def test_prod_gauss_nd():
test_values = np.random.random(size=(10, 3))
lower = ((-5, -5, -5),)
upper = ((4, 4, 4),)
obs1 = ['a', 'b', 'c']
norm_range_3d = Space(obs=obs1, limits=(lower, upper))
test_values_data = Data.from_tensor(obs=obs1, tensor=test_values)
product_pdf = product_gauss_3d()
probs = product_pdf.pdf(x=test_values_data, norm_range=norm_range_3d)
true_probs = np.prod(
[gauss.pdf(test_values[:, i], norm_range=(-5, 4)) for i, gauss in enumerate(create_gaussians())], axis=0)
probs_np = probs.numpy()
np.testing.assert_allclose(true_probs, probs_np, rtol=1e-2)
def test_prod_gauss_nd():
test_values = np.random.random(size=(10, 3))
test_values_data = Data.from_tensor(obs=obs1 * obs2 * obs3, tensor=test_values)
product_pdf = product_gauss_3d()
assert product_pdf.n_obs == 3
probs = product_pdf.pdf(x=test_values_data)
gaussians = create_gaussians()
for gauss, space in zip(gaussians, [obs1, obs2, obs3]):
gauss.set_norm_range(space.rect_limits)
true_probs = np.prod(
[gauss.pdf(test_values[:, i]) for i, gauss in enumerate(gaussians)], axis=0)
probs_np = probs.numpy()
np.testing.assert_allclose(true_probs, probs_np, rtol=1e-2)
def run_no_nan(func, x):
from zfit.core.data import Data
value_with_nans = func(x=x)
if value_with_nans.dtype in (tf.complex128, tf.complex64):
value_with_nans = tf.math.real(value_with_nans) + tf.math.imag(value_with_nans) # we care only about NaN or not
finite_bools = tf.math.is_finite(tf.cast(value_with_nans, dtype=tf.float64))
finite_indices = tf.where(finite_bools)
new_x = tf.gather_nd(params=x, indices=finite_indices)
new_x = Data.from_tensor(obs=x.obs, tensor=new_x)
vals_no_nan = func(x=new_x)
result = tf.scatter_nd(indices=finite_indices, updates=vals_no_nan,
shape=tf.shape(input=value_with_nans, out_type=finite_indices.dtype))
return result
def test_prod_gauss_nd():
# return
test_values = np.random.random(size=(10, 3))
lower = ((-5, -5, -5), )
upper = ((4, 4, 4), )
obs1 = ['a', 'b', 'c']
norm_range_3d = Space(obs=obs1, limits=(lower, upper))
test_values_data = Data.from_tensor(obs=obs1, tensor=test_values)
probs = prod_gauss_3d.pdf(x=test_values_data, norm_range=norm_range_3d)
true_probs = np.prod([
gauss.pdf(test_values[:, i], norm_range=(-5, 4))
for i, gauss in enumerate(gauss_dists)
])
probs_np = zfit.run(probs)
np.testing.assert_allclose(zfit.run(true_probs), probs_np, rtol=1e-2)
def test_prod_gauss_nd_mixed():
norm_range = (-5, 4)
low, high = norm_range
test_values = np.random.uniform(low=low, high=high, size=(1000, 4))
obs4d = ['a', 'b', 'c', 'd']
test_values_data = Data.from_tensor(obs=obs4d, tensor=test_values)
# prod_gauss_4d.update_integration_options(draws_per_dim=10)
limits_4d = Space(limits=(((-5, ) * 4, ), ((4, ) * 4, )), obs=obs4d)
prod_gauss_4d = product_gauss_4d()
prod_gauss_4d.set_norm_range(limits_4d)
probs = prod_gauss_4d.pdf(x=test_values_data, norm_range=limits_4d)
gausses = create_gaussians()
for gauss in (gausses):
gauss.set_norm_range(norm_range)
gauss1, gauss2, gauss3 = gausses
prod_gauss_3d = product_gauss_3d()
def probs_4d(values):
true_prob = [gauss1.pdf(values[:, 3])]
true_prob += [gauss2.pdf(values[:, 0])]
true_prob += [gauss3.pdf(values[:, 2])]
true_prob += [
prod_gauss_3d.pdf(values[:, (0, 1, 2)],
norm_range=Space(limits=(((-5, ) * 3, ),
((4, ) * 3, )),
obs=['a', 'b', 'c']))
]
return np.prod(true_prob, axis=0)
true_unnormalized_probs = probs_4d(values=test_values)
normalization_probs = limits_4d.area() * probs_4d(
np.random.uniform(low=low, high=high, size=(40**4, 4)))
true_probs = true_unnormalized_probs / tf.reduce_mean(
input_tensor=normalization_probs)
grad = tf.gradients(ys=probs, xs=list(prod_gauss_4d.get_dependents()))
probs_np = zfit.run(probs)
grad_np = zfit.run(grad)
print("Gradients", grad_np)
print(np.average(probs_np))
true_probs_np = zfit.run(true_probs)
assert np.average(probs_np * limits_4d.area()) == pytest.approx(
1., rel=0.33) # low n mc
np.testing.assert_allclose(true_probs_np, probs_np, rtol=2e-2)
def test_prod_gauss_nd_mixed():
norm = (-5, 4)
low, high = norm
test_values = np.random.uniform(low=low, high=high, size=(1000, 4))
obs4d = ["obs1", "obs2", "obs3", "obs4"]
test_values_data = Data.from_tensor(obs=obs4d, tensor=test_values)
limits_4d = Space(limits=(((-5, ) * 4, ), ((4, ) * 4, )), obs=obs4d)
prod_gauss_4d = product_gauss_4d()
prod_gauss_4d.set_norm_range(limits_4d)
probs = prod_gauss_4d.pdf(x=test_values_data, norm=limits_4d)
gausses = create_gaussians()
for gauss in gausses:
gauss.set_norm_range(norm)
gauss1, gauss2, gauss3 = gausses
prod_gauss_3d = product_gauss_3d()
def probs_4d(values):
true_prob = [gauss1.pdf(values[:, 3])]
true_prob += [gauss2.pdf(values[:, 0])]
true_prob += [gauss3.pdf(values[:, 2])]
true_prob += [
prod_gauss_3d.pdf(
values[:, 0:3],
norm=Space(limits=(((-5, ) * 3, ), ((4, ) * 3, )),
obs=["obs1", "obs2", "obs3"]),
)
]
return tf.math.reduce_prod(true_prob, axis=0)
true_unnormalized_probs = probs_4d(values=test_values)
normalization_probs = limits_4d.area() * probs_4d(
tf.random.uniform(minval=low, maxval=high, shape=(40**4, 4)))
true_probs = true_unnormalized_probs / tf.reduce_mean(
input_tensor=normalization_probs)
probs_np = probs.numpy()
true_probs_np = true_probs.numpy()
assert np.average(probs_np * limits_4d.area()) == pytest.approx(
1.0, rel=0.33) # low n mc
np.testing.assert_allclose(true_probs_np, probs_np, rtol=2e-2)