0], image.shape[1]
# Normalization
input = (image.astype(np.float32) / 255.) * 2. - 1.
input = input.transpose(2, 0, 1)
output_height = input_height // self._down_ratio
output_width = input_width // self._down_ratio
# trans_output = get_affine_transform(center, scale, 0, [output_width, output_height])
heatmap = np.zeros((self._num_classes, output_height, output_width), dtype=np.float32)
width_height = np.zeros((self._max_objects, 2), dtype=np.float32)
reg = np.zeros((self._max_objects, 2), dtype=np.float32)
ind = np.zeros(self._max_objects, dtype=np.int64)
reg_mask = np.zeros(self._max_objects, dtype=np.uint8)
draw_gaussian = draw_umich_gaussian
new_bboxes = []
num_objs = min(len(bboxes), self._max_objects)
for i in range(num_objs):
bbox = np.array(bboxes[i], dtype=np.float32) / self._down_ratio
class_id = labels[i]
bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, output_width - 1)
bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, output_height - 1)
h, w = bbox[3] - bbox[1], bbox[2] - bbox[0]
new_bboxes.append(bbox)
if h > 0 and w > 0:
radius = gaussian_radius((math.ceil(h), math.ceil(w)))
radius = max(0, int(radius))
_center = np.array(
[(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2],
dtype=np.float32
)
_center_int = _center.astype(np.int32)
draw_gaussian(heatmap[class_id], _center_int, radius)
width_height[i] = 1. * w, 1. * h
ind[i] = _center_int[1] * output_width + _center_int[0]
reg[i] = _center - _center_int
reg_mask[i] = 1
result = {
"filename": image_name,
"input": torch.from_numpy(input),
"hm": torch.from_numpy(heatmap),
"reg_mask": torch.from_numpy(reg_mask),
"ind": torch.from_numpy(ind),
"wh": torch.from_numpy(width_height),
"reg": torch.from_numpy(reg),
"bboxes": np.array(bboxes),
"labels": np.array(labels),
}
return result
from catalyst import utils
from catalyst.data.cv import BlurMixin, FlareMixin, RotateMixin
jpg_rgb_uri = (
"https://raw.githubusercontent.com/catalyst-team/catalyst-pics/master"
"/test_images/catalyst_icon.jpg")
image = utils.imread(jpg_rgb_uri)
def test_blur_mixin():
"""@TODO: Docs. Contribution is welcome."""
global image
image_dump = image.copy()
mixin = BlurMixin()
input = {"image": image_dump} # noqa: WPS125
output = mixin(input)
assert mixin.input_key in output
assert mixin.output_key in output
assert output[mixin.input_key].shape == image_dump.shape
assert 0 <= output[mixin.output_key] < mixin.blur_max
def test_flare_mixin():
"""@TODO: Docs. Contribution is welcome."""
global image
image_dump = image.copy()
def __call__(self, row):
file_dir = str(row[self.input_key])
if self.datapath is not None:
file_dir = (file_dir if file_dir.startswith(self.datapath) else
os.path.join(self.datapath, file_dir))
frames = [os.path.join(file_dir, img) for img in os.listdir(file_dir)]
fps = int(frames[0].split("/")[-1].split("_")[1].split(".")[0])
# Step 1: sorting
frames = sorted(frames)
# Step 2: sampling by time
if self.time_window:
frame_window = self.time_window * fps
if self.with_offset:
start_frame = row["offset"] * fps
elif len(frames) > frame_window:
start_frame = np.random.randint(0, len(frames) - frame_window)
else:
start_frame = 0
# mb less than frame_window
frames = frames[start_frame:start_frame + frame_window]
else:
frame_window = len(frames)
# Step 3: choosing frames from time range
# Option a: uniform time sampling (with constant time step)
if self.uniform_time_sample:
frames_indexes = [
int(frame_window / self.num_frames * i)
for i in range(self.num_frames)
]
# If frames less than needed - duplicate last.
# Important to keep constant time step
frames = [frames[min(i, len(frames) - 1)] for i in frames_indexes]
# Option b: (self.num_frames is None) use all frames (bad idea)
elif self.num_frames is not None:
tmp_frames = []
if self.num_segments is not None:
# Option c: random n sample from each successive interval
frames = np.array_split(frames, self.num_segments)
else:
# Option d: random n sample (at all) from all frames
frames = [frames]
for frames_ in frames:
replace_ = self.num_frames > len(frames_)
frames_ = np.random.choice(frames_,
self.num_frames,
replace=replace_).tolist()
tmp_frames.extend(frames_)
tmp_frames = sorted(tmp_frames) # because of random.choice
frames = tmp_frames
frames = [
imread(x, rootpath=self.datapath, grayscale=self.grayscale)
for x in frames
]
result = {self.output_key: frames}
return result