CUQIpy 1.0.0.post0.dev202__tar.gz → 1.0.0.post0.dev229__tar.gz

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/CUQIpy.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: CUQIpy
-Version: 1.0.0.post0.dev202
+Version: 1.0.0.post0.dev229
 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.dev202 → cuqipy-1.0.0.post0.dev229}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: CUQIpy
-Version: 1.0.0.post0.dev202
+Version: 1.0.0.post0.dev229
 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.dev202 → cuqipy-1.0.0.post0.dev229}/cuqi/_version.py

@@ -8,11 +8,11 @@ import json
 
 version_json = '''
 {
- "date": "2024-04-30T14:38:06+0200",
+ "date": "2024-05-21T12:48:50+0200",
  "dirty": false,
  "error": null,
- "full-revisionid": "e5596bf72d78672549c26c4921e473c12fcf2553",
- "version": "1.0.0.post0.dev202"
+ "full-revisionid": "eb9519734f2558f66772895e41f2cd0c3cd61767",
+ "version": "1.0.0.post0.dev229"
 }
 '''  # END VERSION_JSON
 

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/cuqi/distribution/_gaussian.py

@@ -9,7 +9,7 @@ import scipy.linalg as splinalg
 
 from cuqi import config
 from cuqi.geometry import _get_identity_geometries
-from cuqi.utilities import force_ndarray, sparse_cholesky
+from cuqi.utilities import force_ndarray, sparse_cholesky, check_if_conditional_from_attr
 from cuqi.distribution import Distribution
 
 # We potentially allow the use of sksparse.cholmod for sparse Cholesky
@@ -191,7 +191,7 @@ class Gaussian(Distribution):
         value = force_ndarray(value)
         self._sqrtcov = value
         self._cov = None # Reset covariance (in case it was computed before)      
-        if (value is not None) and (not callable(value)):  
+        if (value is not None) and (not callable(value)):
             if self.dim > config.MIN_DIM_SPARSE:
                 sparse_flag = True # do sparse computations
             else:
@@ -214,7 +214,7 @@ class Gaussian(Distribution):
         value = force_ndarray(value)
         self._sqrtprec = value
         self._cov = None # Reset covariance (in case it was computed before)
-        if (value is not None) and (not callable(value)):  
+        if not check_if_conditional_from_attr(value):
             if self.dim > config.MIN_DIM_SPARSE:
                 sparse_flag = True # do sparse computations
             else:
@@ -631,7 +631,7 @@ def get_sqrtprec_from_sqrtprec(dim, sqrtprec, sparse_flag):
     dim : int
         Dimension of the sqrtprec matrix.
     
-    sqrtprec : 1-d or 2-d ndarray or sparse matrix
+    sqrtprec : 1-d or 2-d ndarray or sparse matrix or scipy.sparse.linalg.LinearOperator
         Square root of precision matrix. If 1-dimensional, then assumed to be a diagonal matrix.
 
     sparse_flag: bool
@@ -666,6 +666,14 @@ def get_sqrtprec_from_sqrtprec(dim, sqrtprec, sparse_flag):
         logdet = np.sum(-np.log(sqrtprec.data**2))
         rank = dim
 
+    # sqrtprec is LinearOperator
+    elif isinstance(sqrtprec, spa.linalg.LinearOperator):
+        if hasattr(sqrtprec, 'logdet'):
+            logdet = sqrtprec.logdet
+        else:
+            logdet = None
+        rank = dim
+
     # sqrtprec diagonal
     elif np.count_nonzero(sqrtprec-np.diag(sqrtprec.diagonal())) == 0:
         stdinv = sqrtprec.diagonal()

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/cuqi/experimental/mcmc/__init__.py

@@ -3,7 +3,7 @@
 from ._sampler import SamplerNew, ProposalBasedSamplerNew
 from ._langevin_algorithm import ULANew, MALANew
 from ._mh import MHNew
-from ._pcn import pCNNew
+from ._pcn import PCNNew
 from ._rto import LinearRTONew, RegularizedLinearRTONew
 from ._cwmh import CWMHNew
 from ._laplace_approximation import UGLANew

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/cuqi/experimental/mcmc/_pcn.py

@@ -3,9 +3,9 @@ import cuqi
 from cuqi.experimental.mcmc import SamplerNew
 from cuqi.array import CUQIarray
 
-class pCNNew(SamplerNew):  # Refactor to Proposal-based sampler?
+class PCNNew(SamplerNew):  # Refactor to Proposal-based sampler?
 
-    _STATE_KEYS = SamplerNew._STATE_KEYS.union({'scale', 'current_likelihood_logd'})
+    _STATE_KEYS = SamplerNew._STATE_KEYS.union({'scale', 'current_likelihood_logd', 'lambd'})
 
     def __init__(self, target, scale=1.0, **kwargs):
 
@@ -17,6 +17,11 @@ class pCNNew(SamplerNew):  # Refactor to Proposal-based sampler?
 
         self._acc = [1] # TODO. Check if we need this
 
+        # parameters used in the Robbins-Monro recursion for tuning the scale parameter
+        # see details and reference in the tune method
+        self.lambd = self.scale
+        self.star_acc = 0.44 #TODO: 0.234 # target acceptance rate
+
     def validate_target(self):
         try:
             if isinstance(self.prior, (cuqi.distribution.Gaussian, cuqi.distribution.Normal)):
@@ -29,7 +34,7 @@ class pCNNew(SamplerNew):  # Refactor to Proposal-based sampler?
     def step(self):
         # propose state
         xi = self.prior.sample(1).flatten()   # sample from the prior
-        x_star = np.sqrt(1-self.scale**2)*self.current_point + self.scale*xi   # pCN proposal
+        x_star = np.sqrt(1-self.scale**2)*self.current_point + self.scale*xi   # PCN proposal
 
         # evaluate target
         loglike_eval_star =  self._loglikelihood(x_star) 
@@ -74,10 +79,6 @@ class pCNNew(SamplerNew):  # Refactor to Proposal-based sampler?
             self._loglikelihood = lambda x : self.likelihood.logd(x)
         else:
             raise ValueError(f"To initialize an object of type {self.__class__}, 'target' need to be of type 'cuqi.distribution.Posterior'.")
-        
-        #TODO:
-        #if not isinstance(self.prior,(cuqi.distribution.Gaussian, cuqi.distribution.Normal)):
-        #    raise ValueError("The prior distribution of the target need to be Gaussian")
 
     @property
     def dim(self): # TODO. Check if we need this. Implemented in base class
@@ -88,4 +89,21 @@ class pCNNew(SamplerNew):  # Refactor to Proposal-based sampler?
         return self._dim
 
     def tune(self, skip_len, update_count):
-        pass
+        """
+        Tune the scale parameter of the PCN sampler.
+        The tuning is based on algorithm 4 in Andrieu, Christophe, and Johannes Thoms. 
+        "A tutorial on adaptive MCMC." Statistics and computing 18 (2008): 343-373.
+        Note: the tuning algorithm here is the same as the one used in MH sampler.
+        """
+
+        # average acceptance rate in the past skip_len iterations
+        hat_acc = np.mean(self._acc[-skip_len:])
+
+        # new scaling parameter zeta to be used in the Robbins-Monro recursion
+        zeta = 1/np.sqrt(update_count+1)
+
+        # Robbins-Monro recursion to ensure that the variation of lambd vanishes
+        self.lambd = np.exp(np.log(self.lambd) + zeta*(hat_acc-self.star_acc))
+
+        # update scale parameter
+        self.scale = min(self.lambd, 1)

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/cuqi/utilities/__init__.py

@@ -10,6 +10,7 @@ from ._utilities import (
     ProblemInfo,
     sparse_cholesky,
     approx_derivative,
+    check_if_conditional_from_attr,
 )
 
 from ._get_python_variable_name import _get_python_variable_name

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/cuqi/utilities/_utilities.py

@@ -64,13 +64,27 @@ def get_indirect_variables(dist):
     attributes = []
     for attribute in dist.get_mutable_variables():
         value = getattr(dist, attribute)
-        if callable(value):
+        if check_if_conditional_from_attr(value):
             keys = get_non_default_args(value)
             for key in keys:
                 if key not in attributes: #Ensure we did not already find this key
                     attributes.append(key)
     return attributes 
 
+def check_if_conditional_from_attr(value):
+    """
+    Check if a distribution is conditional from a given attribute.
+    So far, we assume that a distribution is conditional if
+    - the given attribute is a callable function and
+    - the given attribute is not a LinearOperator.
+    """
+    if isinstance(value, spslinalg.LinearOperator):
+        return False
+    elif callable(value):
+        return True
+    else:
+        return False
+
 def get_writeable_attributes(dist):
     """ Get writeable attributes of object instance. """
     attributes = []

{cuqipy-1.0.0.post0.dev202 → cuqipy-1.0.0.post0.dev229}/tests/test_distribution.py

@@ -4,6 +4,7 @@ import numpy as np
 import scipy as sp
 import scipy.stats as scipy_stats
 import scipy.sparse as sps
+import numpy.linalg as nplinalg
 
 from pytest import approx
 import pytest
@@ -738,3 +739,22 @@ def test_Gaussian_from_sparse_sqrtprec():
 
     assert y_from_dense.logpdf(np.ones(N)) == y_from_sparse.logpdf(np.ones(N))
 
+def test_Gaussian_from_linear_operator_sqrtprec():
+    """ Test Gaussian distribution from LinearOperator sqrtprec is equal to dense sqrtprec """
+    N = 10; M = 5
+
+    sqrtprec = sp.sparse.spdiags(np.random.randn(N), 0, N, N)
+    prec = sqrtprec.todense()@sqrtprec.todense().T
+
+    def matvec(x):
+        return sqrtprec @ x
+    def rmatvec(x):
+        return sqrtprec.T @ x
+    
+    sqrtprec_operator = sp.sparse.linalg.LinearOperator((N, N), matvec=matvec, rmatvec=rmatvec)
+    sqrtprec_operator.logdet = -np.log(nplinalg.det(prec))
+
+    y_from_sparse = cuqi.distribution.Gaussian(mean = np.zeros(N), sqrtprec = sqrtprec_operator)
+    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)))