def test_learn_h_graph(self):
experiment = load_test_experiment()
emission = likelihood.NegBinBetaBinLikelihood()
emission.h = experiment.h
emission.learn_parameters(experiment.x, experiment.l)
N = experiment.l.shape[0]
M = experiment.h.shape[0]
prior = cn_model.CopyNumberPrior(N, M, perfect_cn_prior(experiment.cn))
prior.set_lengths(experiment.l)
model = cn_model.HiddenMarkovModel(N, M, emission, prior)
_, cn_init = model.optimal_state()
graph = genome_graph.GenomeGraph(emission, prior, experiment.adjacencies, experiment.breakpoints)
graph.init_copy_number(cn_init)
h_init = experiment.h + experiment.h * 0.05 * np.random.randn(*experiment.h.shape)
estimator = em.HardAssignmentEstimator(num_em_iter=1)
estimator.learn_param(graph, 'h', h_init)
def test_evaluate_q_derivative(self):
cn, h, l, phi, r, x = generate_simple_data()
emission = likelihood.NegBinBetaBinLikelihood(x, l)
emission.h = h
N = l.shape[0]
M = h.shape[0]
prior = cn_model.CopyNumberPrior(perfect_cn_prior(cn))
prior.set_lengths(l)
model = cn_model.HiddenMarkovModel(N, M, emission, prior, [(0, N)])
log_posterior, cns, resps = model.posterior_marginals()
estimator = em.ExpectationMaximizationEstimator()
params = [emission.h_param, emission.r_param, emission.M_param, emission.z_param, emission.hdel_mu_param, emission.loh_p_param]
value = np.concatenate([p.value for p in params])
idxs = np.array([p.length for p in params]).cumsum() - params[0].length
remixt.tests.utils.assert_grad_correct(
estimator.evaluate_q, estimator.evaluate_q_derivative,
value, model, cns, resps, params, idxs)
def test_learn_h(self):
experiment = load_test_experiment()
h_init = experiment.h * (1. +
0.05 * np.random.randn(*experiment.h.shape))
emission = likelihood.NegBinBetaBinLikelihood(experiment.x,
experiment.l)
emission.h = h_init
print experiment.h, h_init
N = experiment.l.shape[0]
M = experiment.h.shape[0]
prior = cn_model.CopyNumberPrior(experiment.l)
model = cn_model.HiddenMarkovModel(N,
M,
emission,
prior,
experiment.chains,
normal_contamination=False)
estimator = em.ExpectationMaximizationEstimator()
estimator.learn_param(model, emission.h_param)
def test_log_likelihood_cn_betabinnegbin_cornercases(self):
cn, h, l, phi, r, x = self.generate_simple_data()
emission = likelihood.NegBinBetaBinLikelihood()
emission.h = np.array([1e-16, 10., 1e-16])
cn[:, :, :] = np.array([[1., 1.], [0., 1.], [1., 0.]])
ll = emission.log_likelihood(x, l, cn)
ll_partial_h = emission.log_likelihood_partial_param(x, l, cn, 'h')
self.assertTrue(not np.any(np.isnan(ll)))
self.assertTrue(not np.any(np.isnan(ll_partial_h)))
def test_simple_genome_graph(self):
x, cn, h, l, phi = generate_tiny_data()
emission = likelihood.NegBinBetaBinLikelihood(x, l)
emission.h = h
emission.phi = phi
N = l.shape[0]
M = h.shape[0]
prior = cn_model.CopyNumberPrior(l)
adjacencies = [(0, 1), (1, 2)]
breakpoints = [frozenset([(0, 1), (2, 0)])]
for i in xrange(1000):
print i
random.seed(i)
np.random.seed(i)
# Modify copy number from known true
cn_init = cn.copy()
for n in xrange(N):
m = np.random.randint(low=1, high=M)
for allele in xrange(2):
cn_init[n, m, allele] += np.random.randint(low=-1, high=1)
# print np.random.randint(low=-1, high=1, size=cn_init[:, 1:, :].shape)
# cn_init[:, 1:, :] += np.random.randint(low=-1, high=1, size=cn_init[:, 1:, :].shape)
graph = genome_graph.GenomeGraphModel(N, M, emission, prior,
adjacencies, breakpoints)
graph.init_copy_number(cn_init, init_breakpoints=True)
graph.optimize()
if not np.all(graph.breakpoint_copy_number[['cn_1', 'cn_2']].values
== np.array([1, 0])):
print cn
print cn_init
print graph.segment_cn
print graph.breakpoint_copy_number
import IPython
IPython.embed()
raise
self.assertTrue(
np.all(graph.breakpoint_copy_number[['cn_1', 'cn_2']].values ==
np.array([1, 0])))
def test_log_likelihood_cn_betabinnegbin_partial_h(self):
cn, h, l, phi, r, x = self.generate_simple_data()
emission = likelihood.NegBinBetaBinLikelihood()
def evaluate_log_likelihood(h, x, l, cn, phi):
emission.h = h
return emission.log_likelihood(x, l, cn)
def evaluate_log_likelihood_partial_h(h, x, l, cn, phi):
emission.h = h
return emission._log_likelihood_partial_h(x, l, cn)
remixt.tests.utils.assert_grad_correct(
evaluate_log_likelihood, evaluate_log_likelihood_partial_h, h, x,
l, cn, phi)
