diff options
Diffstat (limited to 'src/python/gudhi/representations/kernel_methods.py')
| -rw-r--r-- | src/python/gudhi/representations/kernel_methods.py | 200 |
1 files changed, 148 insertions, 52 deletions
diff --git a/src/python/gudhi/representations/kernel_methods.py b/src/python/gudhi/representations/kernel_methods.py index bfc83aff..23fd23c7 100644 --- a/src/python/gudhi/representations/kernel_methods.py +++ b/src/python/gudhi/representations/kernel_methods.py @@ -9,27 +9,100 @@ import numpy as np from sklearn.base import BaseEstimator, TransformerMixin -from sklearn.metrics import pairwise_distances -from .metrics import SlicedWassersteinDistance, PersistenceFisherDistance +from sklearn.metrics import pairwise_distances, pairwise_kernels +from .metrics import SlicedWassersteinDistance, PersistenceFisherDistance, _sklearn_wrapper, _pairwise, pairwise_persistence_diagram_distances, _sliced_wasserstein_distance, _persistence_fisher_distance +from .preprocessing import Padding ############################################# # Kernel methods ############################ ############################################# +def _persistence_weighted_gaussian_kernel(D1, D2, weight=lambda x: 1, kernel_approx=None, bandwidth=1.): + """ + This is a function for computing the persistence weighted Gaussian kernel value from two persistence diagrams. The persistence weighted Gaussian kernel is computed by convolving the persistence diagram points with weighted Gaussian kernels. See http://proceedings.mlr.press/v48/kusano16.html for more details. + + Parameters: + D1: (n x 2) numpy.array encoding the (finite points of the) first diagram. Must not contain essential points (i.e. with infinite coordinate). + D2: (m x 2) numpy.array encoding the second diagram. + bandwidth (double): bandwidth of the Gaussian kernel with which persistence diagrams will be convolved + weight: weight function for the persistence diagram points (default constant function, ie lambda x: 1). This function must be defined on 2D points, ie lists or numpy arrays of the form [p_x,p_y]. + kernel_approx: kernel approximation class used to speed up computation. Common kernel approximations classes can be found in the scikit-learn library (such as RBFSampler for instance). + + Returns: + float: the persistence weighted Gaussian kernel value between persistence diagrams. + """ + ws1 = np.array([weight(D1[j,:]) for j in range(len(D1))]) + ws2 = np.array([weight(D2[j,:]) for j in range(len(D2))]) + if kernel_approx is not None: + approx1 = np.sum(np.multiply(ws1[:,np.newaxis], kernel_approx.transform(D1)), axis=0) + approx2 = np.sum(np.multiply(ws2[:,np.newaxis], kernel_approx.transform(D2)), axis=0) + return (1./(np.sqrt(2*np.pi)*bandwidth)) * np.matmul(approx1, approx2.T) + else: + W = np.matmul(ws1[:,np.newaxis], ws2[np.newaxis,:]) + E = (1./(np.sqrt(2*np.pi)*bandwidth)) * np.exp(-np.square(pairwise_distances(D1,D2))/(2*bandwidth*bandwidth)) + return np.sum(np.multiply(W, E)) + +def _persistence_scale_space_kernel(D1, D2, kernel_approx=None, bandwidth=1.): + """ + This is a function for computing the persistence scale space kernel value from two persistence diagrams. The persistence scale space kernel is computed by adding the symmetric to the diagonal of each point in each persistence diagram, with negative weight, and then convolving the points with a Gaussian kernel. See https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Reininghaus_A_Stable_Multi-Scale_2015_CVPR_paper.pdf for more details. + + Parameters: + D1: (n x 2) numpy.array encoding the (finite points of the) first diagram. Must not contain essential points (i.e. with infinite coordinate). + D2: (m x 2) numpy.array encoding the second diagram. + bandwidth (double): bandwidth of the Gaussian kernel with which persistence diagrams will be convolved + kernel_approx: kernel approximation class used to speed up computation. Common kernel approximations classes can be found in the scikit-learn library (such as RBFSampler for instance). + + Returns: + float: the persistence scale space kernel value between persistence diagrams. + """ + DD1 = np.concatenate([D1, D1[:,[1,0]]], axis=0) + DD2 = np.concatenate([D2, D2[:,[1,0]]], axis=0) + weight_pss = lambda x: 1 if x[1] >= x[0] else -1 + return 0.5 * _persistence_weighted_gaussian_kernel(DD1, DD2, weight=weight_pss, kernel_approx=kernel_approx, bandwidth=bandwidth) + +def pairwise_persistence_diagram_kernels(X, Y=None, kernel="sliced_wasserstein", n_jobs=None, **kwargs): + """ + This function computes the kernel matrix between two lists of persistence diagrams given as numpy arrays of shape (nx2). + + Parameters: + X (list of n numpy arrays of shape (numx2)): first list of persistence diagrams. + Y (list of m numpy arrays of shape (numx2)): second list of persistence diagrams (optional). If None, pairwise kernel values are computed from the first list only. + kernel: kernel to use. It can be either a string ("sliced_wasserstein", "persistence_scale_space", "persistence_weighted_gaussian", "persistence_fisher") or a function taking two numpy arrays of shape (nx2) and (mx2) as inputs. If it is a function, make sure that it is symmetric. + n_jobs (int): number of jobs to use for the computation. This uses joblib.Parallel(prefer="threads"), so kernels that do not release the GIL may not scale unless run inside a `joblib.parallel_backend <https://joblib.readthedocs.io/en/latest/parallel.html#joblib.parallel_backend>`_ block. + **kwargs: optional keyword parameters. Any further parameters are passed directly to the kernel function. See the docs of the various kernel classes in this module. + + Returns: + numpy array of shape (nxm): kernel matrix. + """ + XX = np.reshape(np.arange(len(X)), [-1,1]) + YY = None if Y is None or Y is X else np.reshape(np.arange(len(Y)), [-1,1]) + if kernel == "sliced_wasserstein": + return np.exp(-pairwise_persistence_diagram_distances(X, Y, metric="sliced_wasserstein", num_directions=kwargs["num_directions"], n_jobs=n_jobs) / kwargs["bandwidth"]) + elif kernel == "persistence_fisher": + return np.exp(-pairwise_persistence_diagram_distances(X, Y, metric="persistence_fisher", kernel_approx=kwargs["kernel_approx"], bandwidth=kwargs["bandwidth"], n_jobs=n_jobs) / kwargs["bandwidth_fisher"]) + elif kernel == "persistence_scale_space": + return _pairwise(pairwise_kernels, False, XX, YY, metric=_sklearn_wrapper(_persistence_scale_space_kernel, X, Y, **kwargs), n_jobs=n_jobs) + elif kernel == "persistence_weighted_gaussian": + return _pairwise(pairwise_kernels, False, XX, YY, metric=_sklearn_wrapper(_persistence_weighted_gaussian_kernel, X, Y, **kwargs), n_jobs=n_jobs) + else: + return _pairwise(pairwise_kernels, False, XX, YY, metric=_sklearn_wrapper(metric, **kwargs), n_jobs=n_jobs) + class SlicedWassersteinKernel(BaseEstimator, TransformerMixin): """ This is a class for computing the sliced Wasserstein kernel matrix from a list of persistence diagrams. The sliced Wasserstein kernel is computed by exponentiating the corresponding sliced Wasserstein distance with a Gaussian kernel. See http://proceedings.mlr.press/v70/carriere17a.html for more details. """ - def __init__(self, num_directions=10, bandwidth=1.0): + def __init__(self, num_directions=10, bandwidth=1.0, n_jobs=None): """ Constructor for the SlicedWassersteinKernel class. Parameters: bandwidth (double): bandwidth of the Gaussian kernel applied to the sliced Wasserstein distance (default 1.). num_directions (int): number of lines evenly sampled from [-pi/2,pi/2] in order to approximate and speed up the kernel computation (default 10). + n_jobs (int): number of jobs to use for the computation. See :func:`pairwise_persistence_diagram_kernels` for details. """ self.bandwidth = bandwidth - self.sw_ = SlicedWassersteinDistance(num_directions=num_directions) + self.num_directions = num_directions + self.n_jobs = n_jobs def fit(self, X, y=None): """ @@ -39,7 +112,7 @@ class SlicedWassersteinKernel(BaseEstimator, TransformerMixin): X (list of n x 2 numpy arrays): input persistence diagrams. y (n x 1 array): persistence diagram labels (unused). """ - self.sw_.fit(X, y) + self.diagrams_ = X return self def transform(self, X): @@ -52,13 +125,26 @@ class SlicedWassersteinKernel(BaseEstimator, TransformerMixin): Returns: numpy array of shape (number of diagrams in **diagrams**) x (number of diagrams in X): matrix of pairwise sliced Wasserstein kernel values. """ - return np.exp(-self.sw_.transform(X)/self.bandwidth) + return pairwise_persistence_diagram_kernels(X, self.diagrams_, kernel="sliced_wasserstein", bandwidth=self.bandwidth, num_directions=self.num_directions, n_jobs=self.n_jobs) + + def __call__(self, diag1, diag2): + """ + Apply SlicedWassersteinKernel on a single pair of persistence diagrams and outputs the result. + + Parameters: + diag1 (n x 2 numpy array): first input persistence diagram. + diag2 (n x 2 numpy array): second input persistence diagram. + + Returns: + float: sliced Wasserstein kernel value. + """ + return np.exp(-_sliced_wasserstein_distance(diag1, diag2, num_directions=self.num_directions)) / self.bandwidth class PersistenceWeightedGaussianKernel(BaseEstimator, TransformerMixin): """ This is a class for computing the persistence weighted Gaussian kernel matrix from a list of persistence diagrams. The persistence weighted Gaussian kernel is computed by convolving the persistence diagram points with weighted Gaussian kernels. See http://proceedings.mlr.press/v48/kusano16.html for more details. """ - def __init__(self, bandwidth=1., weight=lambda x: 1, kernel_approx=None): + def __init__(self, bandwidth=1., weight=lambda x: 1, kernel_approx=None, n_jobs=None): """ Constructor for the PersistenceWeightedGaussianKernel class. @@ -66,9 +152,11 @@ class PersistenceWeightedGaussianKernel(BaseEstimator, TransformerMixin): bandwidth (double): bandwidth of the Gaussian kernel with which persistence diagrams will be convolved (default 1.) weight (function): weight function for the persistence diagram points (default constant function, ie lambda x: 1). This function must be defined on 2D points, ie lists or numpy arrays of the form [p_x,p_y]. kernel_approx (class): kernel approximation class used to speed up computation (default None). Common kernel approximations classes can be found in the scikit-learn library (such as RBFSampler for instance). + n_jobs (int): number of jobs to use for the computation. See :func:`pairwise_persistence_diagram_kernels` for details. """ self.bandwidth, self.weight = bandwidth, weight self.kernel_approx = kernel_approx + self.n_jobs = n_jobs def fit(self, X, y=None): """ @@ -78,10 +166,7 @@ class PersistenceWeightedGaussianKernel(BaseEstimator, TransformerMixin): X (list of n x 2 numpy arrays): input persistence diagrams. y (n x 1 array): persistence diagram labels (unused). """ - self.diagrams_ = list(X) - self.ws_ = [ np.array([self.weight(self.diagrams_[i][j,:]) for j in range(self.diagrams_[i].shape[0])]) for i in range(len(self.diagrams_)) ] - if self.kernel_approx is not None: - self.approx_ = np.concatenate([np.sum(np.multiply(self.ws_[i][:,np.newaxis], self.kernel_approx.transform(self.diagrams_[i])), axis=0)[np.newaxis,:] for i in range(len(self.diagrams_))]) + self.diagrams_ = X return self def transform(self, X): @@ -94,45 +179,36 @@ class PersistenceWeightedGaussianKernel(BaseEstimator, TransformerMixin): Returns: numpy array of shape (number of diagrams in **diagrams**) x (number of diagrams in X): matrix of pairwise persistence weighted Gaussian kernel values. """ - Xp = list(X) - Xfit = np.zeros((len(Xp), len(self.diagrams_))) - if len(self.diagrams_) == len(Xp) and np.all([np.array_equal(self.diagrams_[i], Xp[i]) for i in range(len(Xp))]): - if self.kernel_approx is not None: - Xfit = (1./(np.sqrt(2*np.pi)*self.bandwidth)) * np.matmul(self.approx_, self.approx_.T) - else: - for i in range(len(self.diagrams_)): - for j in range(i+1, len(self.diagrams_)): - W = np.matmul(self.ws_[i][:,np.newaxis], self.ws_[j][np.newaxis,:]) - E = (1./(np.sqrt(2*np.pi)*self.bandwidth)) * np.exp(-np.square(pairwise_distances(self.diagrams_[i], self.diagrams_[j]))/(2*np.square(self.bandwidth))) - Xfit[i,j] = np.sum(np.multiply(W, E)) - Xfit[j,i] = Xfit[i,j] - else: - ws = [ np.array([self.weight(Xp[i][j,:]) for j in range(Xp[i].shape[0])]) for i in range(len(Xp)) ] - if self.kernel_approx is not None: - approx = np.concatenate([np.sum(np.multiply(ws[i][:,np.newaxis], self.kernel_approx.transform(Xp[i])), axis=0)[np.newaxis,:] for i in range(len(Xp))]) - Xfit = (1./(np.sqrt(2*np.pi)*self.bandwidth)) * np.matmul(approx, self.approx_.T) - else: - for i in range(len(Xp)): - for j in range(len(self.diagrams_)): - W = np.matmul(ws[i][:,np.newaxis], self.ws_[j][np.newaxis,:]) - E = (1./(np.sqrt(2*np.pi)*self.bandwidth)) * np.exp(-np.square(pairwise_distances(Xp[i], self.diagrams_[j]))/(2*np.square(self.bandwidth))) - Xfit[i,j] = np.sum(np.multiply(W, E)) - - return Xfit + return pairwise_persistence_diagram_kernels(X, self.diagrams_, kernel="persistence_weighted_gaussian", bandwidth=self.bandwidth, weight=self.weight, kernel_approx=self.kernel_approx, n_jobs=self.n_jobs) + + def __call__(self, diag1, diag2): + """ + Apply PersistenceWeightedGaussianKernel on a single pair of persistence diagrams and outputs the result. + + Parameters: + diag1 (n x 2 numpy array): first input persistence diagram. + diag2 (n x 2 numpy array): second input persistence diagram. + + Returns: + float: persistence weighted Gaussian kernel value. + """ + return _persistence_weighted_gaussian_kernel(diag1, diag2, weight=self.weight, kernel_approx=self.kernel_approx, bandwidth=self.bandwidth) class PersistenceScaleSpaceKernel(BaseEstimator, TransformerMixin): """ This is a class for computing the persistence scale space kernel matrix from a list of persistence diagrams. The persistence scale space kernel is computed by adding the symmetric to the diagonal of each point in each persistence diagram, with negative weight, and then convolving the points with a Gaussian kernel. See https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Reininghaus_A_Stable_Multi-Scale_2015_CVPR_paper.pdf for more details. """ - def __init__(self, bandwidth=1., kernel_approx=None): + def __init__(self, bandwidth=1., kernel_approx=None, n_jobs=None): """ Constructor for the PersistenceScaleSpaceKernel class. Parameters: bandwidth (double): bandwidth of the Gaussian kernel with which persistence diagrams will be convolved (default 1.) kernel_approx (class): kernel approximation class used to speed up computation (default None). Common kernel approximations classes can be found in the scikit-learn library (such as RBFSampler for instance). + n_jobs (int): number of jobs to use for the computation. See :func:`pairwise_persistence_diagram_kernels` for details. """ - self.pwg_ = PersistenceWeightedGaussianKernel(bandwidth=bandwidth, weight=lambda x: 1 if x[1] >= x[0] else -1, kernel_approx=kernel_approx) + self.bandwidth, self.kernel_approx = bandwidth, kernel_approx + self.n_jobs = n_jobs def fit(self, X, y=None): """ @@ -142,11 +218,7 @@ class PersistenceScaleSpaceKernel(BaseEstimator, TransformerMixin): X (list of n x 2 numpy arrays): input persistence diagrams. y (n x 1 array): persistence diagram labels (unused). """ - self.diagrams_ = list(X) - for i in range(len(self.diagrams_)): - op_D = self.diagrams_[i][:,[1,0]] - self.diagrams_[i] = np.concatenate([self.diagrams_[i], op_D], axis=0) - self.pwg_.fit(X) + self.diagrams_ = X return self def transform(self, X): @@ -159,17 +231,26 @@ class PersistenceScaleSpaceKernel(BaseEstimator, TransformerMixin): Returns: numpy array of shape (number of diagrams in **diagrams**) x (number of diagrams in X): matrix of pairwise persistence scale space kernel values. """ - Xp = list(X) - for i in range(len(Xp)): - op_X = Xp[i][:,[1,0]] - Xp[i] = np.concatenate([Xp[i], op_X], axis=0) - return self.pwg_.transform(Xp) + return pairwise_persistence_diagram_kernels(X, self.diagrams_, kernel="persistence_scale_space", bandwidth=self.bandwidth, kernel_approx=self.kernel_approx, n_jobs=self.n_jobs) + + def __call__(self, diag1, diag2): + """ + Apply PersistenceScaleSpaceKernel on a single pair of persistence diagrams and outputs the result. + + Parameters: + diag1 (n x 2 numpy array): first input persistence diagram. + diag2 (n x 2 numpy array): second input persistence diagram. + + Returns: + float: persistence scale space kernel value. + """ + return _persistence_scale_space_kernel(diag1, diag2, bandwidth=self.bandwidth, kernel_approx=self.kernel_approx) class PersistenceFisherKernel(BaseEstimator, TransformerMixin): """ This is a class for computing the persistence Fisher kernel matrix from a list of persistence diagrams. The persistence Fisher kernel is computed by exponentiating the corresponding persistence Fisher distance with a Gaussian kernel. See papers.nips.cc/paper/8205-persistence-fisher-kernel-a-riemannian-manifold-kernel-for-persistence-diagrams for more details. """ - def __init__(self, bandwidth_fisher=1., bandwidth=1., kernel_approx=None): + def __init__(self, bandwidth_fisher=1., bandwidth=1., kernel_approx=None, n_jobs=None): """ Constructor for the PersistenceFisherKernel class. @@ -177,9 +258,11 @@ class PersistenceFisherKernel(BaseEstimator, TransformerMixin): bandwidth (double): bandwidth of the Gaussian kernel applied to the persistence Fisher distance (default 1.). bandwidth_fisher (double): bandwidth of the Gaussian kernel used to turn persistence diagrams into probability distributions by PersistenceFisherDistance class (default 1.). kernel_approx (class): kernel approximation class used to speed up computation (default None). Common kernel approximations classes can be found in the scikit-learn library (such as RBFSampler for instance). + n_jobs (int): number of jobs to use for the computation. See :func:`pairwise_persistence_diagram_kernels` for details. """ self.bandwidth = bandwidth - self.pf_ = PersistenceFisherDistance(bandwidth=bandwidth_fisher, kernel_approx=kernel_approx) + self.bandwidth_fisher, self.kernel_approx = bandwidth_fisher, kernel_approx + self.n_jobs = n_jobs def fit(self, X, y=None): """ @@ -189,7 +272,7 @@ class PersistenceFisherKernel(BaseEstimator, TransformerMixin): X (list of n x 2 numpy arrays): input persistence diagrams. y (n x 1 array): persistence diagram labels (unused). """ - self.pf_.fit(X, y) + self.diagrams_ = X return self def transform(self, X): @@ -202,5 +285,18 @@ class PersistenceFisherKernel(BaseEstimator, TransformerMixin): Returns: numpy array of shape (number of diagrams in **diagrams**) x (number of diagrams in X): matrix of pairwise persistence Fisher kernel values. """ - return np.exp(-self.pf_.transform(X)/self.bandwidth) + return pairwise_persistence_diagram_kernels(X, self.diagrams_, kernel="persistence_fisher", bandwidth=self.bandwidth, bandwidth_fisher=self.bandwidth_fisher, kernel_approx=self.kernel_approx, n_jobs=self.n_jobs) + + def __call__(self, diag1, diag2): + """ + Apply PersistenceFisherKernel on a single pair of persistence diagrams and outputs the result. + + Parameters: + diag1 (n x 2 numpy array): first input persistence diagram. + diag2 (n x 2 numpy array): second input persistence diagram. + + Returns: + float: persistence Fisher kernel value. + """ + return np.exp(-_persistence_fisher_distance(diag1, diag2, bandwidth=self.bandwidth, kernel_approx=self.kernel_approx)) / self.bandwidth_fisher |