def runOnePoint(pdf, data):
if isinstance(pyhf.tensorlib, pyhf.tensor.tensorflow_backend):
# Reset the TensorFlow graph and session for each run
tf.reset_default_graph()
pyhf.tensorlib.session = tf.Session()
return pyhf.runOnePoint(1.0, data, pdf, pdf.config.suggested_init(),
pdf.config.suggested_bounds())
def validate_runOnePoint(pdf, data, mu_test, expected_result, tolerance=1e-5):
init_pars = pdf.config.suggested_init()
par_bounds = pdf.config.suggested_bounds()
CLs_obs, CLs_exp = pyhf.runOnePoint(mu_test, data, pdf, init_pars,
par_bounds)[-2:]
assert (CLs_obs - expected_result['obs']) / \
expected_result['obs'] < tolerance
for result, expected_result in zip(CLs_exp, expected_result['exp']):
assert (result - expected_result) / \
expected_result < tolerance
def test_runOnePoint_q_mu(n_bins,
tolerance={
'numpy': 1e-02,
'tensors': 5e-03
}):
"""
Check that the different backends all compute a test statistic
that is within a specific tolerance of each other.
Args:
n_bins: `list` of number of bins given by pytest parameterization
tolerance: `dict` of the maximum differences the test statistics
can differ relative to each other
Returns:
None
"""
source = generate_source_static(n_bins)
pdf = hepdata_like(source['bindata']['sig'], source['bindata']['bkg'],
source['bindata']['bkgerr'])
data = source['bindata']['data'] + pdf.config.auxdata
backends = [
pyhf.tensor.numpy_backend(poisson_from_normal=True),
pyhf.tensor.tensorflow_backend(session=tf.Session()),
pyhf.tensor.pytorch_backend(),
# mxnet_backend()
]
test_statistic = []
for backend in backends:
pyhf.set_backend(backend)
if isinstance(pyhf.tensorlib, pyhf.tensor.tensorflow_backend):
tf.reset_default_graph()
pyhf.tensorlib.session = tf.Session()
q_mu = pyhf.runOnePoint(1.0, data, pdf, pdf.config.suggested_init(),
pdf.config.suggested_bounds())[0]
test_statistic.append(q_mu)
# compare to NumPy/SciPy
test_statistic = np.array(test_statistic)
numpy_ratio = np.divide(test_statistic, test_statistic[0])
numpy_ratio_delta_unity = np.absolute(np.subtract(numpy_ratio, 1))
# compare tensor libraries to each other
tensors_ratio = np.divide(test_statistic[1], test_statistic[2])
tensors_ratio_delta_unity = np.absolute(np.subtract(tensors_ratio, 1))
try:
assert (numpy_ratio_delta_unity < tolerance['numpy']).all()
except AssertionError:
print('Ratio to NumPy+SciPy exceeded tolerance of {}: {}'.format(
tolerance['numpy'], numpy_ratio_delta_unity.tolist()))
assert False
try:
assert (tensors_ratio_delta_unity < tolerance['tensors']).all()
except AssertionError:
print('Ratio between tensor backends exceeded tolerance of {}: {}'.
format(tolerance['tensors'], tensors_ratio_delta_unity.tolist()))
assert False