API¶
benzina.torch.dataloader¶
benzina.torch.dataset¶
benzina.torch.operations¶
-
class
benzina.torch.operations.
WarpTransform
[source] Interface class that represents a warp transformation as a combined rotation, scale, skew and translation 3 x 3 matrix. The transformation is called for each sample of a batch.
-
class
benzina.torch.operations.
NormTransform
[source] Interface class that represents a normalization transformation. The transformation is called for each sample of a batch.
-
class
benzina.torch.operations.
BiasTransform
[source] Interface class that represents a bias transformation. The transformation is called for each sample of a batch.
-
class
benzina.torch.operations.
ConstantWarpTransform
(warp=(1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0))[source] Represents a constant warp transformation to be applied on each sample of a batch independently of its index.
Parameters: warp (iterable of numerics, optional) – a flatten, row-major 3 x 3 warp matrix (default: flatten identity matrix).
-
class
benzina.torch.operations.
ConstantNormTransform
(norm=(1.0, 1.0, 1.0))[source] Represents a constant norm transformation to be applied on each sample of a batch independently of its index.
Parameters: norm (numeric or iterable of numerics, optional) – an iterable in RGB order containing the normalization constant of a sample’s RGB channels. Components will be multiplied to the respective channels of a sample (default: (1.0, 1.0, 1.0)).
-
class
benzina.torch.operations.
ConstantBiasTransform
(bias=(0.0, 0.0, 0.0))[source] Represents a constant bias transformation to be applied on each sample of a batch independently of its index.
Parameters: bias (numeric or iterable of numerics, optional) – an iterable in RGB order containing the bias of a sample’s RGB channels. Components will be substracted to the respective channels of a sample (default: (0.0, 0.0, 0.0)).
-
class
benzina.torch.operations.
SimilarityTransform
(scale=(1.0, 1.0), ratio=None, degrees=(-0.0, 0.0), translate=(0.0, 0.0), flip_h=0.0, flip_v=0.0, resize=False, keep_ratio=False, random_crop=False)[source] Similarity warp transformation of the image keeping center invariant.
A crop of random size, aspect ratio and location is made. This crop can then be flipped and/or rotated to finally be resized to output size.
Parameters: - scale (Sequence or float or int, optional) – crop area scaling factor
interval, e.g (a, b), then scale is randomly sampled from the range
a <= scale <= b. If scale is a number instead of sequence, the
range of scale will be (scale^-1, scale).
(default:
(+1.0, +1.0)
) - ratio (Sequence or float or int, optional) – range of crop aspect ratio. If ratio is a number instead of sequence like (min, max), the range of aspect ratio will be (ratio^-1, ratio). Will keep original aspect ratio by default.
- degrees (Sequence or float or int, optional) – range of degrees to
select from. If degrees is a number instead of sequence like
(min, max), the range of degrees will be (-degrees, +degrees).
(default:
(-0.0, +0.0)
) - translate (Sequence or float or int, optional) – sequence of maximum
absolute fraction for horizontal and vertical translations. For
example translate=(a, b), then horizontal shift is randomly sampled
in the range -output_width * a < dx < output_width * a and vertical
shift is randomly sampled in the range
-output_height * b < dy < output_height * b. If translate is a
number instead of sequence, translate will be
(translate, translate). These translations are applied
independently from
random_crop
. (default:(0.0, 0.0)
) - flip_h (bool, optional) – probability of the image being flipped
horizontally. (default:
+0.0
) - flip_v (bool, optional) – probability of the image being flipped
vertically. (default:
+0.0
) - resize (bool, optional) – resize the cropped image to fit the output
size. It is forced to
True
ifscale
orratio
are specified. (default:False
) - keep_ratio (bool, optional) – match the smaller edge to the
corresponding output edge size, keeping the aspect ratio after
resize. Has no effect if
resize
isFalse
. (default:False
) - random_crop (bool, optional) – randomly crop the image instead of
a center crop. (default:
False
)
- scale (Sequence or float or int, optional) – crop area scaling factor
interval, e.g (a, b), then scale is randomly sampled from the range
a <= scale <= b. If scale is a number instead of sequence, the
range of scale will be (scale^-1, scale).
(default:
-
class
benzina.torch.operations.
RandomResizedCrop
(scale=(0.08, 1.0), ratio=(0.75, 1.3333333333333333))[source] Crop to random size, aspect ratio and location.
A crop of random size, aspect ratio and location is made. This crop is finally resized to output size.
This is popularly used to train the Inception networks.
Parameters: - scale (Sequence or float or int, optional) – crop area scaling factor
interval, e.g (a, b), then scale is randomly sampled from the range
a <= scale <= b. If scale is a number instead of sequence, the
range of scale will be (scale^-1, scale).
(default:
(+0.08, +1.0)
) - ratio (Sequence or float or int, optional) – range of crop aspect ratio.
If ratio is a number instead of sequence like (min, max), the range
of aspect ratio will be (ratio^-1, ratio). Will keep original
aspect ratio by default. (default:
(3./4., 4./3.)
)
- scale (Sequence or float or int, optional) – crop area scaling factor
interval, e.g (a, b), then scale is randomly sampled from the range
a <= scale <= b. If scale is a number instead of sequence, the
range of scale will be (scale^-1, scale).
(default:
-
class
benzina.torch.operations.
CenterResizedCrop
(scale=1.0, keep_ratio=True)[source] Crops at the center and resize.
A crop at the center is made then resized to the output size.
Parameters: - scale (float or int, optional) – edges scaling factor.
(default:
+1.0
) - keep_ratio (bool, optional) – match the smaller edge to the
corresponding output edge size, keeping the aspect ratio after
resize. (default:
False
)
- scale (float or int, optional) – edges scaling factor.
(default:
-
benzina.torch.operations.
compute_affine_matrix
(in_shape, out_shape, crop=None, degrees=0.0, translate=(0.0, 0.0), flip_h=False, flip_v=False, resize=False, keep_ratio=False)[source] Similarity warp transformation of the image keeping center invariant.
Parameters: - in_shape (Sequence) – the shape of the input image
- out_shape (Sequence) – the shape of the output image
- crop (Sequence, optional) – crop center location, width and height. The
center location is relative to the center of the image. If
resize
is notTrue
, crop is simply a translation in thein_shape
space. - degrees (float or int, optional) – degrees to rotate the crop.
(default:
(0.0)
) - translate (Sequence, optional) – horizontal and vertical translations.
(default:
(0.0, 0.0)
) - flip_h (bool, optional) – flip the image horizontally.
(default:
False
) - flip_v (bool, optional) – flip the image vertically.
(default:
False
) - resize (bool, optional) – resize the cropped image to fit the output’s
size. (default:
False
) - keep_ratio (bool, optional) – match the smaller edge to the
corresponding output edge size, keeping the aspect ratio after
resize. Has no effect if
resize
isFalse
. (default:False
)