CUQIpy 1.0.0.post0.dev337__tar.gz → 1.0.0.post0.dev360__tar.gz

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/CUQIpy.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: CUQIpy
-Version: 1.0.0.post0.dev337
+Version: 1.0.0.post0.dev360
 Summary: Computational Uncertainty Quantification for Inverse problems in Python
 Maintainer-email: "Nicolai A. B. Riis" <nabr@dtu.dk>, "Jakob S. Jørgensen" <jakj@dtu.dk>, "Amal M. Alghamdi" <amaal@dtu.dk>, Chao Zhang <chaz@dtu.dk>
 License:                                  Apache License

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/CUQIpy.egg-info/SOURCES.txt

@@ -38,6 +38,7 @@ cuqi/distribution/_joint_distribution.py
 cuqi/distribution/_laplace.py
 cuqi/distribution/_lmrf.py
 cuqi/distribution/_lognormal.py
+cuqi/distribution/_modifiedhalfnormal.py
 cuqi/distribution/_normal.py
 cuqi/distribution/_posterior.py
 cuqi/distribution/_uniform.py
@@ -55,6 +56,7 @@ cuqi/experimental/mcmc/_mh.py
 cuqi/experimental/mcmc/_pcn.py
 cuqi/experimental/mcmc/_rto.py
 cuqi/experimental/mcmc/_sampler.py
+cuqi/experimental/mcmc/_utilities.py
 cuqi/geometry/__init__.py
 cuqi/geometry/_geometry.py
 cuqi/implicitprior/__init__.py

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: CUQIpy
-Version: 1.0.0.post0.dev337
+Version: 1.0.0.post0.dev360
 Summary: Computational Uncertainty Quantification for Inverse problems in Python
 Maintainer-email: "Nicolai A. B. Riis" <nabr@dtu.dk>, "Jakob S. Jørgensen" <jakj@dtu.dk>, "Amal M. Alghamdi" <amaal@dtu.dk>, Chao Zhang <chaz@dtu.dk>
 License:                                  Apache License

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/cuqi/_version.py

@@ -8,11 +8,11 @@ import json
 
 version_json = '''
 {
- "date": "2024-06-11T10:55:36+0200",
+ "date": "2024-07-01T14:33:09+0200",
  "dirty": false,
  "error": null,
- "full-revisionid": "b3c53f2ac21c807186dc35053421615c1cb3f205",
- "version": "1.0.0.post0.dev337"
+ "full-revisionid": "4da44104c9163182cd1eeb31fb4678a2893111fe",
+ "version": "1.0.0.post0.dev360"
 }
 '''  # END VERSION_JSON
 

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/cuqi/distribution/__init__.py

@@ -3,6 +3,7 @@ from ._beta import Beta
 from ._cauchy import Cauchy
 from ._cmrf import CMRF
 from ._gamma import Gamma
+from ._modifiedhalfnormal import ModifiedHalfNormal
 from ._gaussian import Gaussian, JointGaussianSqrtPrec
 from ._gmrf import GMRF
 from ._inverse_gamma import InverseGamma

cuqipy-1.0.0.post0.dev360/cuqi/distribution/_modifiedhalfnormal.py

@@ -0,0 +1,184 @@
+import numpy as np
+import scipy.stats as sps
+import scipy.special as special
+from cuqi.distribution import Distribution
+from cuqi.utilities import force_ndarray
+
+class ModifiedHalfNormal(Distribution):
+    """
+    Represents a modified half-normal (MHN) distribution, a three-parameter family of distributions generalizing the Gamma distribution.
+    The distribution is continuous with pdf 
+    f(x; alpha, beta, gamma) propto x^(alpha-1) * exp(-beta * x^2 + gamma * x)
+
+    The MHN generalizes the half-normal distribution, because
+    f(x; 1, beta, 0) propto exp(-beta * x^2)
+
+    The MHN generalizes the gamma distribution because
+    f(x; alpha, 0, -gamma) propto x^(alpha-1) * exp(- gamma * x)
+
+    Reference:
+    [1] Sun, et al. "The Modified-Half-Normal distribution: Properties and an efficient sampling scheme." Communications in Statistics-Theory and Methods
+    
+    Parameters
+    ----------
+    alpha : float
+        The polynomial exponent parameter of the MHN distribution. Must be positive.
+
+    beta : float
+        The quadratic exponential parameter of the MHN distribution. Must be positive.
+
+    gamma : float
+        The linear exponential parameter of the MHN distribution.
+
+    """
+    def __init__(self, alpha=None, beta=None, gamma=None, is_symmetric=False, **kwargs):
+        # Init from abstract distribution class
+        super().__init__(is_symmetric=is_symmetric, **kwargs) 
+
+        self._alpha = alpha
+        self._beta = beta
+        self._gamma = gamma
+
+    @property
+    def alpha(self):
+        """ The polynomial exponent parameter of the MHN distribution. Must be positive. """
+        return self._alpha
+    
+    @alpha.setter
+    def shape(self, value):
+        self._shape = force_ndarray(value, flatten=True)
+
+    @property
+    def beta(self):
+        """ The quadratic exponential parameter of the MHN distribution. Must be positive. """
+        return self._alpha
+    
+    @beta.setter
+    def beta(self, value):
+        self._beta = force_ndarray(value, flatten=True)
+
+    @property
+    def gamma(self):
+        """ The linear exponential parameter of the MHN distribution. """
+        return self._alpha
+    
+    @gamma.setter
+    def gamma(self, value):
+        self._gamma = force_ndarray(value, flatten=True)
+
+    def logpdf(self, x): # Unnormalized
+        return np.sum((self.alpha - 1)*np.log(x) - self.beta * x * x + self.gamma * x)
+
+    def _gradient_scalar(self, val):
+        if val <= 0.0:
+            return np.nan
+        return (self.alpha - 1)/val - 2*self.beta*val + self.gamma
+
+    def _gradient(self, val, *args, **kwargs):
+        if hasattr(self.alpha, '__iter__'):
+            return np.array([self._gradient_scalar(v) for v in val])
+        else:
+            return np.array([self.dim*[self._gradient_scalar(v)] for v in val])
+    
+    def _MHN_sample_gamma_proposal(self, alpha, beta, gamma, rng, delta=None):
+        """
+            Sample from a modified half-normal distribution using a Gamma distribution proposal.
+        """
+        if delta is None:
+            delta = beta + (gamma*gamma - gamma*np.sqrt(gamma*gamma + 8*beta*alpha))/(4*alpha)
+            
+        while True:
+            T = rng.gamma(alpha/2, 1.0/delta)
+            X = np.sqrt(T)
+            U = rng.uniform()
+            if X > 0 and np.log(U) < -(beta-delta)*T + gamma*X - gamma*gamma/(4*(beta-delta)):
+                return X
+            
+    def _MHN_sample_normal_proposal(self, alpha, beta, gamma, mu, rng):
+        """
+            Sample from a modified half-normal distribution using a Normal/Gaussian distribution proposal.
+        """
+        if mu is None:
+            mu = (gamma + np.sqrt(gamma*gamma + 8*beta*(alpha - 1)))/(4*beta)    
+        
+        while True:
+            X = rng.normal(mu, np.sqrt(0.5/beta))
+            U = rng.uniform()
+            if X > 0 and np.log(U) < (alpha-1)*np.log(X) - np.log(mu) + (2*beta*mu-gamma)*(mu-X):
+                return X
+
+    def _MHN_sample_positive_gamma_1(self, alpha, beta, gamma, rng):
+        """
+            Sample from a modified half-normal distribution, assuming alpha is greater than one and gamma is positive.
+        """
+        if gamma <= 0.0:
+            raise ValueError("gamma needs to be positive")
+            
+        if alpha <= 1.0:
+            raise ValueError("alpha needs to be greater than 1.0")
+
+        # Decide whether to use Normal or sqrt(Gamma) proposals for acceptance-rejectance scheme
+        mu = (gamma + np.sqrt(gamma*gamma + 8*beta*(alpha - 1)))/(4*beta)
+        K1  = 2*np.sqrt(np.pi)
+        K1 *= np.power((np.sqrt(beta)*(alpha-1))/(2*beta*mu-gamma), alpha - 1)
+        K1 *= np.exp(-(alpha-1)+beta*mu*mu)
+        
+        delta = beta + (gamma*gamma - gamma*np.sqrt(gamma*gamma + 8*beta*alpha))/(4*alpha)
+        K2  = np.power(beta/delta, 0.5*alpha)
+        K2 *= special.gamma(alpha/2.0)
+        K2 *= np.exp(gamma*gamma/(4*(beta-delta)))
+
+        if K2 > K1: # Use normal proposal
+            return self._MHN_sample_normal_proposal(alpha, beta, gamma, mu, rng)
+        else: # Use sqrt(gamma) proposal
+            return self._MHN_sample_gamma_proposal(alpha, beta, gamma, rng, delta)
+
+    def _MHN_sample_negative_gamma(self, alpha, beta, gamma, rng, m=None):
+        """
+            Sample from a modified half-normal distribution, assuming gamma is negative.
+            The argument 'm' is the matching point, see Algorithm 3 from [1] for details.
+        """
+        if gamma > 0.0:
+            raise ValueError("gamma needs to be negative")
+            
+        if m is None:
+            if alpha <= 1.0:
+                m = 1.0
+            else:
+                m = "mode"
+        
+        # The acceptance rate of this choice is at least 0.5*sqrt(2) approx 70.7 percent, according to Theorem 4 from [1].
+        if isinstance(m, str) and m.lower() == "mode":
+            m = (gamma + np.sqrt(gamma*gamma + 8*beta*alpha))/(4*beta)
+
+        while True:
+            val1 = (beta*m-gamma)/(2*beta*m-gamma)
+            val2 = m*(beta*m-gamma)
+            T = rng.gamma(alpha*val1, 1.0/val2)
+            X = m*np.power(T,val1)
+            U = rng.uniform()
+            if np.log(U) < val2*T-beta*X*X+gamma*X:
+                return X
+            
+    def _MHN_sample(self, alpha, beta, gamma, m=None, rng=None):
+        """
+        Sample from a modified half-normal distribution using an algorithm from [1].
+        """
+        if rng == None:
+            rng = np.random
+
+        if gamma <= 0.0:
+            return self._MHN_sample_negative_gamma(alpha, beta, gamma, m=m, rng=rng)
+        
+        if alpha > 1:
+            return self._MHN_sample_positive_gamma_1(alpha, beta, gamma, rng=rng)
+        
+        return self._MHN_sample_gamma_proposal(alpha, beta, gamma, rng=rng)
+
+    def _sample(self, N, rng=None):
+        if hasattr(self.alpha, '__getitem__'):
+            return np.array([self._MHN_sample(self.alpha[i], self.beta[i], self.gamma[i], rng=rng) for i in range(N)])
+        else:
+            return np.array([self._MHN_sample(self.alpha, self.beta, self.gamma, rng=rng) for i in range(N)])
+
+            

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/cuqi/experimental/mcmc/__init__.py

@@ -12,3 +12,4 @@ from ._gibbs import HybridGibbsNew
 from ._conjugate import ConjugateNew
 from ._conjugate_approx import ConjugateApproxNew
 from ._direct import DirectNew
+from ._utilities import find_valid_samplers

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/cuqi/experimental/mcmc/_cwmh.py

@@ -100,6 +100,9 @@ class CWMHNew(ProposalBasedSamplerNew):
             raise ValueError(
                 "Target should be an instance of "+\
                 f"{cuqi.density.Density.__class__.__name__}")
+        # Fail when there is no log density, which is currently assumed to be the case in case NaN is returned.
+        if np.isnan(self.target.logd(self._default_initial_point)):
+            raise ValueError("Target does not have valid logd")
         
     def validate_proposal(self):
         if not isinstance(self.proposal, cuqi.distribution.Distribution):

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/cuqi/experimental/mcmc/_mh.py

@@ -32,7 +32,9 @@ class MHNew(ProposalBasedSamplerNew):
         self._scale_temp = self.scale
 
     def validate_target(self):
-        pass # All targets are valid
+        # Fail only when there is no log density, which is currently assumed to be the case in case NaN is returned.
+        if np.isnan(self.target.logd(self._default_initial_point)):
+            raise ValueError("Target does not have valid logd")
 
     def validate_proposal(self):
         if not isinstance(self.proposal, cuqi.distribution.Distribution):

cuqipy-1.0.0.post0.dev360/cuqi/experimental/mcmc/_utilities.py

@@ -0,0 +1,17 @@
+import cuqi
+import inspect
+
+def find_valid_samplers(target):
+    """ Finds all samplers in the cuqi.experimental.mcmc module that accept the provided target. """
+
+    all_samplers = [(name, cls) for name, cls in inspect.getmembers(cuqi.experimental.mcmc, inspect.isclass) if issubclass(cls, cuqi.experimental.mcmc.SamplerNew)]
+    valid_samplers = []
+
+    for name, sampler in all_samplers:
+        try:
+            sampler(target)
+            valid_samplers += [name]
+        except:
+            pass
+
+    return valid_samplers

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/tests/test_distribution.py

@@ -758,3 +758,25 @@ def test_Gaussian_from_linear_operator_sqrtprec():
     y_from_dense = cuqi.distribution.Gaussian(mean = np.zeros(N), sqrtprec = sqrtprec.todense())
 
     assert np.allclose(y_from_dense.logpdf(np.ones(N)), y_from_sparse.logpdf(np.ones(N)))
+
+@pytest.mark.parametrize("alpha, beta, gamma, expected",[
+                        (1.0, 2.0, 3.0, [0.77974597, 0.77361298, 0.5422682, 0.81054637, 1.35205349]),
+                        (128.0, 3.0, -4.0, [1.02250649, 1.05735833, 0.98794758, 1.04533337, 1.00500678])
+                        ])
+def test_MHN_sample_regression(alpha, beta, gamma, expected):
+    rng = np.random.RandomState(0)
+    dist = cuqi.distribution.ModifiedHalfNormal(alpha, beta, gamma)
+    samples = dist.sample(5,rng=rng).samples
+    assert np.allclose( samples, np.array(expected))
+
+@pytest.mark.parametrize("alpha, beta, gamma, expected_logpdf, expected_gradient",[
+                        (1.0, 2.0, 3.0, -40.0, [[-1.], [-3.], [-5.], [-7.], [-9.]]),
+                        (64.0, 3.0, -4.0, -2258.388020204731, [[-1.], [-160.5], [-299.], [-432.25], [-563.4 ]])
+                        ])
+def test_MHN_regression(alpha, beta, gamma, expected_logpdf, expected_gradient):
+    rng = np.random.RandomState(0)
+    dist = cuqi.distribution.ModifiedHalfNormal(alpha, beta, gamma)
+    logpdf = dist.logpdf(np.array([1.0, 2.0, 3.0, 4.0, 5.0]))
+    gradient = dist._gradient(np.array([1.0, 2.0, 3.0, 4.0, 5.0]))
+    assert np.allclose( logpdf, np.array(expected_logpdf))
+    assert np.allclose( gradient, np.array(expected_gradient))

{cuqipy-1.0.0.post0.dev337 → cuqipy-1.0.0.post0.dev360}/tests/test_distributions_shape.py

@@ -17,7 +17,7 @@ ignore_list = [
 
 # Define cases to skip (these are TODO)
 skip_logd = [
-    cuqi.distribution.Gamma # Missing force_ndarray
+    cuqi.distribution.Gamma, # Missing force_ndarray
 ]
 skip_sample = [
     cuqi.distribution.Gamma, # Missing force_ndarray
@@ -67,7 +67,8 @@ def test_multivariate_scalar_vars_logd(dist):
     val = np.random.randn(5)
     assert np.allclose(
         dist_from_vec.logd(val),
-        dist_from_dim.logd(val)
+        dist_from_dim.logd(val),
+        equal_nan=True
     )