Feature Generator
Aydin use a set of well-engineered feature generators to implement image translators internally. Aydin also provides a public API on feature generators to enable developers who might want to use same feature:
- StandardFeatureGenerator
- ExtensibleFeatureGenerator
Extensible Feature Generator class
Adds a feature to this feature generator.
ExtensibleFeatureGenerator.compute(image[, ...])Computes the features given an image.
Creates a feature array of the right size and possibly in a 'lazy' way using memory mapping.
Clears the features group list
Returns the number of features when considering translations
Returns the receptive field radius in pixels
Returns a 'all-batteries-inlcuded' feature generator from a given path (folder)
Saves a 'all-batteries-included' feature generator at a given path (folder)
- FeatureGeneratorBase
Feature Generator base class
FeatureGeneratorBase.compute(image[, ...])Computes the features given an image.
Creates a feature array of the right size and possibly in a 'lazy' way using memory mapping.
Returns the receptive field radius in pixels
Returns a 'all-batteries-inlcuded' feature generator from a given path (folder)
Saves a 'all-batteries-included' feature generator at a given path (folder)
StandardFeatureGenerator
ExtensibleFeatureGenerator
- class aydin.features.extensible_features.ExtensibleFeatureGenerator[source]
Extensible Feature Generator class
- add_feature_group(feature_group: aydin.features.groups.base.FeatureGroupBase, *args, **kwargs)[source]
Adds a feature to this feature generator.
- Parameters
- feature_groupFeatureGroupBase
feature group
- args
additional arguments for function
- kwargs
additional keyword arguments for function
- compute(image, exclude_center_feature: bool = False, exclude_center_value: bool = False, features: Optional[numpy.ndarray] = None, feature_last_dim: bool = True, passthrough_channels: Optional[Tuple[bool]] = None, num_reserved_features: int = 0, excluded_voxels: Optional[List[Tuple[int]]] = None, spatial_feature_offset: Optional[Tuple[float, ...]] = None, spatial_feature_scale: Optional[Tuple[float, ...]] = None)[source]
Computes the features given an image. If the input image is of shape (d,h,w), resulting features are of shape (n,d,h,w) where n is the number of features.
- Parameters
- imagenumpy.ndarray
image for which features are computed
- exclude_center_featurebool
If true, features that use the image patch’s center pixel are entirely excluded from teh set of computed features.
- exclude_center_valuebool
If true, the center pixel is never used to compute any feature, different feature generation algorithms can take different approaches to achieve that.
- featuresndarray
If None the feature array is allocated internally, if not None the provided array is used to store the features.
- feature_last_dimbool
If True the last dimension of the feature array is the feature dimension, if False then it is the first dimension.
- passthrough_channelsOptional[Tuple[bool]]
Optional tuple of booleans that specify which channels are ‘pass-through’ channels, i.e. channels that are not featurised and directly used as features.
- num_reserved_featuresint
Number of features to be left as blank, useful when adding features separately.
- excluded_voxelsOptional[List[Tuple[int]]]
List of pixel coordinates – expressed as tuple of ints relative to the central pixel – that will be excluded from any computed features. This is used for implementing ‘extended blind-spot’ N2S denoising approaches.
- spatial_feature_offset: Optional[Tuple[float, …]]
Offset vector to be applied (added) to the spatial features (if used).
- spatial_feature_scale: Optional[Tuple[float, …]]
Scale vector to be applied (multiplied) to the spatial features (if used).
- Returns
- feature arraynumpy.ndarray
- create_feature_array(image, nb_features)
Creates a feature array of the right size and possibly in a ‘lazy’ way using memory mapping.
- Parameters
- imagenumpy.ndarray
image for which features are created
- nb_featuresint
- Returns
- feature arraynumpy.ndarray
- get_num_features(ndim: int) int[source]
Returns the number of features when considering translations
- Parameters
- ndimint
number of dimensions
- Returns
- nb_featuresint
- get_receptive_field_radius() int[source]
Returns the receptive field radius in pixels
- Returns
- resultint
receptive field radius in pixels
- static load(path: str)
Returns a ‘all-batteries-inlcuded’ feature generator from a given path (folder)
- Parameters
- pathstr
path to load from
- Returns
- thawed
- save(path: str)
Saves a ‘all-batteries-included’ feature generator at a given path (folder)
- Parameters
- pathstr
path to save to
- Returns
- frozen
FeatureGeneratorBase
- class aydin.features.base.FeatureGeneratorBase[source]
Feature Generator base class
- abstract compute(image, exclude_center_feature: bool = False, exclude_center_value: bool = False, features: Optional[numpy.ndarray] = None, feature_last_dim: bool = True, passthrough_channels: Optional[Tuple[bool]] = None, num_reserved_features: int = 0, excluded_voxels: Optional[List[Tuple[int]]] = None, spatial_feature_offset: Optional[Tuple[float, ...]] = None, spatial_feature_scale: Optional[Tuple[float, ...]] = None)[source]
Computes the features given an image. If the input image is of shape (d,h,w), resulting features are of shape (n,d,h,w) where n is the number of features.
- Parameters
- imagenumpy.ndarray
image for which features are computed
- exclude_center_featurebool
If true, features that use the image patch’s center pixel are entirely excluded from teh set of computed features.
- exclude_center_valuebool
If true, the center pixel is never used to compute any feature, different feature generation algorithms can take different approaches to acheive that.
- featuresndarray
If None the feature array is allocated internally, if not None the provided array is used to store the features.
- feature_last_dimbool
If True the last dimension of the feature array is the feature dimension, if False then it is the first dimension.
- passthrough_channelsOptional[Tuple[bool]]
Optional tuple of booleans that specify which channels are ‘pass-through’ channels, i.e. channels that are not featurised and directly used as features.
- num_reserved_featuresint
Number of features to be left as blank, useful when adding features separately.
- excluded_voxelsOptional[List[Tuple[int]]]
List of pixel coordinates – expressed as tuple of ints relative to the central pixel – that will be excluded from any computed features. This is used for implementing ‘extended blind-spot’ N2S denoising approaches.
- spatial_feature_offset: Optional[Tuple[float, …]]
Offset vector to be applied (added) to the spatial features (if used).
- spatial_feature_scale: Optional[Tuple[float, …]]
Scale vector to be applied (multiplied) to the spatial features (if used).
- Returns
- feature arraynumpy.ndarray
- create_feature_array(image, nb_features)[source]
Creates a feature array of the right size and possibly in a ‘lazy’ way using memory mapping.
- Parameters
- imagenumpy.ndarray
image for which features are created
- nb_featuresint
- Returns
- feature arraynumpy.ndarray