Results output

discretesampling/base/algorithms/mcmc.py

@@ -2,6 +2,7 @@
 import copy
 from tqdm.auto import tqdm
 from discretesampling.base.random import RNG
+from discretesampling.base.output import MCMCOutput
 
 
 class DiscreteVariableMCMC():
@@ -12,15 +13,20 @@ def __init__(self, variableType, target, initialProposal):
         self.initialProposal = initialProposal
         self.target = target
 
-    def sample(self, N, seed=0, verbose=True):
+    def sample(self, N, N_warmup=None, seed=0, include_warmup=True, verbose=True):
         rng = RNG(seed)
+
+        if N_warmup is None:
+            N_warmup = int(N/2)
+
         initialSample = self.initialProposal.sample(rng)
         current = initialSample
 
         samples = []
 
         display_progress_bar = verbose
         progress_bar = tqdm(total=N, desc="MCMC sampling", disable=not display_progress_bar)
+        acceptance_probabilities = []
 
         for i in range(N):
             forward_proposal = self.proposalType(current, rng)
@@ -49,7 +55,13 @@ def sample(self, N, seed=0, verbose=True):
                 pass
 
             samples.append(copy.copy(current))
+            acceptance_probabilities.append(acceptance_probability)
             progress_bar.update(1)
 
+        if not include_warmup:
+            samples = samples[(N_warmup-1):N]
+            acceptance_probabilities = acceptance_probabilities[(N_warmup-1):N]
+
+        results = MCMCOutput(samples, acceptance_probabilities, include_warmup, N, N_warmup)
         progress_bar.close()
-        return samples
+        return results

discretesampling/base/algorithms/smc.py

@@ -7,6 +7,7 @@
 from discretesampling.base.algorithms.smc_components.normalisation import normalise
 from discretesampling.base.algorithms.smc_components.effective_sample_size import ess
 from discretesampling.base.algorithms.smc_components.resampling import systematic_resampling
+from discretesampling.base.output import SMCOutput
 
 
 class DiscreteVariableSMC():
@@ -29,7 +30,7 @@ def __init__(self, variableType, target, initialProposal,
         self.initialProposal = initialProposal
         self.target = target
 
-    def sample(self, Tsmc, N, seed=0, verbose=True):
+    def sample(self, Tsmc, N, seed=0, gather_results=True, verbose=True):
         loc_n = int(N/self.exec.P)
         rank = self.exec.rank
         mvrs_rng = RNG(seed)
@@ -79,5 +80,10 @@ def sample(self, Tsmc, N, seed=0, verbose=True):
             current_particles = new_particles
             progress_bar.update(1)
 
+        if gather_results:
+            current_particles = self.exec.gather_all(current_particles)
+            logWeights = self.exec.gather(logWeights, [N])
+
+        results = SMCOutput(current_particles, logWeights)
         progress_bar.close()
-        return current_particles
+        return results

discretesampling/base/algorithms/smc_components/effective_sample_size.py

@@ -1,7 +1,8 @@
 import numpy as np
+from discretesampling.base.executor import Executor
 
 
-def ess(logw, exec):
+def ess(logw, exec: Executor = Executor()):
     """
     Description
     -----------

discretesampling/base/executor/executor.py

@@ -31,3 +31,6 @@ def redistribute(self, particles, ncopies):
             [[particles[i]]*ncopies[i] for i in range(len(particles))]
         ))
         return particles
+
+    def gather_all(self, particles):
+        return particles

discretesampling/base/executor/executor_MPI.py

@@ -2,6 +2,7 @@
 import numpy as np
 from scipy.special import logsumexp
 from discretesampling.base.executor import Executor
+from discretesampling.base.util import gather_all
 from discretesampling.base.executor.MPI.distributed_fixed_size_redistribution.prefix_sum import (
     inclusive_prefix_sum
 )
@@ -61,3 +62,7 @@ def cumsum(self, x):
 
     def redistribute(self, particles, ncopies):
         return variable_size_redistribution(particles, ncopies, self)
+
+    def gather_all(self, particles):
+        particles = gather_all(particles, exec=self)
+        return particles

discretesampling/base/output/__init__.py

@@ -0,0 +1,3 @@
+from .base import BaseOutput  # noqa
+from .smc_output import SMCOutput  # noqa
+from .mcmc_output import MCMCOutput  # noqa

discretesampling/base/output/base.py

@@ -0,0 +1,13 @@
+from abc import ABC
+import numpy as np
+
+
+class BaseOutput(ABC):
+    def __init__(self, samples):
+        self.samples = samples
+        self.log_weights = np.repeat([-len(samples)], len(samples))
+
+    def __eq__(self, other):
+        samples_eq = np.array_equal(self.samples, other.samples)
+        weights_eq = np.array_equal(self.log_weights, other.log_weights)
+        return samples_eq and weights_eq

discretesampling/base/output/mcmc_output.py

@@ -0,0 +1,24 @@
+import numpy as np
+from discretesampling.base.types import DiscreteVariable
+from discretesampling.base.output import BaseOutput
+
+
+class MCMCOutput(BaseOutput):
+    def __init__(
+        self,
+        samples: list[DiscreteVariable],
+        acceptance_probabilities=None,
+        include_warmup=None,
+        N=None, N_warmup=None
+    ):
+        super().__init__(samples)
+        self.acceptance_probabilities = acceptance_probabilities
+        self.include_warmup = include_warmup
+        if N is None:
+            N = len(samples)
+        self.N = N
+        self.N_warmup = N_warmup
+        acceptance_rate = None
+        if acceptance_probabilities is not None:
+            acceptance_rate = np.mean(acceptance_probabilities)
+        self.acceptance_rate = acceptance_rate

discretesampling/base/output/smc_output.py

@@ -0,0 +1,17 @@
+import numpy as np
+from typing import Union
+from discretesampling.base.types import DiscreteVariable
+from discretesampling.base.output import BaseOutput
+from discretesampling.base.algorithms.smc_components import ess
+
+
+class SMCOutput(BaseOutput):
+    def __init__(
+        self,
+        samples: list[DiscreteVariable],
+        log_weights: Union[np.ndarray, None] = None
+    ):
+        super().__init__(samples)
+        if log_weights is not None:
+            self.log_weights = log_weights
+        self.ess = ess(self.log_weights)

examples/additive_structure_example.py

@@ -37,9 +37,9 @@ def func(x):
 
 
 # Comparison between the three methods based on sampling the true structure
-mcmc = [x.discrete_set for x in asSamplesMCMC]
-lkern = [x.discrete_set for x in asSamplesLk]
-smc = [x.discrete_set for x in asSamplesSMC]
+mcmc = [x.discrete_set for x in asSamplesMCMC.samples]
+lkern = [x.discrete_set for x in asSamplesLk.samples]
+smc = [x.discrete_set for x in asSamplesSMC.samples]
 
 mcmc.count(true_str)
 lkern.count(true_str)

examples/decision_tree_example.py

@@ -29,7 +29,7 @@
 try:
     treeSamples = dtMCMC.sample(500)
 
-    mcmcLabels = dt.stats(treeSamples, X_test).predict(X_test, use_majority=True)
+    mcmcLabels = dt.stats(treeSamples.samples, X_test).predict(X_test, use_majority=True)
     mcmcAccuracy = [dt.accuracy(y_test, mcmcLabels)]
     print("MCMC mean accuracy: ", (mcmcAccuracy))
 except ZeroDivisionError:
@@ -40,7 +40,7 @@
 try:
     treeSMCSamples = dtSMC.sample(10, 1000)
 
-    smcLabels = dt.stats(treeSMCSamples, X_test).predict(X_test, use_majority=True)
+    smcLabels = dt.stats(treeSMCSamples.samples, X_test).predict(X_test, use_majority=True)
     smcAccuracy = [dt.accuracy(y_test, smcLabels)]
     print("SMC mean accuracy: ", np.mean(smcAccuracy))
 

examples/mcmc_example.py

@@ -24,7 +24,7 @@
 
 try:
     treeSamples = dtMCMC.sample(N=1000)
-    mcmcLabels = dt.stats(treeSamples[500:999], X_test).predict(X_test)
+    mcmcLabels = dt.stats(treeSamples.samples[500:999], X_test).predict(X_test)
     mcmc_acc = dt.accuracy(y_test, mcmcLabels)
     print(numpy.mean(mcmc_acc))
 except ZeroDivisionError:

examples/mpi/decision_tree_example.py

@@ -5,7 +5,6 @@
 from discretesampling.base.algorithms import DiscreteVariableSMC
 from discretesampling.domain import decision_tree as dt
 from discretesampling.base.executor.executor_MPI import Executor_MPI
-from discretesampling.base.util import gather_all
 
 data = datasets.load_wine()
 
@@ -35,10 +34,7 @@
     MPI.COMM_WORLD.Barrier()
     end = MPI.Wtime()
 
-    if MPI.COMM_WORLD.Get_size() > 1:
-        treeSMCSamples = gather_all(treeSMCSamples, exec)
-
-    smcLabels = dt.stats(treeSMCSamples, X_test).predict(X_test)
+    smcLabels = dt.stats(treeSMCSamples.samples, X_test).predict(X_test)
     smcAccuracy = dt.accuracy(y_test, smcLabels)
 
     if MPI.COMM_WORLD.Get_rank() == 0:

examples/mpi/regression_tree_example.py

@@ -5,7 +5,6 @@
 from discretesampling.base.algorithms import DiscreteVariableSMC
 from discretesampling.domain import decision_tree as dt
 from discretesampling.base.executor.executor_MPI import Executor_MPI
-from discretesampling.base.util import gather_all
 
 data = datasets.load_diabetes()
 
@@ -32,10 +31,8 @@
     treeSMCSamples = dtSMC.sample(T, N, seed)
     MPI.COMM_WORLD.Barrier()
     end = MPI.Wtime()
-    if MPI.COMM_WORLD.Get_size() > 1:
-        treeSMCSamples = gather_all(treeSMCSamples, exec)
 
-    smcLabels = dt.RegressionStats(treeSMCSamples, X_test).predict(X_test)
+    smcLabels = dt.RegressionStats(treeSMCSamples.samples, X_test).predict(X_test)
     smcAccuracy = dt.accuracy_mse(y_test, smcLabels)
 
     if MPI.COMM_WORLD.Get_rank() == 0:

examples/regression_decision_tree_example.py

@@ -19,7 +19,7 @@
 try:
     treeSamples = dtMCMC.sample(100)
 
-    mcmcLabels = dt.RegressionStats(treeSamples, X_test).predict(X_test, use_majority=True)
+    mcmcLabels = dt.RegressionStats(treeSamples.samples, X_test).predict(X_test, use_majority=True)
     mcmcAccuracy = [dt.accuracy_mse(y_test, mcmcLabels)]
 
     print("MCMC mean MSE: ", (mcmcAccuracy))
@@ -31,7 +31,7 @@
 try:
     treeSMCSamples = dtSMC.sample(10, 10)
 
-    smcLabels = dt.RegressionStats(treeSMCSamples, X_test).predict(X_test, use_majority=True)
+    smcLabels = dt.RegressionStats(treeSMCSamples.samples, X_test).predict(X_test, use_majority=True)
     smcAccuracy = [dt.accuracy_mse(y_test, smcLabels)]
 
     print("SMC mean MSE: ", (smcAccuracy))

examples/smc_example.py

@@ -25,7 +25,7 @@
 
 try:
     treeSamples = dtSMC.sample(10, 10)
-    smcLabels = dt.stats(treeSamples, X_test).predict(X_test, use_majority=True)
+    smcLabels = dt.stats(treeSamples.samples, X_test).predict(X_test, use_majority=True)
     smc_acc = dt.accuracy(y_test, smcLabels)
     print(numpy.mean(smc_acc))
 except ZeroDivisionError:

tests/test_mcmc.py

@@ -1,12 +1,18 @@
 import pytest
 from discretesampling.base.algorithms import DiscreteVariableMCMC
 from discretesampling.domain import spectrum
+from discretesampling.base.output import MCMCOutput
 
 
 @pytest.mark.parametrize(
     "seed,T,expected",
-    [(0, 10, [spectrum.SpectrumDimension(i) for i in [31, 30, 30, 30, 30, 29, 28, 27, 28, 27]]),
-     (1, 10, [spectrum.SpectrumDimension(i) for i in [27, 28, 27, 26, 25, 26, 27, 26, 25, 24]])]
+    [
+        (0, 10, MCMCOutput(
+            [spectrum.SpectrumDimension(i) for i in [31, 30, 30, 30, 30, 29, 28, 27, 28, 27]]
+        )),
+        (1, 10, MCMCOutput(
+         [spectrum.SpectrumDimension(i) for i in [27, 28, 27, 26, 25, 26, 27, 26, 25, 24]]))
+    ]
 
 )
 def test_mcmc(seed, T, expected):

tests/test_smc.py

@@ -1,16 +1,31 @@
 import pytest
+import numpy as np
 from discretesampling.base.algorithms import DiscreteVariableSMC
-from discretesampling.base.util import gather_all
 from discretesampling.base.executor.executor_MPI import Executor_MPI
 from discretesampling.domain import spectrum
+from discretesampling.base.output import SMCOutput
 
 # Test reproducibility
 
 
 @pytest.mark.parametrize(
     "seed,T,N,expected",
-    [(0, 3, 16, [spectrum.SpectrumDimension(i) for i in [15, 13, 15, 6, 6, 14, 8, 8, 8, 6, 6, 6, 14, 14, 12, 10]]),
-     (1, 3, 16, [spectrum.SpectrumDimension(i) for i in [13, 15, 15, 6, 2, 6, 18, 8, 6, 6, 6, 8, 14, 12, 10, 12]])]
+    [
+        (0, 3, 16, SMCOutput(
+            [spectrum.SpectrumDimension(i) for i in [15, 13, 15, 6, 6, 14, 8, 8, 8, 6, 6, 6, 14, 14, 12, 10]],
+            np.array([-3.3423949417970995, -2.6927167512983843, -3.3423949417970995, -2.5485352600349263, -2.5485352600349263,
+                      -1.7791312350561306, -2.807685770214025, -2.807685770214025, -2.807685770214025, -3.2906867661536747,
+                      -3.2906867661536747, -3.2906867661536747, -3.286288420776657, -3.286288420776657, -2.753770144750778,
+                      -2.4895429738738315])
+        )),
+        (1, 3, 16, SMCOutput(
+            [spectrum.SpectrumDimension(i) for i in [13, 15, 15, 6, 2, 6, 18, 8, 6, 6, 6, 8, 14, 12, 10, 12]],
+            np.array([-2.577088421206954, -3.2267666117056693, -3.2267666117056693, -2.432906929943496, -5.389379226012296,
+                      -2.432906929943496, -3.37139647864896, -2.692057440122595, -3.1750584360622445, -3.1750584360622445,
+                      -3.1750584360622445, -2.692057440122595, -3.1706600906852267, -2.6381418146593476, -2.3739146437824012,
+                      -2.6381418146593476])
+        )
+        )]
 )
 def test_smc(seed, T, N, expected):
     target = spectrum.SpectrumDimensionTarget(10, 3.4)  # NB with mean 10 and variance 3.4^2
@@ -25,15 +40,28 @@ def test_smc(seed, T, N, expected):
 @pytest.mark.mpi
 @pytest.mark.parametrize(
     "seed,T,N,expected",
-    [(0, 3, 16, [spectrum.SpectrumDimension(i) for i in [15, 13, 15, 6, 6, 14, 8, 8, 8, 6, 6, 6, 14, 14, 12, 10]]),
-     (1, 3, 16, [spectrum.SpectrumDimension(i) for i in [13, 15, 15, 6, 2, 6, 18, 8, 6, 6, 6, 8, 14, 12, 10, 12]])]
+    [
+        (0, 3, 16, SMCOutput(
+            [spectrum.SpectrumDimension(i) for i in [15, 13, 15, 6, 6, 14, 8, 8, 8, 6, 6, 6, 14, 14, 12, 10]],
+            np.array([-3.3423949417970995, -2.6927167512983843, -3.3423949417970995, -2.5485352600349263, -2.5485352600349263,
+                      -1.7791312350561306, -2.807685770214025, -2.807685770214025, -2.807685770214025, -3.2906867661536747,
+                      -3.2906867661536747, -3.2906867661536747, -3.286288420776657, -3.286288420776657, -2.753770144750778,
+                      -2.4895429738738315])
+        )),
+        (1, 3, 16, SMCOutput(
+            [spectrum.SpectrumDimension(i) for i in [13, 15, 15, 6, 2, 6, 18, 8, 6, 6, 6, 8, 14, 12, 10, 12]],
+            np.array([-2.577088421206954, -3.2267666117056693, -3.2267666117056693, -2.432906929943496, -5.389379226012296,
+                      -2.432906929943496, -3.37139647864896, -2.692057440122595, -3.1750584360622445, -3.1750584360622445,
+                      -3.1750584360622445, -2.692057440122595, -3.1706600906852267, -2.6381418146593476, -2.3739146437824012,
+                      -2.6381418146593476])
+        ))
+    ]
 )
 def test_smc_MPI(seed, T, N, expected):
     target = spectrum.SpectrumDimensionTarget(10, 3.4)  # NB with mean 10 and variance 3.4^2
     initialProposal = spectrum.SpectrumDimensionInitialProposal(50)  # Uniform sampling from 0-50
 
     exec = Executor_MPI()
     specSMC = DiscreteVariableSMC(spectrum.SpectrumDimension, target, initialProposal, exec=exec)
-    local_samples = specSMC.sample(T, N, seed=seed)
-    samples = gather_all(local_samples, exec)
+    samples = specSMC.sample(T, N, seed=seed)
     assert samples == expected