lightly-ai
Expose RandomResizedCrop ratio argument
Currently, https://github.com/lightly-ai/lightly/blob/d893cb592936d80e3d7b921da13e6f513efebf53/lightly/transforms/dino_transform.py#L219-L224 doesn't allow for changing the aspect ratio argument.
If users would like to use the DINO transform but change the aspect ratio of the RandomResizedCrop, they would have to reimplement the transform, something like this
show possible workaround
from typing import Dict, List, Optional, Tuple, Union
import PIL
from PIL.Image import Image
from torch import Tensor
from lightly.transforms.gaussian_blur import GaussianBlur
from lightly.transforms.multi_view_transform import MultiViewTransform
from lightly.transforms.rotation import random_rotation_transform
from lightly.transforms.solarize import RandomSolarization
from lightly.transforms.torchvision_v2_compatibility import torchvision_transforms as T
from lightly.transforms.utils import IMAGENET_NORMALIZE
class DINOTransformWithRatio(MultiViewTransform):
"""Implements the global and local view augmentations for DINO [0].
Input to this transform:
PIL Image or Tensor.
Output of this transform:
List of Tensor of length 2 * global + n_local_views. (8 by default)
Applies the following augmentations by default:
- Random resized crop
- Random horizontal flip
- Color jitter
- Random gray scale
- Gaussian blur
- Random solarization
- ImageNet normalization
This class generates two global and a user defined number of local views
for each image in a batch. The code is adapted from [1].
- [0]: DINO, 2021, https://arxiv.org/abs/2104.14294
- [1]: https://github.com/facebookresearch/dino
Attributes:
global_crop_size:
Crop size of the global views.
global_crop_scale:
Tuple of min and max scales relative to global_crop_size.
local_crop_size:
Crop size of the local views.
local_crop_scale:
Tuple of min and max scales relative to local_crop_size.
n_local_views:
Number of generated local views.
hf_prob:
Probability that horizontal flip is applied.
vf_prob:
Probability that vertical flip is applied.
rr_prob:
Probability that random rotation is applied.
rr_degrees:
Range of degrees to select from for random rotation. If rr_degrees is None,
images are rotated by 90 degrees. If rr_degrees is a (min, max) tuple,
images are rotated by a random angle in [min, max]. If rr_degrees is a
single number, images are rotated by a random angle in
[-rr_degrees, +rr_degrees]. All rotations are counter-clockwise.
cj_prob:
Probability that color jitter is applied.
cj_strength:
Strength of the color jitter. `cj_bright`, `cj_contrast`, `cj_sat`, and
`cj_hue` are multiplied by this value.
cj_bright:
How much to jitter brightness.
cj_contrast:
How much to jitter constrast.
cj_sat:
How much to jitter saturation.
cj_hue:
How much to jitter hue.
random_gray_scale:
Probability of conversion to grayscale.
gaussian_blur:
Tuple of probabilities to apply gaussian blur on the different
views. The input is ordered as follows:
(global_view_0, global_view_1, local_views)
kernel_size:
Will be deprecated in favor of `sigmas` argument. If set, the old behavior applies and `sigmas` is ignored.
Used to calculate sigma of gaussian blur with kernel_size * input_size.
kernel_scale:
Old argument. Value is deprecated in favor of sigmas. If set, the old behavior applies and `sigmas` is ignored.
Used to scale the `kernel_size` of a factor of `kernel_scale`
sigmas:
Tuple of min and max value from which the std of the gaussian kernel is sampled.
Is ignored if `kernel_size` is set.
solarization:
Probability to apply solarization on the second global view.
normalize:
Dictionary with 'mean' and 'std' for torchvision.transforms.Normalize.
"""
def __init__(
self,
global_crop_size: int = 224,
global_crop_scale: Tuple[float, float] = (0.4, 1.0),
global_crop_ratio: Tuple[float, float] = (0.75, 1.3333),
local_crop_size: int = 96,
local_crop_scale: Tuple[float, float] = (0.05, 0.4),
local_crop_ratio: Tuple[float, float] = (0.75, 1.3333),
n_local_views: int = 6,
hf_prob: float = 0.5,
vf_prob: float = 0,
rr_prob: float = 0,
rr_degrees: Optional[Union[float, Tuple[float, float]]] = None,
cj_prob: float = 0.8,
cj_strength: float = 0.5,
cj_bright: float = 0.8,
cj_contrast: float = 0.8,
cj_sat: float = 0.4,
cj_hue: float = 0.2,
random_gray_scale: float = 0.2,
gaussian_blur: Tuple[float, float, float] = (1.0, 0.1, 0.5),
kernel_size: Optional[float] = None,
kernel_scale: Optional[float] = None,
sigmas: Tuple[float, float] = (0.1, 2),
solarization_prob: float = 0.2,
normalize: Union[None, Dict[str, List[float]]] = IMAGENET_NORMALIZE,
):
# first global crop
global_transform_0 = DINOViewTransformWithRatio(
crop_size=global_crop_size,
crop_scale=global_crop_scale,
crop_ratio=global_crop_ratio,
hf_prob=hf_prob,
vf_prob=vf_prob,
rr_prob=rr_prob,
rr_degrees=rr_degrees,
cj_prob=cj_prob,
cj_strength=cj_strength,
cj_bright=cj_bright,
cj_contrast=cj_contrast,
cj_hue=cj_hue,
cj_sat=cj_sat,
random_gray_scale=random_gray_scale,
gaussian_blur=gaussian_blur[0],
kernel_size=kernel_size,
kernel_scale=kernel_scale,
sigmas=sigmas,
solarization_prob=0,
normalize=normalize,
)
# second global crop
global_transform_1 = DINOViewTransformWithRatio(
crop_size=global_crop_size,
crop_scale=global_crop_scale,
crop_ratio=global_crop_ratio,
hf_prob=hf_prob,
vf_prob=vf_prob,
rr_prob=rr_prob,
rr_degrees=rr_degrees,
cj_prob=cj_prob,
cj_bright=cj_bright,
cj_contrast=cj_contrast,
cj_hue=cj_hue,
cj_sat=cj_sat,
random_gray_scale=random_gray_scale,
gaussian_blur=gaussian_blur[1],
kernel_size=kernel_size,
kernel_scale=kernel_scale,
sigmas=sigmas,
solarization_prob=solarization_prob,
normalize=normalize,
)
# transformation for the local small crops
local_transform = DINOViewTransformWithRatio(
crop_size=local_crop_size,
crop_scale=local_crop_scale,
crop_ratio=local_crop_ratio,
hf_prob=hf_prob,
vf_prob=vf_prob,
rr_prob=rr_prob,
rr_degrees=rr_degrees,
cj_prob=cj_prob,
cj_strength=cj_strength,
cj_bright=cj_bright,
cj_contrast=cj_contrast,
cj_hue=cj_hue,
cj_sat=cj_sat,
random_gray_scale=random_gray_scale,
gaussian_blur=gaussian_blur[2],
kernel_size=kernel_size,
kernel_scale=kernel_scale,
sigmas=sigmas,
solarization_prob=0,
normalize=normalize,
)
local_transforms = [local_transform] * n_local_views
transforms = [global_transform_0, global_transform_1]
transforms.extend(local_transforms)
super().__init__(transforms)
class DINOViewTransformWithRatio:
def __init__(
self,
crop_size: int = 224,
crop_scale: Tuple[float, float] = (0.4, 1.0),
crop_ratio: Tuple[float, float] = (0.75, 1.3333),
hf_prob: float = 0.5,
vf_prob: float = 0,
rr_prob: float = 0,
rr_degrees: Optional[Union[float, Tuple[float, float]]] = None,
cj_prob: float = 0.8,
cj_strength: float = 0.5,
cj_bright: float = 0.8,
cj_contrast: float = 0.8,
cj_sat: float = 0.4,
cj_hue: float = 0.2,
random_gray_scale: float = 0.2,
gaussian_blur: float = 1.0,
kernel_size: Optional[float] = None,
kernel_scale: Optional[float] = None,
sigmas: Tuple[float, float] = (0.1, 2),
solarization_prob: float = 0.2,
normalize: Union[None, Dict[str, List[float]]] = IMAGENET_NORMALIZE,
):
transform = [
T.RandomResizedCrop(
size=crop_size,
scale=crop_scale,
ratio=crop_ratio,
# Type ignore needed because BICUBIC is not recognized as an attribute.
interpolation=PIL.Image.BICUBIC, # type: ignore[attr-defined]
),
T.RandomHorizontalFlip(p=hf_prob),
T.RandomVerticalFlip(p=vf_prob),
random_rotation_transform(rr_prob=rr_prob, rr_degrees=rr_degrees),
T.RandomApply(
[
T.ColorJitter(
brightness=cj_strength * cj_bright,
contrast=cj_strength * cj_contrast,
saturation=cj_strength * cj_sat,
hue=cj_strength * cj_hue,
)
],
p=cj_prob,
),
T.RandomGrayscale(p=random_gray_scale),
GaussianBlur(
kernel_size=kernel_size,
scale=kernel_scale,
sigmas=sigmas,
prob=gaussian_blur,
),
RandomSolarization(prob=solarization_prob),
T.ToTensor(),
]
if normalize:
transform += [T.Normalize(mean=normalize["mean"], std=normalize["std"])]
self.transform = T.Compose(transform)
def __call__(self, image: Union[Tensor, Image]) -> Tensor:
"""
Applies the transforms to the input image.
Args:
image:
The input image to apply the transforms to.
Returns:
The transformed image.
"""
transformed: Tensor = self.transform(image)
return transformed
For me, this workaround works, but it a bit cumbersome for just one small change.
There may be other arguments to transforms used by the DINOTransform that are not exposed but can still be used for interesting experiments.
Also, there may be other transforms like DINOTransform that have the same limitation.
Would it be possible to implement this feature?
Hi @siemdejong, That looks like a sensible approach! What other arguments that are not exposed did you find limiting? Would you be willing to create a PR for it?
I'd be happy to contribute. Just a quick look reveals more examples of parameters that may be limiting to others, like
- the scale and ratio parameters in https://github.com/lightly-ai/lightly/blob/d893cb592936d80e3d7b921da13e6f513efebf53/lightly/transforms/simclr_transform.py#L158 and https://github.com/lightly-ai/lightly/blob/d893cb592936d80e3d7b921da13e6f513efebf53/lightly/transforms/wmse_transform.py#L92-L96
- Maybe the interpolation parameter can be added too
I can create a PR and list unexposed parameters there.
Thanks @siemdejong !
I am not sure yet if we really want to pack all possible parameters (interpolation, antialias, ...) into the constructors of those transforms or if we instead want to go a route of more composability in the future anyway, e.g. with dataclasses or pydantic models that would keep the hierarchy of the configuration intact (instead of flattening them and having to prefix with cj_ or crop_, which does not scale well). That said, imo crop_ratio would indeed be a sensible option to add here, but at the same time it would technically introduce a breaking change (unless we put it as the last argument of the constructor, where it makes little sense).
I will discuss this with the team and get back to you towards the end of next week!
Hey @siemdejong! Sorry for the delayed response: I created an internal issue when you submitted this, however unfortunately your request is currently not at the top of the priority list. However, we will make sure to add better abstractions when we add new methods and we will hopefully get to refactoring this in the not so distant future.