def test_multisource_likelihood():
lf = UnbinnedLogLikelihood(test_conf(n_sources=2))
lf.add_shape_parameter('some_multiplier', (0.5, 1, 2, 4))
lf.add_rate_parameter('s0')
lf.add_rate_parameter('s1')
lf.prepare()
d = lf.base_model.simulate()
lf.set_data(d)
assert lf(s0_rate_multiplier=1, s1_rate_multiplier=1, some_multiplier=1) == lf()
assert lf(s0_rate_multiplier=1, s1_rate_multiplier=1) == lf()
assert lf(s0_rate_multiplier=1) == lf()
assert lf(some_multiplier=1) == lf()
# Equivalence of rate parameters
assert almost_equal(lf(s0_rate_multiplier=2), lf(s1_rate_multiplier=2))
assert almost_equal(lf(s0_rate_multiplier=4), lf(s0_rate_multiplier=2.5, s1_rate_multiplier=2.5))
# Equivalence of rate and shape parameters
assert lf(s0_rate_multiplier=2, s1_rate_multiplier=2) == lf(some_multiplier=2)
# Likelihood goes in right direction
assert lf(some_multiplier=2) < lf()
def test_noninterpolated_pdf():
#
conf = test_conf(n_sources=1)
conf['some_multiplier']=3e-3
lf = UnbinnedLogLikelihood(conf)
lf.add_shape_parameter('mu',(0.,1.))
lf.add_shape_parameter('sigma',(1.,2.))
lf.prepare()
d = np.zeros(1,dtype=[('x',np.float)])
lf.set_data(d)
assert almost_equal(lf(compute_pdf=True,mu=0.5,sigma=1.5),sps.poisson(3).logpmf(1)+sps.norm(0.5,1.5).logpdf(0),1e-5)
assert not almost_equal(lf(compute_pdf=False,mu=0.5,sigma=1.5),sps.poisson(3).logpmf(1)+sps.norm(0.5,1.5).logpdf(0),1e-5)
def test_shape_uncertainty():
lf = UnbinnedLogLikelihood(test_conf(events_per_day=1))
with pytest.raises(InvalidParameterSpecification):
lf.add_shape_uncertainty('strlen_multiplier', 0.5, {1: 'x', 2: 'hi', 3: 'wha'})
lf.add_shape_uncertainty(setting_name='strlen_multiplier',
fractional_uncertainty=0.5,
anchor_zs={1: 'x', 2: 'hi', 3: 'wha'},
base_value=1)
# Make a single event at x=0
lf.prepare()
lf.set_data(np.zeros(1,
dtype=[('x', np.float), ('source', np.int)]))
log_prior = stats.norm(1, 0.5).logpdf
assert lf() == -1 + stats.norm.logpdf(0) + log_prior(1)
assert lf(strlen_multiplier=2) == -2 + np.log(2 * stats.norm.pdf(0)) + log_prior(2)
def test_error_handling():
lf = UnbinnedLogLikelihood(test_conf())
d = lf.base_model.simulate()
lf.add_shape_parameter('some_multiplier', (0.5, 1, 2))
with pytest.raises(NotPreparedException):
lf.set_data(d)
with pytest.raises(NotPreparedException):
lf()
lf.prepare()
with pytest.raises(NotPreparedException):
lf()
lf.set_data(d)
lf()
with pytest.raises(InvalidParameter):
lf(blargh=41)
def test_limit():
"""Test the limit setting code
For now just tests if it runs, does not test whether the results are correct...
"""
lf = LogLikelihood(test_conf(n_sources=2))
lf.add_rate_parameter('s0')
lf.prepare()
lf.set_data(lf.base_model.simulate())
# Test upper limits
one_parameter_interval(lf, target='s0_rate_multiplier', kind='upper', bound=40)
one_parameter_interval(lf, target='s0_rate_multiplier', kind='lower', bound=0.1)
one_parameter_interval(lf, target='s0_rate_multiplier', kind='central', bound=(0.1, 20))
def test_likelihood_value():
"""Just a sanity check to show we get the right likelihood values"""
lf = UnbinnedLogLikelihood(test_conf(events_per_day=1))
lf.add_rate_parameter('s0')
# Make a single event at x=0
lf.set_data(np.zeros(1,
dtype=[('x', np.float), ('source', np.int)]))
assert lf() == -1 + stats.norm.logpdf(0)
assert lf(s0_rate_multiplier=2) == -2 + np.log(2 * stats.norm.pdf(0))
def test_rate_uncertainty():
lf = UnbinnedLogLikelihood(test_conf(events_per_day=1))
lf.add_rate_uncertainty('s0', 0.5)
# Make a single event at x=0
lf.set_data(np.zeros(1,
dtype=[('x', np.float), ('source', np.int)]))
log_prior = stats.norm(1, 0.5).logpdf
assert lf() == -1 + stats.norm.logpdf(0) + log_prior(1)
assert lf(s0_rate_multiplier=2) == -2 + np.log(2 * stats.norm.pdf(0)) + log_prior(2)
def test_limit():
"""Test the limit setting code
For now just tests if it runs, does not test whether the results are correct...
"""
lf = LogLikelihood(test_conf(n_sources=2))
lf.add_rate_parameter('s0')
lf.prepare()
lf.set_data(lf.base_model.simulate())
# Test upper limits
one_parameter_interval(lf,
target='s0_rate_multiplier',
kind='upper',
bound=40)
one_parameter_interval(lf,
target='s0_rate_multiplier',
kind='lower',
bound=0.1)
one_parameter_interval(lf,
target='s0_rate_multiplier',
kind='central',
bound=(0.1, 20))
def test_fit_minuit():
# Single rate parameter
lf = LogLikelihood(test_conf())
lf.add_rate_parameter('s0')
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert isinstance(fit_result, dict)
assert 's0_rate_multiplier' in fit_result
# Don't fit
res, ll = bestfit_minuit(lf, s0_rate_multiplier=1)
assert len(res) == 0
assert ll == lf(s0_rate_multiplier=1)
# Single shape parameter
lf = LogLikelihood(test_conf())
lf.add_shape_parameter('some_multiplier', (0.5, 1, 1.5, 2))
lf.prepare()
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert 'some_multiplier' in fit_result
# Shape and rate parameter
lf = LogLikelihood(test_conf())
lf.add_rate_parameter('s0')
lf.add_shape_parameter('some_multiplier', (0.5, 1, 1.5, 2))
lf.prepare()
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert 'some_multiplier' in fit_result
assert 's0_rate_multiplier' in fit_result
# Non-numeric shape parameter
lf = LogLikelihood(test_conf())
lf.add_shape_parameter('strlen_multiplier', {
1: 'x',
2: 'hi',
3: 'wha'
},
base_value=1)
lf.prepare()
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert 'strlen_multiplier' in fit_result
def test_shape_params():
lf = UnbinnedLogLikelihood(test_conf(n_sources=1))
lf.add_rate_parameter('s0')
with pytest.raises(InvalidParameterSpecification):
lf.add_shape_parameter('strlen_multiplier', {1: 'x', 2: 'hi', 3:'wha'})
lf.add_shape_parameter('strlen_multiplier', {1: 'q', 2: 'hi', 3: 'wha'}, base_value=1)
d = lf.base_model.simulate()
lf.prepare()
lf.set_data(d)
assert len(lf.anchor_models) == 3
# Can't call with the raw setting
with pytest.raises(ValueError):
lf(strlen_multiplier='hi')
# But can call with representative number:
lf(strlen_multiplier=1.5)
# Test for correct use of base_value
assert lf() == lf(strlen_multiplier=1)
# Test for interpolating non-numeric values by their representative settings
assert lf(strlen_multiplier=1.5) < lf()
def test_no_shape_params():
lf = UnbinnedLogLikelihood(test_conf())
d = lf.base_model.simulate()
lf.prepare()
lf.set_data(d)
lf()
# Test a MonteCarloSource, which should trigger a pdf computation
lf = UnbinnedLogLikelihood(test_conf(mc=True))
d = lf.base_model.simulate()
lf.prepare()
lf.set_data(d)
lf()
def test_fit_minuit():
# Single rate parameter
lf = LogLikelihood(test_conf())
lf.add_rate_parameter('s0')
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert isinstance(fit_result, dict)
assert 's0_rate_multiplier' in fit_result
# Don't fit
res, ll = bestfit_minuit(lf, s0_rate_multiplier=1)
assert len(res) == 0
assert ll == lf(s0_rate_multiplier=1)
# Single shape parameter
lf = LogLikelihood(test_conf())
lf.add_shape_parameter('some_multiplier', (0.5, 1, 1.5, 2))
lf.prepare()
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert 'some_multiplier' in fit_result
# Shape and rate parameter
lf = LogLikelihood(test_conf())
lf.add_rate_parameter('s0')
lf.add_shape_parameter('some_multiplier', (0.5, 1, 1.5, 2))
lf.prepare()
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert 'some_multiplier' in fit_result
assert 's0_rate_multiplier' in fit_result
# Non-numeric shape parameter
lf = LogLikelihood(test_conf())
lf.add_shape_parameter('strlen_multiplier', {1: 'x', 2: 'hi', 3:'wha'}, base_value=1)
lf.prepare()
lf.set_data(lf.base_model.simulate())
fit_result, ll = bestfit_minuit(lf)
assert 'strlen_multiplier' in fit_result