dml package¶
Submodules¶
dml.anmm module¶
Average Neighborhood Margin Maximization (ANMM)
-
class
dml.anmm.
ANMM
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Average Neighborhood Margin Maximization (ANMM)
A DML Algorithm that obtains a transformer that maximizes the distance between the nearest friends and the nearest enemies for each example.
Parameters: - num_dims : int, default=None
Dimension desired for the transformed data. If None, all features will be taken.
- n_friends : int, default=3
Number of nearest same-class neighbors to compute homogeneus neighborhood.
- n_enemies : int, default=1
Number of nearest different-class neighbors to compute heterogeneus neigborhood.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
num_dims : dimension of the reduced data.
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
-
class
dml.anmm.
KANMM
¶ Bases:
dml.dml_algorithm.KernelDML_Algorithm
The kernelized version of ANMM.
Parameters: - num_dims : int, default=None
Dimension desired for the transformed data.
- n_friends : int, default=3
Number of nearest same-class neighbors to compute homogeneus neighborhood.
- n_enemies : int, default=1
Number of nearest different-class neighbors to compute heterogeneus neigborhood.
- kernel : “linear” | “poly” | “rbf” | “sigmoid” | “cosine” | “precomputed”
Kernel. Default=”linear”.
- gamma : float, default=1/n_features
Kernel coefficient for rbf, poly and sigmoid kernels. Ignored by other kernels.
- degree : int, default=3
Degree for poly kernels. Ignored by other kernels.
- coef0 : float, default=1
Independent term in poly and sigmoid kernels. Ignored by other kernels.
- kernel_params : mapping of string to any, default=None
Parameters (keyword arguments) and values for kernel passed as callable object. Ignored by other kernels.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the kernel transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
transformer
¶ Obtains the learned projection.
Returns: - A : (d’x N) matrix, where d’ is the desired output dimension, and N is the number of samples.
To apply A to a new sample x, A must be multiplied by the kernel vector of dimension N obtained by taking the kernels between x and each training sample.
dml.base module¶
Some basic DML implementations.
Created on Fri Mar 30 19:13:58 2018
@author: jlsuarezdiaz
-
class
dml.base.
Covariance
(tol=1e-15, reg_method='pseudoinverse', alpha=0.001)¶ Bases:
dml.dml_algorithm.DML_Algorithm
A distance metric learning algorithm that learns the distance given by the inverse covariance matrix (the original concept of Mahalanobis distance).
Parameters: - tol: float, default=1e-15
The toleration level to consider the covariance matrix invertible.
- reg_method: string, default=’pseudoinverse’
The strategy to deal with singular matrix. Valid options are:
- ‘pseudoinverse’ : calculates the pseudoinverse matrix when covariance is singular.
- ‘addid’ : adds a multiple of identity to the covariance matrix to make it invertible.
- The value added is given by the ‘alpha’ parameter.
- alpha : regularization constant to calculate the inverse covariance matrix.
Ignored if reg_method is not ‘addid’.
Methods
fit
(X[, y])Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
(X, y=None)¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples. Added for compatibility, but not used.
Returns: - self : object
Returns the instance itself.
-
metric
()¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
-
class
dml.base.
Euclidean
¶ Bases:
dml.dml_algorithm.DML_Algorithm
A basic transformer that represents the euclidean distance.
Methods
fit
(X[, y])Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Obtains the learned projection. -
fit
(X, y=None)¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metric
()¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
-
transform
(X=None)¶ Applies the metric transformation.
Parameters: - X : (N x d) matrix, optional
Data to transform. If not supplied, the training data will be used.
Returns: - transformed : (N x d’) matrix
Input data transformed to the metric space by \(XL^{\top}\)
-
transformer
()¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
-
-
class
dml.base.
Metric
(metric)¶ Bases:
dml.dml_algorithm.DML_Algorithm
A DML algorithm that defines a distance given a PSD metric matrix.
Parameters: - metric : (d x d) matrix. A positive semidefinite matrix, to define a pseudodistance in euclidean d-dimensional space.
Methods
fit
([X, y])Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
(X=None, y=None)¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metric
()¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
-
class
dml.base.
Transformer
(transformer)¶ Bases:
dml.dml_algorithm.DML_Algorithm
A DML algorithm that defines a distance given a linear transformation.
Parameters: - transformer : (d’ x d) matrix, representing a linear transformacion from d-dimensional euclidean space
to d’-dimensional euclidean space.
Methods
fit
([X, y])Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Obtains the learned projection. -
fit
(X=None, y=None)¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
transformer
()¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
dml.dml_algorithm module¶
Distance Metric Algorithm basis.
-
class
dml.dml_algorithm.
DML_Algorithm
¶ Bases:
sklearn.base.BaseEstimator
,sklearn.base.TransformerMixin
Abstract class that defines a distance metric learning algorithm. Distance metric learning are implemented as subclasses of DML_Algorithm. A DML Algorithm can compute either a Mahalanobis metric matrix or an associated linear transformation. DML subclasses must override one of the following methods (metric or transformer), according to their computation way.
Methods
fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
metadata
()¶ Obtains the algorithm metadata. Must be implemented in subclasses.
Returns: - dict : A map from string to any.
-
metric
()¶ Computes the Mahalanobis matrix from the transformation matrix. .. math:: M = L^T L
Returns: - M : (d x d) matrix. M defines a metric whose distace is given by
- ..math:: d(x,y) = sqrt{(x-y)^TM(x-y)}.
-
transform
(X=None)¶ Applies the metric transformation.
Parameters: - X : (N x d) matrix, optional
Data to transform. If not provided, the training data will be used.
Returns: - transformed : (N x d’) matrix
Input data transformed to the metric space. The learned distance can be measured using the euclidean distance with the transformed data.
-
transformer
()¶ Computes a transformation matrix from the Mahalanobis matrix. ..math:: L = M^{1/2}
Returns: - L : (d’ x d) matrix, with d’ <= d. It defines a projection. The distance can be calculated by
- ..math:: d(x,y) = |L(x-y)|_2.
-
-
class
dml.dml_algorithm.
KernelDML_Algorithm
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Abstract class that defines a kernel distance metric learning algorithm. Distance metric learning are implemented as subclasses of KernelDML_Algorithm. A Kernel DML Algorithm can compute a (d’ x n) transformer that maps the high dimensional data using the kernel trick. Kernel DML subclasses must override the transformer method, providing the matrix A that performs the kernel trick, that is
\[Lx = A(K(x_1,x),\dots,K(x_n,x)),\]where L is the high dimensional transformer and K is the kernel function.
Methods
fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the kernel transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
transform
(X=None)¶ Applies the kernel transformation.
Parameters: - X : (N x d) matrix, optional
Data to transform. If not supplied, the training data will be used.
Returns: - transformed: (N x d’) matrix.
Input data transformed by the learned mapping.
-
dml.dml_eig module¶
Distance Metric Learning with Eigenvalue Optimization
Created on Fri Mar 9 10:18:35 2018
@author: jlsuarezdiaz
-
class
dml.dml_eig.
DML_eig
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Distance Metric Learning with Eigenvalue Optimization (DML-eig)
A DML Algorithm that learns a metric that minimizes the minimum distance between different-class points constrained to the sum of distances at same-class points be non higher than a constant.
Parameters: - mu : float, default=1e-4
Smoothing parameter.
- tol : float, default=1e-5
Tolerance stop criterion (difference between two point iterations at gradient descent).
- eps : float, default=1e-10
Precision stop criterion (norm of gradient at gradient descent).
- max_it: int, default=25
Number of iterations at gradient descent.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
initial_error : initial value of the objective error function.
final_error : final value of the objective error function.
-
metric
¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
dml.dml_plot module¶
Plot utilies for classifiers and distance metric learners.
Created on Sat Feb 3 16:45:28 2018
@author: jlsuarezdiaz
-
dml.dml_plot.
classifier_pairplots
(X, y, clf, attrs=None, xattrs=None, yattrs=None, diag='hist', sections='mean', fitted=False, title=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='center right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=False, **fig_kw)¶ This function allows multiple 2D-scatter plots for different pairs of attributes of the same dataset, and to plot regions defined by different classifiers.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- clf : A
ClassifierMixin
. A classifier. It must support the methods fit(X,y) and predict(X), as specified in
ClassifierMixin
- attrs : List, default=None
A list specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, and xattrs and yattrs are None, all the attributes will be taken.
- xattrs : List, default=None
A list specifying the dataset attributes to show in X axis. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. Ignored if attrs is specified.
- yattrs : List, default=None
A list specifying the dataset attributes to show in Y axis. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. Ignored if attrs is specified.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
classifier_plot
(X, y, clf, attrs=None, sections='mean', fitted=False, f=None, ax=None, title=None, subtitle=None, xrange=None, yrange=None, xlabel=None, ylabel=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='lower right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=True, **fig_kw)¶ A 2D-scatter plot for a labeled dataset, together with a classifier that allows to plot each clasiffier region.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- clf : A classifier. It must support the methods fit(X,y) and predict(X), as specified in
ClassifierMixin
- attrs : List, default=None
A list of two items specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, the two first attributes will be taken.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- subtitle : String, default=None. An optional subtitle for the plot.
- xrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the X axis. If None, it will be calculated according to the maximum and minimum of the X feature.
- yrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the Y axis. If None, it will be calculated according to the maximum and minimum of the Y feature.
- xlabel : String, default=None. An optional title for the X axis.
- ylabel : String, default=None. An optional title for the Y axis.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
classifier_plot_3d
(X, y, clf, attrs=None, sections='mean', fitted=False, f=None, ax=None, elev=0.0, azim=0.0, title=None, subtitle=None, xrange=None, yrange=None, zrange=None, xlabel=None, ylabel=None, zlabel=None, grid_split=[40, 40, 40], grid_step=[0.1, 0.1, 0.1], label_legend=True, legend_loc='lower right', cmap=None, label_colors=None, plot_points=True, plot_regions='all', region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=True, **fig_kw)¶ A 3D-scatter plot for a labeled dataset, together with a classifier that allows to plot each clasiffier region.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- clf : A classifier. It must support the methods fit(X,y) and predict(X), as specified in
ClassifierMixin
- attrs : List, default=None
A list of three items specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, the three first attributes will be taken.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- elev : Float, default=0.0
The elevation parameter for the 3D plot.
- azim : Float, default=0.0
The azimut parameter for the 3D plot.
- title : String, default=None. An optional title for the plot.
- subtitle : String, default=None. An optional subtitle for the plot.
- xrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the X axis. If None, it will be calculated according to the maximum and minimum of the X feature.
- yrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the Y axis. If None, it will be calculated according to the maximum and minimum of the Y feature.
- zrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the Ζ axis. If None, it will be calculated according to the maximum and minimum of the Ζ feature.
- xlabel : String, default=None. An optional title for the X axis.
- ylabel : String, default=None. An optional title for the Y axis.
- zlabel : String, default=None. An optional title for the Ζ axis.
- grid_split : List, default=[40, 40, 40]
A list with three items, specifying the number of partitions, in the X, Y and Z axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1, 0.1, 0.1]
A list with three items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
dml_multiplot
(X, y, nrow=None, ncol=None, ks=None, clfs=None, attrs=None, sections='mean', fitted=False, metrics=None, transformers=None, dmls=None, dml_fitted=False, transforms=None, title=None, subtitles=None, xlabels=None, ylabels=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='center right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=False, **fig_kw)¶ This functions allows multiple 2D-scatter plots for a labeled dataset, to plot regions defined by different classifiers and distances. The distances can be provided by a metric PSD matrix, a matrix of a linear transformation, or by a distance metric learning algorithm, that can learn the distance during the plotting, or it can be fitted previously.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- nrow : integer, default=None.
Number of rows of the figure. If any of nrow or ncol is None, it will be generated automatically.
- ncol : integer, default=None.
Number of columns of the figure. If any of nrow or ncol is None, it will be generated automatically.
- ks : List of int, default=None.
The number of neighbors for the k-NN classifier in each plot. List size must be equal to the number of plots. ks[i] is ignored if clfs[i] is specified.
- clfs : List of
ClassifierMixin
. Default=None. The classifier to use in each plot. List size must be equal to the number of plots.
- attrs : List, default=None
A list of two items specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, the two first attributes will be taken.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- metrics : List of Matrix, or 2D-Array. Default=None.
The metric PSD matrix to use in each plot. List size must be equal to the number of plots. metrics[i] is ignored if transformers[i] or dmls[i] are provided.
- transformers : List of Matrix, or 2D-Array. Default=None.
A linear transformation to use in each plot. List size must be equal to the number of plots. transformers[i] will be ignored if dmls[i] is provided.
- dmls : List of DML_Algorithm, default=None.
A distance metric learning algorithm for each plot. List size must be equal to the number of plots.
- dml_fitted : Boolean, default=True.
Specifies if the DML algorithms are already fitted. If True, the algorithms’ fit method will not be called.
- transforms: List of Boolean, default=True.
For each plot where the list item is True, it projects the data by the learned transformer and plots the transform data. Else, the classifier region will be ploted with the original data, but the regions will change according to the learned distance. List size must be equal to the number of plots.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- subtitles : List of String, default=None.
Optional titles for each subplot. List size must be equal to the number of plots.
- xlabels : List of String, default=None.
Optional titles for the X axis. List size must be equal to the number of plots.
- ylabels : List of String, default=None.
Optional titles for the Y axis. List size must be equal to the number of plots.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=False.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
dml_pairplots
(X, y, clf, attrs=None, xattrs=None, yattrs=None, diag='hist', sections='mean', fitted=False, metric=None, transformer=None, dml=None, dml_fitted=False, title=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='center right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=False, **fig_kw)¶ This function allows multiple 2D-scatter plots for different pairs of attributes of the same dataset, and to plot regions defined by different classifiers and a distance. The distance can be provided by a metric PSD matrix, a matrix of a linear transformation, or by a distance metric learning algorithm, that can learn the distance during the plotting, or it can be fitted previously.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- clf : A
ClassifierMixin
. A classifier. It must support the methods fit(X,y) and predict(X), as specified in
ClassifierMixin
- attrs : List, default=None
A list specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, and xattrs and yattrs are None, all the attributes will be taken.
- xattrs : List, default=None
A list specifying the dataset attributes to show in X axis. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. Ignored if attrs is specified.
- yattrs : List, default=None
A list specifying the dataset attributes to show in Y axis. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. Ignored if attrs is specified.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- metric : Matrix, or 2D-Array. Default=None.
A positive semidefinite matrix of size (d x d), where d is the number of features. Ignored if dml or transformer is specified.
- transformer : Matrix, or 2D-Array. Default=None.
A matrix of size (d’ x d), where d is the number of features and d’ is the desired dimension. Ignored if dml is specified.
- dml : DML_Algorithm, default=None.
A distance metric learning algorithm. If metric, transformer and dml are None, no distances are used in the plot.
- dml_fitted : Boolean, default=True.
Specifies if the DML algorithm is already fitted. If True, the algorithm’s fit method will not be called.
- transform: Boolean, default=True.
If True, projects the data by the learned transformer and plots the transform data. Else, the classifier region will be ploted with the original data, but the regions will change according to the learned distance.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
dml_plot
(X, y, clf, attrs=None, sections='mean', fitted=False, metric=None, transformer=None, dml=None, dml_fitted=False, transform=True, f=None, ax=None, title=None, subtitle=None, xrange=None, yrange=None, xlabel=None, ylabel=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='lower right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=True, **fig_kw)¶ A 2D-scatter plot for a labeled dataset, together with a classifier that allows to plot each clasiffier region, and a distance that can be used by the classifier. The distance can be provided by a metric PSD matrix, a matrix of a linear transformation, or by a distance metric learning algorithm, that can learn the distance during the plotting, or it can be fitted previously.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- clf : A classifier. It must support the methods fit(X,y) and predict(X), as specified in
ClassifierMixin
- attrs : List, default=None
A list of two items specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, the two first attributes will be taken.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- metric : Matrix, or 2D-Array. Default=None.
A positive semidefinite matrix of size (d x d), where d is the number of features. Ignored if dml or transformer is specified.
- transformer : Matrix, or 2D-Array. Default=None.
A matrix of size (d’ x d), where d is the number of features and d’ is the desired dimension. Ignored if dml is specified.
- dml : DML_Algorithm, default=None.
A distance metric learning algorithm. If metric, transformer and dml are None, no distances are used in the plot.
- transform: Boolean, default=True.
If True, projects the data by the learned transformer and plots the transform data. Else, the classifier region will be ploted with the original data, but the regions will change according to the learned distance.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- subtitle : String, default=None. An optional subtitle for the plot.
- xrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the X axis. If None, it will be calculated according to the maximum and minimum of the X feature.
- yrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the Y axis. If None, it will be calculated according to the maximum and minimum of the Y feature.
- xlabel : String, default=None. An optional title for the X axis.
- ylabel : String, default=None. An optional title for the Y axis.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
knn_pairplots
(X, y, k=1, attrs=None, xattrs=None, yattrs=None, diag='hist', sections='mean', knn_clf=None, fitted=False, metric=None, transformer=None, dml=None, dml_fitted=False, title=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='center right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=False, **fig_kw)¶ This function allows multiple 2D-scatter plots for different pairs of attributes of the same dataset, and to plot regions defined by a k-NN classifier and a distance. The distance can be provided by a metric PSD matrix, a matrix of a linear transformation, or by a distance metric learning algorithm, that can learn the distance during the plotting, or it can be fitted previously.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- k : int, default=1.
The number of neighbors for the k-NN classifier. Ignored if knn_clf is specified.
- knn_clf : A
ClassifierMixin
. An already defined kNN classifier. It can be any other classifier, but then options are the same as in :meth: ~dml_plot.
- attrs : List, default=None
A list specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, and xattrs and yattrs are None, all the attributes will be taken.
- xattrs : List, default=None
A list specifying the dataset attributes to show in X axis. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. Ignored if attrs is specified.
- yattrs : List, default=None
A list specifying the dataset attributes to show in Y axis. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. Ignored if attrs is specified.
- diag : String
What to plot on the diagonal subplots. Allowed options are:
- “hist” : An histogram of the data will be plot for the attribute.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- metric : Matrix, or 2D-Array. Default=None.
A positive semidefinite matrix of size (d x d), where d is the number of features. Ignored if dml or transformer is specified.
- transformer : Matrix, or 2D-Array. Default=None.
A matrix of size (d’ x d), where d is the number of features and d’ is the desired dimension. Ignored if dml is specified.
- dml : DML_Algorithm, default=None.
A distance metric learning algorithm. If metric, transformer and dml are None, no distances are used in the plot.
- dml_fitted : Boolean, default=True.
Specifies if the DML algorithm is already fitted. If True, the algorithm’s fit method will not be called.
- transform: Boolean, default=True.
If True, projects the data by the learned transformer and plots the transform data. Else, the classifier region will be ploted with the original data, but the regions will change according to the learned distance.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot.classifier Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for
suplots()
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
-
dml.dml_plot.
knn_plot
(X, y, k=1, attrs=None, sections='mean', knn_clf=None, fitted=False, metric=None, transformer=None, dml=None, dml_fitted=False, transform=True, f=None, ax=None, title=None, subtitle=None, xrange=None, yrange=None, xlabel=None, ylabel=None, grid_split=[400, 400], grid_step=[0.1, 0.1], label_legend=True, legend_loc='lower right', cmap=None, label_colors=None, plot_points=True, plot_regions=True, region_intensity=0.4, legend_plot_points=True, legend_plot_regions=True, legend_on_axis=True, **fig_kw)¶ A 2D-scatter plot for a labeled dataset to plot regions defined by a k-NN classifier and a distance. The distance can be provided by a metric PSD matrix, a matrix of a linear transformation, or by a distance metric learning algorithm, that can learn the distance during the plotting, or it can be fitted previously.
Parameters: - X : array-like of size (N x d), where N is the number of samples, and d is the number of features.
- y : array-like of size N, where N is the number of samples.
- k : int, default=1.
The number of neighbors for the k-NN classifier. Ignored if knn_clf is specified.
- knn_clf : A
ClassifierMixin
. An already defined kNN classifier. It can be any other classifier, but then options are the same as in :meth: ~dml_plot.
- attrs : List, default=None
A list of two items specifying the dataset attributes to show in the scatter plot. The items can be the keys, if X is a pandas dataset, or integer indexes with the attribute position. If None, the two first attributes will be taken.
- sections : String, default=fitted
It specifies how to take sections in the features space, if there are more than two features in the dataset. It is used to plot the classifier fixing the non-ploting attributes in this space section. Allowed values are:
‘mean’ : takes the mean of the remaining attributes to plot the classifier region.
‘zeros’ : takes the remaining attributes as zeros to plot the classifier region.
- fitted : Boolean, default=False.
It indicates if the classifier has already been fitted. If it is false, the function will call the classifier fit method with parameters X,y.
- metric : Matrix, or 2D-Array. Default=None.
A positive semidefinite matrix of size (d x d), where d is the number of features. Ignored if dml or transformer is specified.
- transformer : Matrix, or 2D-Array. Default=None.
A matrix of size (d’ x d), where d is the number of features and d’ is the desired dimension. Ignored if dml is specified.
- dml : DML_Algorithm, default=None.
A distance metric learning algorithm. If metric, transformer and dml are None, no distances are used in the plot.
- dml_fitted : Boolean, default=True.
Specifies if the DML algorithm is already fitted. If True, the algorithm’s fit method will not be called.
- transform: Boolean, default=True.
If True, projects the data by the learned transformer and plots the transform data. Else, the classifier region will be ploted with the original data, but the regions will change according to the learned distance.
- f : The :class: ~matplotlib.figure.Figure object to paint. If None, a new object will be created.
- ax : The :class: ~matplotlib.axes.Axes object to paint. If None, a new object will be created.
- title : String, default=None. An optional title for the plot.
- subtitle : String, default=None. An optional subtitle for the plot.
- xrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the X axis. If None, it will be calculated according to the maximum and minimum of the X feature.
- yrange : List, default=None
A list with two items, specifying the minimum and maximum range to plot in the Y axis. If None, it will be calculated according to the maximum and minimum of the Y feature.
- xlabel : String, default=None. An optional title for the X axis.
- ylabel : String, default=None. An optional title for the Y axis.
- grid_split : List, default=[400, 400]
A list with two items, specifying the number of partitions, in the X and Y axis, to make in the plot to paint the classifier region. Each split will define a point where the predict method of the classifier is evaluated. It can be None. In this case, the grid_step parameter will be considered.
- grid_step : List, default=[0.1,0.1]
A list with two items, specifying the distance between the points in the grid that defines the classifier plot. Each created point in this way will define a point where the predict method of the classifier is evaluated. It is ignored if the parameter grid_split is not None.
- label_legend : Boolean, default=True. If True, a legend with the labels and its colors will be ploted.
- legend_loc : int, string of a pair of floats, default=”lower right”.
Specifies the legend position. Ignored if legend is not plotted. Allowed values are: ‘best’ (0), ‘upper right’ (1), ‘upper left’ (2), ‘lower left’ (3), ‘lower right’ (4), ‘right’ (5), ‘center left’ (6), ‘center right’ (7), ‘lower center’ (8), ‘upper center’ (9), ‘center’ (10).
Alternatively can be a 2-tuple giving x, y of the lower-left corner of the legend in axes coordinates.
- cmap : Colormap, default=None.
A
Colormap
instance or None. If cmap is None and label_colors is None, a default Colormap is used.- label_colors : List, default=None.
A list of size C with matplotlib colors, or strings specitying a color, where C is the number of classes in y. Each class will be plotted with the corresponding color. If cmap is None and label_colors is None, a default Colormap is used.
- plot_points : Boolean, default=True.
If True, points will be plotted.
- plot_regions : Boolean, default=True.
If True, the classifier regions will be plotted.
- region_intensity : Float, default=0.4.
A float between 0 and 1, indicating the transparency of the colors in the classifier regions respect the point colors.
- legend_plot_points : Boolean, default=True.
If True, points are plotted in the legend.
- legend_plot_regions : Boolean, default=True.
If True, classifier regions are plotted in the legend.
- legend_on_axis : Boolean, default=True.
If True, the legend is plotted inside the scatter plot. Else, it is plotted out of the scatter plot.
- fig_kw : dict
Additional keyword args for \(~Matplotlib.suplots\)
Returns: - f : The plotted :class: ~matplotlib.figure.Figure object.
dml.dml_utils module¶
Utility functions for different DML algoritms
-
dml.dml_utils.
SDProject
()¶ Projects a symmetric matrix onto the positive semidefinite cone (considering the Frobenius norm). The projection is made by taking the non negative eigenvalues after diagonalizing.
Parameters: - M : 2D-Array or Matrix
A symmetric matrix.
Returns: - Mplus : 2D-Array
The projection of M onto the positive semidefinite cone.
-
dml.dml_utils.
calc_outers
()¶ Calculates the outer products between two datasets. All outer products are calculated, so memory may be not enough. To avoid memory errors the output of this function should be used in the input of
calc_outers_i()
.Parameters: - X : Numpy array, shape (N x d)
A 2D-array, where N is the number of samples and d is the number of features.
- Y : Numpy array, shape (M x d), default=None
A 2D-array, where M is the number of samples in Y and d is the number of features. If None, Y is taken as X.
Returns: - outers : A 4D-array, of shape (N x M x d x d), where outers[i,j] is the outer product between X[i] and Y[j].
It can also be None, if memory was not enough. In this case, outers will be calculated in
calc_outers_i()
.
-
dml.dml_utils.
calc_outers_i
()¶ Obtains a subset of outer products from the calculated in
calc_outers()
. If memory was enough, this function just returns a row of outer products from the calculated matrix of outer products. Else, this method calculates this row.Parameters: - X : Numpy array, shape (N x d)
A 2D-array, where N is the number of samples and d is the number of features.
- outers : Numpy array, or None
The output of the function
calc_outers()
.- i : int
The row to fetch from outers, from 0 to N-1.
- Y : Numpy array, shape (M x d), default=None
A 2D-array, where M is the number of samples in Y and d is the number of features. If None, Y is taken as X.
Returns: - outers_i : A 3D-Array, of shape (M x d x d), where outers_i[j] is the outer product between X[i] and Y[j].
It can also be None, if memory was not enough. In this case, outers will be calculated in
calc_outers_ij()
.
-
dml.dml_utils.
calc_outers_ij
()¶ Obtains an outer product between two elements in datasets, from the output calculated in
calc_outers()
.Parameters: - X : Numpy array, shape (N x d)
A 2D-array, where N is the number of samples and d is the number of features.
- outers_i : Numpy array, or None
The output of the function
calc_outers_i()
.- i : int
The row to fetch from outers, from 0 to N-1.
- j : int
The column to fetch from outers, from 0 to M-1.
- Y : Numpy array, shape (M x d), default=None
A 2D-array, where M is the number of samples in Y and d is the number of features. If None, Y is taken as X.
Returns: - outers_i : A 2D-Array, of shape (d x d), with the outer product between X[i] and Y[j].
-
dml.dml_utils.
calc_regularized_outers
()¶ Calculates the outer products between two datasets. All outer products are calculated, so memory may be not enough. To avoid memory errors the output of this function should be used in the input of
calc_outers_i()
.Parameters: - X : Numpy array, shape (N x d)
A 2D-array, where N is the number of samples and d is the number of features.
- Y : Numpy array, shape (M x d), default=None
A 2D-array, where M is the number of samples in Y and d is the number of features. If None, Y is taken as X.
Returns: - outers : A 4D-array, of shape (N x M x d x d), where outers[i,j] is the outer product between X[i] and Y[j].
It can also be None, if memory was not enough. In this case, outers will be calculated in
calc_outers_i()
.
-
dml.dml_utils.
calc_regularized_outers_i
()¶ Obtains a subset of outer products from the calculated in
calc_outers()
. If memory was enough, this function just returns a row of outer products from the calculated matrix of outer products. Else, this method calculates this row.Parameters: - X : Numpy array, shape (N x d)
A 2D-array, where N is the number of samples and d is the number of features.
- outers : Numpy array, or None
The output of the function
calc_outers()
.- i : int
The row to fetch from outers, from 0 to N-1.
- Y : Numpy array, shape (M x d), default=None
A 2D-array, where M is the number of samples in Y and d is the number of features. If None, Y is taken as X.
Returns: - outers_i : A 3D-Array, of shape (M x d x d), where outers_i[j] is the outer product between X[i] and Y[j].
It can also be None, if memory was not enough. In this case, outers will be calculated in
calc_outers_ij()
.
-
dml.dml_utils.
calc_regularized_outers_ij
()¶ Obtains an outer product between two elements in datasets, from the output calculated in
calc_outers()
.Parameters: - X : Numpy array, shape (N x d)
A 2D-array, where N is the number of samples and d is the number of features.
- outers_i : Numpy array, or None
The output of the function
calc_outers_i()
.- i : int
The row to fetch from outers, from 0 to N-1.
- j : int
The column to fetch from outers, from 0 to M-1.
- Y : Numpy array, shape (M x d), default=None
A 2D-array, where M is the number of samples in Y and d is the number of features. If None, Y is taken as X.
Returns: - outers_i : A 2D-Array, of shape (d x d), with the outer product between X[i] and Y[j].
-
dml.dml_utils.
local_scaling_affinity_matrix
()¶ Local scaling affinity matrix.
Computes a local scaling affinity matrix A for the dataset (X, y), where .. math:
A[i, j] = exp(-\|x_i - x_j\|^2/\sigma_i \sigma_j).
The sigma values represent the local scaling of the samples around x_i, and are given by .. math:
\sigma_i = \|x_i - x_i^{(K)}\|,
where \(x_i^{(K)}\) is the K-th nearest neighbor of \(x_i\).
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
- k : int
The value for the k-th nearest neighbor to consider in the local scaling.
Returns: - A : 2D-array
The affinity matrix.
-
dml.dml_utils.
matpack
()¶ Returns a matrix that takes by columns the elements in the vector v.
Parameters: - v : 1D-Array
The vector to fit in a matrix.
- n : int
The matrix rows.
- m : int
The matrix columns.
Returns: - A : 2D-Array, shape (n x m)
The matrix that takes by columns the elements in v.
-
dml.dml_utils.
metric_sq_distance
()¶ Calculates a distance between two points given a metric PSD matrix.
Parameters: - M : 2D-Array or Matrix
A positive semidefinite matrix defining the distance.
- x : Array.
First argument for the distance. It must have the same length as y and the order of M.
- y : Array.
Second argument for the distance. It must have the same length as x and the order of M.
-
dml.dml_utils.
metric_to_linear
()¶ Converts a metric PSD matrix into an associated linear transformation matrix, so the distance defined by the metric matrix is the same as the euclidean distance after projecting by the linear transformation. This implementation takes the linear transformation corresponding to the square root of the matrix M.
Parameters: - M : 2D-Array or Matrix
A positive semidefinite matrix.
Returns: - L : 2D-Array
The matrix associated to the linear transformation that computes the same distance as M.
-
dml.dml_utils.
neighbors_affinity_matrix
()¶ Neighbors affinity matrix.
Computes a neighbors affinity matrix A for the dataset (X, y), where A[i, j] = 1 if x_j is a k-nearest neighbor of the same class as x_i, and 0 otherwise.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
- k : int
The number of neighbors to consider as nearest neighbors.
Returns: - A : 2D-array
The affinity matrix.
-
dml.dml_utils.
pairwise_sq_distances_from_dot
()¶ Calculates the pairwise squared distance between two datasets given the matrix of dot products.
Parameters: - K : 2D-Array or Matrix
A matrix with the dot products between two datasets. It verifies ..math:: K[i,j] = langle x_i, y_j rangle
Returns: - dists : 2D-Array
A matrix with the squared distances between the elements in both datasets. It verifies ..math:: dists[i,j] = d(x_i, y_j)
-
dml.dml_utils.
unroll
()¶ Returns a column vector from a matrix with all its columns concatenated.
Parameters: - A : 2D-Array or Matrix.
The matrix to unroll.
Returns: - v : 1D-Array
The vector with the unrolled matrix.
dml.dmlmj module¶
Distance Metric Learning through the Maximization of the Jeffrey divergence (DMLMJ)
Created on Fri Feb 23 12:34:43 2018
@author: jlsuarezdiaz
-
class
dml.dmlmj.
DMLMJ
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Distance Metric Learning through the Maximization of the Jeffrey divergence (DMLMJ).
A DML Algorithm that obtains a transformer that maximizes the Jeffrey divergence between the distribution of differences of same-class neighbors and the distribution of differences between different-class neighbors.
Parameters: - num_dims : int, default=None
Dimension desired for the transformed data.
- n_neighbors : int, default=3
Number of neighbors to consider in the computation of the difference spaces.
- alpha : float, default=0.001
Regularization parameter for inverse matrix computation.
- reg_tol : float, default=1e-10
Tolerance threshold for applying regularization. The tolerance is compared with the matrix determinant.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
num_dims : dimension of the reduced data.
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
-
class
dml.dmlmj.
KDMLMJ
¶ Bases:
dml.dml_algorithm.KernelDML_Algorithm
The kernelized version of DMLMJ.
Parameters: - num_dims : int, default=None
Dimension desired for the transformed data.
- n_neighbors : int, default=3
Number of neighbors to consider in the computation of the difference spaces.
- alpha : float, default=0.001
Regularization parameter for inverse matrix computation.
- reg_tol : float, default=1e-10
Tolerance threshold for applying regularization. The tolerance is compared with the matrix determinant.
- kernel : “linear” | “poly” | “rbf” | “sigmoid” | “cosine” | “precomputed”
Kernel. Default=”linear”.
- gamma : float, default=1/n_features
Kernel coefficient for rbf, poly and sigmoid kernels. Ignored by other kernels.
- degree : int, default=3
Degree for poly kernels. Ignored by other kernels.
- coef0 : float, default=1
Independent term in poly and sigmoid kernels. Ignored by other kernels.
- kernel_params : mapping of string to any, default=None
Parameters (keyword arguments) and values for kernel passed as callable object. Ignored by other kernels.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the kernel transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
num_dims : dimension of the reduced data.
-
transformer
¶ Obtains the learned projection.
Returns: - A : (d’x N) matrix, where d’ is the desired output dimension, and N is the number of samples.
To apply A to a new sample x, A must be multiplied by the kernel vector of dimension N obtained by taking the kernels between x and each training sample.
dml.itml module¶
Information Theoretic Metric Learning (ITML)
Created on Thu Feb 1 17:19:12 2018
@author: jlsuarezdiaz
-
class
dml.itml.
ITML
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Information Theoretic Metric Learning (ITML).
A DML algorithm that learns a metric associated to the nearest gaussian distribution satisfying similarity constraints. The nearest gaussian distribution is obtained minimizing the Kullback-Leibler divergence.
Parameters: - initial_metric : 2D-Array or Matrix
A positive definite matrix that defines the initial metric used to compare.
- upper_bound : float, default=None
Bound for dissimilarity constraints. If None, it will be estimated from upper_perc.
- lower_bound : float, default=None
Bound for similarity constraints. If None, it will be estimated from lower_perc.
- num_constraints : int, default=None
Number of constraints to generate. If None, it will be taken as 40 * k * (k-1), where k is the number of classes.
- gamma : float, default=1.0
The gamma value for slack variables.
- tol : float, default=0.001
Tolerance stop criterion for the algorithm.
- max_iter : int, default=100000
Maximum number of iterations for the algorithm.
- low_perc : int, default=5
Lower percentile (from 0 to 100) to estimate the lower bound from the dataset. Ignored if lower_bound is provided.
- up_perc : int, default=95
Upper percentile (from 0 to 100) to estimate the upper bound from the dataset. Ignored if upper_bound is provided.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metric
¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
dml.kda module¶
Kernel Discriminant Analysis (KDA)
Created on Sun Feb 18 18:38:16 2018
@author: jlsuarezdiaz
-
class
dml.kda.
KDA
¶ Bases:
dml.dml_algorithm.KernelDML_Algorithm
Kernel Discriminant Analysis (KDA)
Discriminant Analysis in high dimensionality using the kernel trick.
Parameters: - solver : string, default=’eigen’.
- Solver to use, posible values:
- ‘eigen’: Eigenvalue decomposition.
- n_components : int, default=None.
Number of components (lower than number of classes -1) for dimensionality reduction.
- tol : float, default=1e-4
Singularity toleration level.
- alpha : float, default=1e-3
Regularization term for singular within-class matrix.
- kernel : “linear” | “poly” | “rbf” | “sigmoid” | “cosine” | “precomputed”
Kernel. Default=”linear”.
- gamma : float, default=1/n_features
Kernel coefficient for rbf, poly and sigmoid kernels. Ignored by other kernels.
- degree : int, default=3
Degree for poly kernels. Ignored by other kernels.
- coef0 : float, default=1
Independent term in poly and sigmoid kernels. Ignored by other kernels.
- kernel_params : mapping of string to any, default=None
Parameters (keyword arguments) and values for kernel passed as callable object. Ignored by other kernels.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
()Obtains the algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the kernel transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
transformer
¶ Obtains the learned projection.
Returns: - A : (d’x N) matrix, where d’ is the desired output dimension, and N is the number of samples.
To apply A to a new sample x, A must be multiplied by the kernel vector of dimension N obtained by taking the kernels between x and each training sample.
dml.knn module¶
k-Nearest Neighbors (kNN)
An interface for kNN adapted to distance metric learning algorithms.
-
class
dml.knn.
kNN
(n_neighbors, dml_algorithm)¶ Bases:
object
k-Nearest Neighbors (kNN) The nearest neighbors classifier adapted to be used with distance metric learning algorithms.
Parameters: - n_neighbors : int
Number of neighbors to consider in classification.
- dml_algorithm : DML_Algorithm
The distance metric learning algorithm that will provide the distance in kNN.
Methods
fit
(X, y)Fit the model from the data in X and the labels in y. loo_pred
(X)Obtains the predicted for the given data using them as a training and with Leave One Out. loo_prob
(X)Predicts the probabilities for the given data using them as a training and with Leave One Out. loo_score
(X)Obtains the score for the given data using them as a training and with Leave One Out. predict
([X])Predicts the labels for the given data. predict_orig
([X])Predicts the labels for the given data with the Euclidean distance (with no dml transformations). predict_proba
([X])Predicts the probabilities for the given data. predict_proba_orig
([X])Predicts the probabilities for the given data with euclidean distance (with no dml transformations). score
([X, y])Obtains the classification score for the given data. score_orig
([X, y])Obtains the classification score for the given data with euclidean distance (with no dml transformation). -
fit
(X, y)¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
loo_pred
(X)¶ Obtains the predicted for the given data using them as a training and with Leave One Out.
X : 2D-Array or Matrix, default=None
The dataset to be used.Returns: - y : 1D-Array
The vector with the label predictions.
-
loo_prob
(X)¶ Predicts the probabilities for the given data using them as a training and with Leave One Out.
X : 2D-Array or Matrix, default=None
The dataset to be used.Returns: - T : 2D-Array, shape (N x c)
A matrix with the probabilities for each class. N is the number of samples and c is the number of classes. The element i, j shows the probability of sample X[i] to be in class j.
-
loo_score
(X)¶ Obtains the score for the given data using them as a training and with Leave One Out.
X : 2D-Array or Matrix, default=None
The dataset to be used.Returns: - score : float
The classification score at kNN. It is calculated as ..math:: card(y_pred == y_real) / n_samples
-
predict
(X=None)¶ Predicts the labels for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - y : 1D-Array
The vector with the label predictions.
-
predict_orig
(X=None)¶ Predicts the labels for the given data with the Euclidean distance (with no dml transformations). Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - y : 1D-Array
The vector with the label predictions.
-
predict_proba
(X=None)¶ Predicts the probabilities for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - T : 2D-Array, shape (N x c)
A matrix with the probabilities for each class. N is the number of samples and c is the number of classes. The element i, j shows the probability of sample X[i] to be in class j.
-
predict_proba_orig
(X=None)¶ Predicts the probabilities for the given data with euclidean distance (with no dml transformations). Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - T : 2D-Array, shape (N x c)
A matrix with the probabilities for each class. N is the number of samples and c is the number of classes. The element i, j shows the probability of sample X[i] to be in class j.
-
score
(X=None, y=None)¶ Obtains the classification score for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).y : 1D-Array, default=None
The real labels for the dataset. It can be None only if X is None.Returns: - score : float
The classification score at kNN. It is calculated as ..math:: card(y_pred == y_real) / n_samples
-
score_orig
(X=None, y=None)¶ Obtains the classification score for the given data with euclidean distance (with no dml transformation). Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).y : 1D-Array, default=None
The true labels for the dataset. It can be None only if X is None.Returns: - score : float
The classification score at kNN. It is calculated as ..math:: card(y_pred == y_real) / n_samples
dml.lda module¶
Linear Discriminant Analysis (LDA)
-
class
dml.lda.
LDA
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Linear Discriminant Analysis (LDA).
A distance metric learning algorithm for supervised dimensionality reduction, maximizing the ratio of variances between classes and within classes. This class is a wrapper for
LinearDiscriminantAnalysis
.Parameters: - num_dims : int, default=None
Number of components (< n_classes - 1) for dimensionality reduction. If None, it will be taken as n_classes - 1. Ignored if thres is provided.
- thres : float
Fraction of variability to keep, from 0 to 1. Data dimension will be reduced until the lowest dimension that keeps ‘thres’ explained variance.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transformer
Obtains the learned projection. transform -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
num_dims : dimension of the reduced data.
-
transform
¶
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
dml.ldml module¶
Logistic Discriminant Metric Learning (LDML)
Created on Mon Mar 12 18:26:53 2018
@author: jlsuarezdiaz
-
class
dml.ldml.
LDML
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Logistic Discriminant Metric Learning (LDML).
Distance Metric Learning through the likelihood maximization of a logistic based probability distribution.
Parameters: - num_dims : int, default=None.
Number of dimensions for dimensionality reduction. Not supported yet.
- b : float, default=1e-3
Logistic function positive threshold.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : float, default=0.3
The initial value for learning rate.
- initial_metric : 2D-Array or Matrix (d x d), or string, default=None.
If array or matrix, it must be a positive semidefinite matrix with the starting metric for gradient descent, where d is the number of features. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
- max_iter : int, default=10
Maximum number of iterations of gradient descent.
- prec : float, default=1e-3
Precision stop criterion (gradient norm).
- tol : float, default=1e-3
Tolerance stop criterion (difference between two iterations)
- descent_method : string, default=’SDP’
The descent method to use. Allowed values are:
- ‘SDP’ : semidefinite programming, consisting of gradient descent with projections onto the PSD cone.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- ‘num_iters’ : Number of iterations that the descent method took.
- ‘initial_error’ : Initial value of the objective function.
- ‘final_error’ : Final value of the objective function.
-
metric
¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
dml.llda module¶
Local Linear Discriminant Analysis (LLDA)
-
class
dml.llda.
KLLDA
¶ Bases:
dml.dml_algorithm.KernelDML_Algorithm
The kernelized version of LLDA.
Parameters: - num_dims : int, default=None
Number of components for dimensionality reduction. If None, it will be taken as n_classes - 1. Ignored if thres is provided.
- affinity : array-like or string, default=”neighbors”
The affinity matrix, that is, an (N x N) matrix with entries in [0,1], where N is the number of samples, where the (i, j) element specifies the affinity between samples x_i and x_j. It can be also a string. In this case, the affinity matrix will be computed in the algorithm. Valid strings are:
- “neighbors” : An affinity matrix A, where A[i, j] is 1 if x_j is one of the k-nearest neighbors of x_i, will be computed. The value of k
- is determined by the ‘n_neighbors’ attribute.
- “local-scaling” : An affinity matrix is computed according to the local scaling method, using the kth-nearest neighbors. The value of k
- is determined by the ‘n_neighbors’ attribute. A recommended value for this case is n_neighbors=7. See [1] for more information.
- n_neighbors : int, default=1
Number of neighbors to consider in the affinity matrix. Ignored if ‘affinity’ is not equal to “neighbors” or “local-scaling”.
- tol : float, default=1e-4
Singularity toleration level.
- alpha : float, default=1e-3
Regularization term for singular within-class matrix.
- kernel : “linear” | “poly” | “rbf” | “sigmoid” | “cosine” | “precomputed”
Kernel. Default=”linear”.
- gamma : float, default=1/n_features
Kernel coefficient for rbf, poly and sigmoid kernels. Ignored by other kernels.
- degree : int, default=3
Degree for poly kernels. Ignored by other kernels.
- coef0 : float, default=1
Independent term in poly and sigmoid kernels. Ignored by other kernels.
- kernel_params : mapping of string to any, default=None
Parameters (keyword arguments) and values for kernel passed as callable object. Ignored by other kernels.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the kernel transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
num_dims : dimension of the reduced data.
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
-
class
dml.llda.
LLDA
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Local Linear Discriminant Analysis (LDA).
A local version for the Linear Discriminant Analysis.
Parameters: - n_components : int, default=None
Number of components for dimensionality reduction. If None, it will be taken as n_classes - 1. Ignored if thres is provided.
- affinity : array-like or string, default=”neighbors”
The affinity matrix, that is, an (N x N) matrix with entries in [0,1], where N is the number of samples, where the (i, j) element specifies the affinity between samples x_i and x_j. It can be also a string. In this case, the affinity matrix will be computed in the algorithm. Valid strings are:
- “neighbors” : An affinity matrix A, where A[i, j] is 1 if x_j is one of the k-nearest neighbors of x_i, will be computed. The value of k
- is determined by the ‘n_neighbors’ attribute.
- “local-scaling” : An affinity matrix is computed according to the local scaling method, using the kth-nearest neighbors. The value of k
- is determined by the ‘n_neighbors’ attribute. A recommended value for this case is n_neighbors=7. See [1] for more information.
- n_neighbors : int, default=1
Number of neighbors to consider in the affinity matrix. Ignored if ‘affinity’ is not equal to “neighbors” or “local-scaling”.
- tol : float, default=1e-4
Singularity toleration level.
- alpha : float, default=1e-3
Regularization term for singular within-class matrix.
- solver : string, default=”sugiyama”
The resolution method. Valid values are:
- “classic” : the original LLDA problem will be computed (building the within-class and between-class matrices in the usual way).
- “sugiyama” : the algorithm proposed in [1]. It is faster than the classic method and provides the same results. The solver ‘classic’
- is kept for testing, but this solver is the recommended one.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
num_dims : dimension of the reduced data.
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
dml.lmnn module¶
Large Margin Nearest Neighbors (LMNN)
-
class
dml.lmnn.
KLMNN
¶ Bases:
dml.dml_algorithm.KernelDML_Algorithm
The kernelized version of LMNN.
Parameters: - num_dims : int, default=None
Desired value for dimensionality reduction. Ignored if solver is ‘SDP’.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : float, default=0.3
The initial value for learning rate.
- initial_metric : 2D-Array or Matrix (d’ x d), or string, default=None.
If array or matrix, and solver is SDP, it must be a positive semidefinite matrix with the starting metric (d x d) for gradient descent, where d is the number of features. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
If solver is SGD, then the array or matrix will represent a linear map (d’ x d), where d’ is the dimension provided in num_dims.
- max_iter : int, default=100
Maximum number of iterations of gradient descent.
- prec : float, default=1e-8
Precision stop criterion (gradient norm).
- tol : float, default=1e-8
Tolerance stop criterion (difference between two iterations)
- k : int, default=3
Number of target neighbors to take. If this algorithm is used for nearest neighbors classification, a good choice is to take k as the number of neighbors.
- mu : float, default=0.5
The weight of the push error in the minimization algorithm. The objective function is composed of a push error, given by the impostors, with weight mu, and a pull error, given by the target neighbors, with weight (1-mu). It must be between 0.0 and 1.0.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- kernel : “linear” | “poly” | “rbf” | “sigmoid” | “cosine” | “precomputed”
Kernel. Default=”linear”.
- gamma : float, default=1/n_features
Kernel coefficient for rbf, poly and sigmoid kernels. Ignored by other kernels.
- degree : int, default=3
Degree for poly kernels. Ignored by other kernels.
- coef0 : float, default=1
Independent term in poly and sigmoid kernels. Ignored by other kernels.
- kernel_params : mapping of string to any, default=None
Parameters (keyword arguments) and values for kernel passed as callable object. Ignored by other kernels.
- target_selecion : string, default=’kernel’
How to find the target neighbors. Allowed values are:
- ‘kernel’ : using the euclidean distance in the kernel space.
- ‘original’ : using the euclidean distance in the original space.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the kernel transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- ‘num_iters’ : Number of iterations that the descent method took.
- ‘initial_error’ : Initial value of the objective function.
- ‘final_error’ : Final value of the objective function.
-
transformer
¶ Obtains the learned projection.
Returns: - A : (d’x N) matrix, where d’ is the desired output dimension, and N is the number of samples.
To apply A to a new sample x, A must be multiplied by the kernel vector of dimension N obtained by taking the kernels between x and each training sample.
-
class
dml.lmnn.
LMNN
¶ Bases:
dml.dml_algorithm.DML_Algorithm
,sklearn.base.ClassifierMixin
Large Margin Nearest Neighbors (LMNN)
A distance metric learning algorithm that obtains a metric with target neighbors as near as possible and impostors as far as possible
Parameters: - num_dims : int, default=None
Desired value for dimensionality reduction. Ignored if solver is ‘SDP’. If NULL, all features will be kept.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : int, default=0.3
The initial value for learning rate. If solver is ‘SGD’, default value may be too large. In this case it is recommended to use a learning_rate of an order of 1e-3 instead.
- initial_metric : 2D-Array or Matrix (d’ x d), or string, default=None.
If array or matrix, and solver is SDP, it must be a positive semidefinite matrix with the starting metric (d x d) for gradient descent, where d is the number of features. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
If solver is SGD, then the array or matrix will represent a linear map (d’ x d), where d’ is the dimension provided in num_dims.
- max_iter : int, default=100
Maximum number of iterations of gradient descent.
- prec : float, default=1e-8
Precision stop criterion (gradient norm).
- tol : float, default=1e-8
Tolerance stop criterion (difference between two iterations)
- k : int, default=3
Number of target neighbors to take. If this algorithm is used for nearest neighbors classification, a good choice is to take k as the number of neighbors.
- mu : float, default=0.5
The weight of the push error in the minimization algorithm. The objective function is composed of a push error, given by the impostors, with weight mu, and a pull error, given by the target neighbors, with weight (1-mu). It must be between 0.0 and 1.0.
- soft_comp_interval : int, default=1
Intervals of soft computation. The soft computation relaxes the gradient descent conditions, but makes the algorithm more efficient. This value provides the length of a soft computation interval. After soft_comp_interval iterations of gradient descent, a complete gradient step is performed.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- solver : string, default=’SDP’
The algorithm used for minimization. Allowed values are:
- ‘SDP’ : semidefinite programming, consisting of gradient descent with projections onto the positive semidefinite cone.
- It learns a metric.
- ‘SGD’ : stochastic gradient descent. It learns a linear transformer.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. predict
Predict the class labels for the provided data, according to the LMNN energy method. score
(X, y[, sample_weight])Return the mean accuracy on the given test data and labels. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- num_iters : Number of iterations that the descent method took.
- initial_error : Initial value of the objective function.
- final_error : Final value of the objective function.
-
predict
¶ Predict the class labels for the provided data, according to the LMNN energy method.
Parameters: - X : array-like, shape (N x d)
Test samples. N is the number of samples and d the number of features. If None, training set will be used.
Returns: - y : array of shape (N)
Class labels for each data sample.
dml.lsi module¶
Learning with Side Information (LSI)
-
class
dml.lsi.
LSI
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Learning with Side Information (LSI)
A distance metric learning algorithm that minimizes the sum of distances between similar data, with non similar data constrained to be separated.
Parameters: - initial_metric : 2D-Array or Matrix (d x d), or string, default=None.
If array or matrix, it must be a positive semidefinite matrix with the starting metric for gradient descent, where d is the number of features. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : float, default=0.1
The initial value for learning rate.
- max_iter : int, default=100
Number of iterations for gradient descent.
- max_proj_iter : int, default=5000
Number of iterations for iterated projections.
- itproj_err : float, default=1e-3
Convergence error criterion for iterated projections
- err : float, default=1e-3
Convergence error stop criterion for gradient descent.
- supervised : Boolean, default=False
If True, the algorithm will accept a labeled dataset (X,y). Else, it will accept the dataset and the similarity sets, (X,S,D).
Methods
fit
Fit the model from the data in X and the side information in side fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. fD fD1 fS fS1 grad_projection label_to_similarity_set -
fD
¶
-
fD1
¶
-
fS
¶
-
fS1
¶
-
fit
¶ Fit the model from the data in X and the side information in side
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- side : list of array-like, or 1D-array (N)
- The side information, or the label set. Options:
- side = y, the label set (only if supervised = True)
- side = [S,D], where S is the set of indices of similar data and D is the set of indices of dissimilar data.
- side = [S], where S is the set of indices of similar data. The set D will be the complement of S.
Sets S and D are represented as a boolean matrix (S[i,j]==True iff (i,j) in S)
- Returns
- ——-
- self : object
Returns the instance itself.
-
grad_projection
¶
-
label_to_similarity_set
¶
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- ‘initial_objective’ : Initial value of the objective function.
- ‘initial_constraint’ : Initial calue of the constraint function.
- ‘final_objective’ : Final value of the objective function.
- ‘final_constraint’ : Final value of the constraint function.
- ‘iterative_projections_conv_exp’ : Convergence ratio, from 0 to 1, of the iterative projections.
- ‘projection_iterations_avg’ : Average iterations needed in iterative projections.
- ‘num_its’ : Number of iterations of gradient descent.
-
metric
¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
dml.mcml module¶
Maximally collapsing metric learning (MCML)
Created on Mon Mar 12 10:47:23 2018
@author: jlsuarezdiaz
-
class
dml.mcml.
MCML
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Maximally Collapsing Metric Learning (MCML)
A distance metric learning algorithm that learns minimizing the KL divergence to the maximally collapsing distribution.
Parameters: - num_dims : int, default=None.
Number of dimensions for dimensionality reduction. Not supported yet.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : float, default=0.01
The initial value for learning rate.
- initial_metric : 2D-Array or Matrix (d x d), or string, default=None.
If array or matrix, it must be a positive semidefinite matrix with the starting metric for gradient descent, where d is the number of features. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
- max_iter : int, default=20
Maximum number of iterations of gradient descent.
- prec : float, default=1e-3
Precision stop criterion (gradient norm).
- tol : float, default=1e-3
Tolerance stop criterion (difference between two iterations)
- descent_method : string, default=’SDP’
The descent method to use. Allowed values are:
- ‘SDP’ : semidefinite programming, consisting of gradient descent with projections onto the PSD cone.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
Obtains the learned metric. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
()Computes a transformation matrix from the Mahalanobis matrix. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- ‘num_iters’ : Number of iterations that the descent method took.
- ‘initial_error’ : Initial value of the objective function.
- ‘final_error’ : Final value of the objective function.
-
metric
¶ Obtains the learned metric.
Returns: - M : (dxd) positive semidefinite matrix, where d is the number of features.
dml.multidml_knn module¶
Multiple-DML k-Nearest Neighbors (kNN)
-
class
dml.multidml_knn.
MultiDML_kNN
(n_neighbors, dmls=None, verbose=False, **knn_args)¶ Bases:
object
Multi-DML k-NN
An interface that allows learning k-NN with different distance metric learners simultaneously.
Parameters: - n_neighbors : int
The number of neighbors for k-NN.
- dmls : list, default=None
A list of distance metric learning algorithms to be learned for k-NN. By default, euclidean distance will be added at the first place of the dml list.
- verbose : boolean, default=False
If True, console message about the algorithms execution will be printed.
Methods
add
(dmls)Adds a new distance metric learning algorithm to the list. dmls_string
()Obtains the strings with the dml names. elapsed
()Obtains the elapsed time of each DML algorithm fit
(X, y)Fit the model from the data in X and the labels in y. predict_all
([X])Predicts the labels for the given data. predict_proba_all
([X])Predicts the probabilities for the given data. score_all
([X, y])Obtains the scores for the given data. -
add
(dmls)¶ Adds a new distance metric learning algorithm to the list.
Parameters: - dmls : DML_Algorithm, or list of DMÑ_Algorithm
The DML algorithm or algorithms to add.
-
dmls_string
()¶ Obtains the strings with the dml names.
Returns: - strings : A list with the names of each dml.
-
elapsed
()¶ Obtains the elapsed time of each DML algorithm
Returns: - elapsed : A list of float with the time of each DML.
-
fit
(X, y)¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
predict_all
(X=None)¶ Predicts the labels for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - y : list of 1D-Arrays
A list with the vectors with the label predictions for each DML.
-
predict_proba_all
(X=None)¶ Predicts the probabilities for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - T : list of 2D-Arrays
A list with the matrices with the label probabilities for each class, for each DML.
-
score_all
(X=None, y=None)¶ Obtains the scores for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - s : list of float
A list with the k-NN scores for each DML.
dml.nca module¶
Neighbourhood Component Analysis (NCA)
-
class
dml.nca.
NCA
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Neighborhood Component Analysis (NCA)
A distance metric learning algorithm that tries to minimize kNN expected error.
Parameters: - num_dims : int, default=None
Desired value for dimensionality reduction. If None, the dimension of transformed data will be the same as the original.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : int, default=0.3
The initial value for learning rate.
- initial_transform : 2D-Array or Matrix (d’ x d), or string, default=None.
If array or matrix that will represent the starting linear map for gradient descent, where d is the number of features, and d’ is the dimension specified in num_dims. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
- max_iter : int, default=100
Maximum number of gradient descent iterations.
- prec : float, default=1e-8
Precision stop criterion (gradient norm).
- tol : float, default=1e-8
Tolerance stop criterion (difference between two iterations)
- descent_method : string, default=’SGD’
The descent method to use. Allowed values are:
- ‘SGD’ : stochastic gradient descent.
- ‘BGD’ : batch gradient descent.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- num_iters : Number of iterations that the descent method took.
- initial_expectance : Initial value of the objective function (the expected LOO score)
- final_expectance : Final value of the objective function (the expected LOO score)
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
dml.ncmc module¶
Nearest Class with Multiple Centroids (NCMC)
Created on Wed Feb 28 16:18:39 2018
@author: jlsuarezdiaz
-
class
dml.ncmc.
NCMC
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Nearest Class with Multiple Centroids distance metric learner (NCMC).
A distance metric learning algorithm to improve the nearest class with multiple centroids classifier.
Parameters: - num_dims : int, default=None
Desired value for dimensionality reduction. If None, the dimension of transformed data will be the same as the original.
- centroids_num : int, or list of int, default=3
If it is a list, it must have the same size as the number of classes. In this case, i-th item will be the number of centroids to take in the i-th class. If it is an int, every class will have the same number of centroids.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : int, default=0.3
The initial value for learning rate.
- initial_transform : 2D-Array or Matrix (d’ x d), or string, default=None.
If array or matrix that will represent the starting linear map for gradient descent, where d is the number of features, and d’ is the dimension specified in num_dims. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
- max_iter : int, default=300
Maximum number of gradient descent iterations.
- prec : float, default=1e-15
Precision stop criterion (gradient norm).
- tol : float, default=1e-15
Tolerance stop criterion (difference between two iterations)
- descent_method : string, default=’SGD’
The descent method to use. Allowed values are:
- ‘SGD’ : stochastic gradient descent.
- ‘BGD’ : batch gradient descent.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- num_iters : Number of iterations that the descent method took.
- initial_expectance : Initial value of the objective function (the expected score)
- final_expectance : Final value of the objective function (the expected score)
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
-
class
dml.ncmc.
NCMC_Classifier
¶ Bases:
sklearn.base.BaseEstimator
,sklearn.base.ClassifierMixin
Nearest Class with Multiple Centroids classifier.
A classifier that makes its predictions by choosing the class who has a centroid the nearest to the point. For each class, an arbitrary number of centroids can be set. This centroids are calculated using k-Means over each class sub-dataset.
Parameters: - centroids_num : int, or list of int, default=3
If it is a list, it must have the same size as the number of classes. In this case, i-th item will be the number of centroids to take in the i-th class. If it is an int, every class will have the same number of centroids.
- kmeans_args : dictionary
Additional keyword args for k-Means.
Methods
fit
Fit the model from the data in X and the labels in y. get_params
([deep])Get parameters for this estimator. predict
Predicts the labels for the given data. score
(X, y[, sample_weight])Return the mean accuracy on the given test data and labels. set_params
(**params)Set the parameters of this estimator. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
predict
¶ Predicts the labels for the given data. Model needs to be already fitted.
X : 2D-Array or Matrix, default=None
The dataset to be used. If None, the training set will be used. In this case, the prediction will be made using Leave One Out (that is, the sample to predict will be taken away from the training set).Returns: - y : 1D-Array
The vector with the label predictions.
dml.ncmml module¶
Nearest Class Mean Metric Learning (NCMML)
Created on Wed Feb 28 12:07:43 2018
@author: jlsuarezdiaz
-
class
dml.ncmml.
NCMML
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Nearest Class Mean Metric Learning (NCMML)
A distance metric learning algorithm to improve the nearest class mean classifier.
Parameters: - num_dims : int, default=None
Desired value for dimensionality reduction. If None, the dimension of transformed data will be the same as the original.
- learning_rate : string, default=’adaptive’
Type of learning rate update for gradient descent. Possible values are:
- ‘adaptive’ : the learning rate will increase if the gradient step is succesful, else it will decrease.
- ‘constant’ : the learning rate will be constant during all the gradient steps.
- eta0 : int, default=0.3
The initial value for learning rate.
- initial_transform : 2D-Array or Matrix (d’ x d), or string, default=None.
If array or matrix that will represent the starting linear map for gradient descent, where d is the number of features, and d’ is the dimension specified in num_dims. If None, euclidean distance will be used. If a string, the following values are allowed:
- ‘euclidean’ : the euclidean distance.
- ‘scale’ : a diagonal matrix that normalizes each attribute according to its range will be used.
- max_iter : int, default=300
Maximum number of gradient descent iterations.
- prec : float, default=1e-15
Precision stop criterion (gradient norm).
- tol : float, default=1e-15
Tolerance stop criterion (difference between two iterations)
- descent_method : string, default=’SGD’
The descent method to use. Allowed values are:
- ‘SGD’ : stochastic gradient descent.
- ‘BGD’ : batch gradient descent.
- eta_thres : float, default=1e-14
A learning rate threshold stop criterion.
- learn_inc : float, default=1.01
Increase factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
- learn_dec : float, default=0.5
Decrease factor for learning rate. Ignored if learning_rate is not ‘adaptive’.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
([X])Applies the metric transformation. transformer
Obtains algorithm metadata. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- num_iters : Number of iterations that the descent method took.
- initial_expectance : Initial value of the objective function (the expected score)
- final_expectance : Final value of the objective function (the expected score)
-
transformer
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
- num_iters : Number of iterations that the descent method took.
- initial_expectance : Initial value of the objective function (the expected score)
- final_expectance : Final value of the objective function (the expected score)
dml.pca module¶
Principal Component Analysis (PCA)
-
class
dml.pca.
PCA
¶ Bases:
dml.dml_algorithm.DML_Algorithm
Principal Component Analysis (PCA)
A distance metric learning algorithm for unsupervised dimensionality reduction, obtaining orthogonal directions that maximize the variance. This class is a wrapper for
PCA
.Parameters: - num_dims : int, default=None
Number of components for dimensionality reduction. If None, all the principal components will be taken. Ignored if thres is provided.
- thres : float
Fraction of variability to keep, from 0 to 1. Data dimension will be reduced until the lowest dimension that keeps ‘thres’ explained variance.
Methods
fit
Fit the model from the data in X and the labels in y. fit_transform
(X[, y])Fit to data, then transform it. get_params
([deep])Get parameters for this estimator. metadata
Obtains algorithm metadata. metric
()Computes the Mahalanobis matrix from the transformation matrix. set_params
(**params)Set the parameters of this estimator. transform
Applies the kernel transformation. transformer
Obtains the learned projection. -
fit
¶ Fit the model from the data in X and the labels in y.
Parameters: - X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
Returns: - self : object
Returns the instance itself.
-
metadata
¶ Obtains algorithm metadata.
Returns: - meta : A dictionary with the following metadata:
num_dims : dimension of the reduced data.
acum_eig : eigenvalue rate accumulated in the learned output respect to the total dimension.
-
transform
¶ Applies the kernel transformation.
Parameters: - X : (N x d) matrix, optional
Data to transform. If not supplied, the training data will be used.
Returns: - transformed: (N x d’) matrix.
Input data transformed by the learned mapping.
-
transformer
¶ Obtains the learned projection.
Returns: - L : (d’xd) matrix, where d’ is the desired output dimension and d is the number of features.
dml.tune module¶
Tune utilities for distance metric learning.
Created on Fri Feb 9 19:29:06 2018
@author: jlsuarezdiaz
-
dml.tune.
cross_validate
(alg, X, y, n_folds=5, n_reps=1, verbose=False, seed=None)¶ Cross validation for a classifier.
Parameters: - alg : object.
A classifier. It must support the methods fit(X,y) and score(X,y), as specified in
ClassifierMixin
.- X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
- n_folds : int, default=5
Number of folds for cross validation.
- n_reps : int, default=1
Number of cross validations to do.
- verbose : boolean, default=False
If True, a console log will be printed.
- seed : int, RandomState instance or None, optional, default=None
If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by np.random. Used when
shuffle
== True.
Returns: - results : Pandas Dataframe
A matrix whose rows represent each fold of the cross validation, including also the mean and the std. The columns represent the score, the fit time and the predict time of the classifier.
-
dml.tune.
metadata_cross_validate
(dml, X, y, metrics, n_folds=5, n_reps=1, verbose=False, seed=None, **knn_args)¶ Cross validation for distance metric learning algorithms using metadata as metrics.
Parameters: - dml : DML_Algorithm
The distance metric learning algorithm to tune.
- X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
- metrics : list of string and int
The metrics to evaluate. If string, it must be a key of the metadata() function of the DML Algorithm, or ‘time’. In this last case, the elapsed fitting time will be returned as a metric. If int, the metric will be the k-NN score, where k is the specified int.
- n_folds : int, default=5
Number of folds for cross validation.
- n_reps : int, default=1
Number of cross validations to do.
- verbose : boolean, default=False
If True, a console log will be printed.
- seed : int, RandomState instance or None, optional, default=None
If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by np.random. Used when
shuffle
== True.- knn_args : dictionary.
Additional keyword arguments for k-NN.
Returns: - results : Pandas Dataframe
A matrix whose rows represent each fold of the cross validation, including also the mean and the std. The columns represent the scores of each of the metrics specified.
-
dml.tune.
tune
(dml, X, y, dml_params, tune_args, metrics, n_folds=5, n_reps=1, verbose=False, seed=None, **knn_args)¶ Tune function for a distance metric learning algorithm, allowing as metrics the algorithm metadata, times and k-NN scores.
Parameters: - dml : A DML_Algorithm subclass
The distance metric algorithm class to tune.
- X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
- dml_params : dictionary
Additional keyword parameters for the distance metric learning algorithm.
- tune_args : dictionary
Parameters of the DML algorithm to tune. Each key has to be a keyword argument of the DML. The associated values have to be lists containing all the desired values for the tuning parameters.
- metrics : list of string and int
The metrics to evaluate. If string, it must be a key of the metadata() function of the DML Algorithm, or ‘time’. In this last case, the elapsed fitting time will be returned as a metric. If int, the metric will be the k-NN score, where k is the specified int.
- n_folds : int, default=5
Number of folds for cross validation.
- n_reps : int, default=1
Number of cross validations to do.
- verbose : boolean, default=False
If True, a console log will be printed.
- seed : int, RandomState instance or None, optional, default=None
If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by np.random. Used when
shuffle
== True.- knn_args : dictionary.
Additional keyword arguments for k-NN.
Returns: - tune_results : Pandas Dataframe
A dataframe whose entries are all the cases considered for the tune parameters, and with a single column that shows the cross validation score for each case.
- best_performance : Tuple
A pair with the best case obtained, together with its corresponding score.
- best_dml : DML_Algorithm
The DML Algorithm object that obtained the best result in the tuning.
- detailed_results : Dictionary
A dictionary whose keys are all the possible cases, and each entry is the cross validation table for the corresponding case, containing the scores for every fold.
-
dml.tune.
tune_knn
(dml, X, y, n_neighbors, dml_params, tune_args, n_folds=5, n_reps=1, verbose=False, seed=None, **knn_args)¶ A tune function for a distance metric learning algorithm, using k-NN score as metric.
Parameters: - dml : A DML_Algorithm subclass
The distance metric algorithm class to tune.
- X : array-like, shape (N x d)
Training vector, where N is the number of samples, and d is the number of features.
- y : array-like, shape (N)
Labels vector, where N is the number of samples.
- n_neighbors : int
Number of neighbors for k-NN.
- dml_params : dictionary
Additional keyword parameters for the distance metric learning algorithm.
- tune_args : dictionary
Parameters of the DML algorithm to tune. Each key has to be a keyword argument of the DML. The associated values have to be lists containing all the desired values for the tuning parameters.
- n_folds : int, default=5
Number of folds for cross validation.
- n_reps : int, default=1
Number of cross validations to do.
- verbose : boolean, default=False
If True, a console log will be printed.
- seed : int, RandomState instance or None, optional, default=None
If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by np.random. Used when
shuffle
== True.- knn_args : dictionary.
Additional keyword arguments for k-NN.
Returns: - tune_results : Pandas Dataframe
A dataframe whose entries are all the cases considered for the tune parameters, and with a single column that shows the cross validation score for each case.
- best_performance : Tuple
A pair with the best case obtained, together with its corresponding score.
- best_dml : DML_Algorithm
The DML Algorithm object that obtained the best result in the tuning.
- detailed_results : Dictionary
A dictionary whose keys are all the possible cases, and each entry is the cross validation table for the corresponding case, containing the scores for every fold.
Module contents¶
The Distance Metric Learning module.