Initial sampling commit

src/original/ETP_SRI/Sampling.py

@@ -0,0 +1,130 @@
+import numpy as np
+import emcee
+import tqdm
+import corner
+from scipy.stats import rice, norm, uniform
+from matplotlib import pyplot as plt
+
+from utilities.data_simulation.GenerateData import GenerateData
+
+class MCMC:
+    """
+    Performs sampling of exponential data
+    """
+    def __init__(self,
+                 data,
+                 b,
+                 data_scale=1e-5,
+                 parameter_scale=(1e-7, 1e-11, 1e-9),
+                 bounds=((0, 1), (0, 1), (0, 1)),
+                 priors=None,
+                 gaussian_noise=False,
+                 nwalkers=16,
+                 nsteps=10000,
+                 burn_in=2000,
+                 progress=True):
+        """
+        Parameters
+        ----------
+
+        """
+        self.data = np.atleast_2d(np.asarray(data))
+        self.b = np.atleast_2d(np.asarray(b)).T
+        self.data_scale = np.asarray(data_scale)
+        self.parameter_scale = np.asarray(parameter_scale)
+        self.bounds = np.asarray(bounds)
+        self.priors = np.asarray(priors)
+        self.nwalkers = nwalkers
+        self.nsteps = nsteps
+        self.burn_in = burn_in
+        self.progress = progress
+        if priors is None:
+            self.prior = self.zero_prior
+        else:
+            self.prior = self.loglike_gauss_prior
+        if gaussian_noise:
+            self.likelihood = self.biexp_loglike_gauss
+        else:
+            self.likelihood = self.biexp_loglike_rice
+        self.ndim = 3
+        self.chain = None
+        self.means = None
+        self.stds = None
+
+    def accepted_dimensions(self):
+        return (1, 1)
+
+    def bounds_prior(self, params):
+        return np.sum(uniform.logpdf(params, loc=self.bounds[:, 0], scale=self.bounds[:, 1] - self.bounds[:, 0]), 1)
+
+    def loglike_gauss_prior(self, params):
+        return np.sum(norm.logpdf(params, loc=self.priors[:, 0], scale=self.priors[:, 1]), 1)
+
+    def zero_prior(self, params):
+        return 0
+
+    def signal(self, f, D, D_star):
+        return (f * np.exp(-self.b * D_star) + (1 - f) * np.exp(-self.b * D)).T
+
+    def biexp_loglike_gauss(self, f, D, D_star):
+        expected = self.signal(f, D, D_star)
+        # check this!
+        # print(f'likelihood {norm.logpdf(self.data, loc=expected/self.data_scale, scale=self.data_scale)}')
+        return np.sum(norm.logpdf(self.data, loc=expected, scale=self.data_scale), 1)
+
+    # def biexp_loglike_gauss_full(self, f, D, D_star):
+    #     expected = self.signal(f, D, D_star)
+    #     print(f'expected {expected}')
+    #     print(f'data {self.data}')
+    #     return norm.logpdf(self.data, loc=expected, scale=self.data_scale)
+
+    def biexp_loglike_rice(self, f, D, D_star):
+        expected = self.signal(f, D, D_star)
+        # print(f'expected {expected}')
+        return np.sum(rice.logpdf(self.data, b=expected/self.data_scale, scale=self.data_scale), 1)
+
+    def posterior(self, params):
+        params = np.atleast_2d(params)
+        total = self.bounds_prior(params)
+        # print(f'bounds params {total}')
+        neginf = np.isneginf(total)
+        # print(f'neginf {neginf}')
+        f = params[~neginf, 0]
+        D = params[~neginf, 1]
+        D_star = params[~neginf, 2]
+        prior = self.prior(params[~neginf, :])
+        # print(f'prior {prior}')
+        likelihood = self.likelihood(f, D, D_star)
+        # print(f'likelihood {likelihood}')
+        total[~neginf] += prior + likelihood
+        return total
+
+    def sample(self, initial_pos):
+        # f = initial_pos[0]
+        # D = initial_pos[1]
+        # D_star = initial_pos[2]
+        # print(f'initial pos likelihood {self.biexp_loglike_gauss_full(f, D, D_star)}')
+        print(f'initial pos likelihood {self.posterior(initial_pos)}')
+        sampler = emcee.EnsembleSampler(self.nwalkers, 3, self.posterior, vectorize=True)
+        pos = initial_pos + self.parameter_scale * np.random.randn(self.nwalkers, self.ndim)
+        # print(f'pos {pos}')
+        # print(f'nsteps {self.nsteps}')
+        sampler.run_mcmc(pos, self.nsteps, progress=True)
+        self.chain = sampler.get_chain(discard=self.burn_in, flat=True)
+        self.means = np.mean(self.chain, 0)
+        self.stds = np.std(self.chain, 0)
+        print(f'final pos likelihood {self.posterior(self.means)}')
+        # print(f'final pos likelihood {self.biexp_loglike_gauss_full(self.means[0], self.means[1], self.means[2])}')
+        # print(f'chain {self.chain}')
+        return self.means, self.stds
+
+    def plot(self, truths=None, labels=('f', 'D', 'D*'), overplot=None):
+        if truths is None:
+            truths = self.means
+        # print(f'chain size {self.chain.shape}')
+        fig = corner.corner(self.chain, labels=labels, truths=truths)
+        fig.suptitle("Sampling of the IVIM data", fontsize=16)
+        if overplot is not None:
+            corner.overplot_lines(fig, overplot, color='r')
+        plt.show()
+

tests/IVIMmodels/unit_tests/simple_test_run_of_algorithm.py

@@ -1,33 +1,81 @@
 import numpy as np
+import matplotlib.pyplot as plt
 import os
 from pathlib import Path
 #from src.standardized.ETP_SRI_LinearFitting import ETP_SRI_LinearFitting
-from src.standardized.IAR_LU_biexp import IAR_LU_biexp
+# from src.standardized.IAR_LU_biexp import IAR_LU_biexp
 #from src.standardized.IAR_LU_segmented_2step import IAR_LU_segmented_2step
 from src.standardized.PvH_KB_NKI_IVIMfit import PvH_KB_NKI_IVIMfit
 #from src.standardized.PV_MUMC_biexp import PV_MUMC_biexp
+from src.original.ETP_SRI.Sampling import MCMC
 
 ## Simple test code... 
 # Used to just do a test run of an algorithm during development
 def dev_test_run(model, **kwargs):
-    bvalues = np.array([0, 50, 100, 150, 200, 500, 800])
+    bvalues = np.array([0, 20, 50, 75, 100, 150, 200, 300, 400, 500, 800, 1000, 1500])
 
-    def ivim_model(b, S0=1, f=0.1, Dstar=0.01, D=0.001):
+    def ivim_model(b, f=0.1, Dstar=0.01, D=0.001, S0=1):
+        # print(f'S0 {S0}')
+        # print(f'Dstar {f*np.exp(-b*Dstar)}')
+        # print(f'D {(1-f)*np.exp(-b*D)}')
+        # print(f'sum {f*np.exp(-b*Dstar) + (1-f)*np.exp(-b*D)}')
+        # print(f'S0 {(f*np.exp(-b*Dstar) + (1-f)*np.exp(-b*D))}')
         return S0*(f*np.exp(-b*Dstar) + (1-f)*np.exp(-b*D))
 
-    signals = ivim_model(bvalues)
+    # TODO: add S0 fitting!
+    true_f = 0.4
+    true_Dstar = 0.01
+    true_D = 0.001
+    truth = [true_f, true_D, true_Dstar]
+    signals_noiseless = ivim_model(bvalues, true_f, true_Dstar, true_D)
+    print(f'noiselss {signals_noiseless}')
+    signals = signals_noiseless + np.abs(1e-1 * (np.random.randn(len(bvalues)) + 1j * np.random.randn(len(bvalues))) / np.sqrt(2))
 
     #model = ETP_SRI_LinearFitting(thresholds=[200])
     if kwargs:
         results = model.osipi_fit(signals, bvalues, **kwargs)
     else:
         results = model.osipi_fit(signals, bvalues)
-    print(results)
+    # print(results) # f, D*, D
+    results_reordered = np.asarray([results[0], results[2], results[1]])
+    print(truth)
+    print(results_reordered)
     #test = model.osipi_simple_bias_and_RMSE_test(SNR=20, bvalues=bvalues, f=0.1, Dstar=0.03, D=0.001, noise_realizations=10)
+    signal_results = ivim_model(bvalues, results[0], results[1], results[2])
+
+
+
+    mcmc = MCMC(signals, bvalues, gaussian_noise=False, data_scale=1e-2) #, priors=((0.07, 1e-1), (0.0135, 1e-1), (0.001, 1e-1)))
+    means, stds = mcmc.sample(truth)
+    print(f'means {means} stds {stds}')
+    print(f'expected {results_reordered}')
+    print(f'truth {truth}')
+
+    signal_means = ivim_model(bvalues, means[0], means[2], means[1])
+    plt.plot(bvalues, signals_noiseless, 'g', label='Noiseless Signal')
+    plt.plot(bvalues, signals, '.g', label='Noisy Signal')
+    plt.plot(bvalues, signal_results, 'r', label='Results Signal')
+    plt.plot(bvalues, signal_means, 'b', label='Means Signal')
+    plt.legend()
+    # plt.show()
+
+    mcmc.plot(overplot=truth)
+
 
 #model1 = ETP_SRI_LinearFitting(thresholds=[200])
 #model2 = IAR_LU_biexp()
 model2 = PvH_KB_NKI_IVIMfit()
 
 #dev_test_run(model1, linear_fit_option=True)
 dev_test_run(model2)
+
+
+def run_sampling():
+    bvalues = np.array([0, 50, 100, 150, 200, 500, 800])
+
+    def ivim_model(b, S0=1, f=0.1, Dstar=0.01, D=0.001):
+        return S0*(f*np.exp(-b*Dstar) + (1-f)*np.exp(-b*D))
+
+    signals = ivim_model(bvalues)
+
+