def _loss_func(self, model, data, fit_range, constraints, log_offset):
nll = super()._loss_func(
model=model,
data=data,
fit_range=fit_range,
constraints=constraints,
log_offset=log_offset,
)
yields = []
nevents_collected = []
for mod, dat in zip(model, data):
if not mod.is_extended:
raise NotExtendedPDFError(
f"The pdf {mod} is not extended but has to be (for an extended fit)"
)
nevents = dat.n_events if dat.weights is None else z.reduce_sum(
dat.weights)
nevents = tf.cast(nevents, tf.float64)
nevents_collected.append(nevents)
yields.append(mod.get_yield())
yields = znp.stack(yields, axis=0)
nevents_collected = znp.stack(nevents_collected, axis=0)
term_new = tf.nn.log_poisson_loss(nevents_collected, znp.log(yields))
if log_offset is not None:
term_new += log_offset
nll += znp.sum(term_new, axis=0)
return nll
def _rel_counts(self, x, norm):
pdf = self.pdfs[0]
edges = [znp.array(edge) for edge in self.axes.edges]
edges_flat = [znp.reshape(edge, [-1]) for edge in edges]
lowers = [edge[:-1] for edge in edges_flat]
uppers = [edge[1:] for edge in edges_flat]
lowers_meshed = znp.meshgrid(*lowers, indexing="ij")
uppers_meshed = znp.meshgrid(*uppers, indexing="ij")
shape = tf.shape(lowers_meshed[0])
lowers_meshed_flat = [
znp.reshape(lower_mesh, [-1]) for lower_mesh in lowers_meshed
]
uppers_meshed_flat = [
znp.reshape(upper_mesh, [-1]) for upper_mesh in uppers_meshed
]
lower_flat = znp.stack(lowers_meshed_flat, axis=-1)
upper_flat = znp.stack(uppers_meshed_flat, axis=-1)
options = {"type": "bins"}
@z.function
def integrate_one(limits):
l, u = tf.unstack(limits)
limits_space = zfit.Space(obs=self.obs, limits=[l, u])
return pdf.integrate(limits_space, norm=False, options=options)
limits = znp.stack([lower_flat, upper_flat], axis=1)
values = tf.vectorized_map(integrate_one, limits)
values = znp.reshape(values, shape)
if norm:
values /= pdf.normalization(norm)
return values
def _sample(self, n, limits: ZfitSpace):
pdf = self.pdfs[0]
# TODO: use real limits, currently not supported in binned sample
sample = pdf.sample(n=n)
edges = sample.space.binning.edges
ndim = len(edges)
edges = [znp.array(edge) for edge in edges]
edges_flat = [znp.reshape(edge, [-1]) for edge in edges]
lowers = [edge[:-1] for edge in edges_flat]
uppers = [edge[1:] for edge in edges_flat]
lowers_meshed = znp.meshgrid(*lowers, indexing="ij")
uppers_meshed = znp.meshgrid(*uppers, indexing="ij")
lowers_meshed_flat = [
znp.reshape(lower_mesh, [-1]) for lower_mesh in lowers_meshed
]
uppers_meshed_flat = [
znp.reshape(upper_mesh, [-1]) for upper_mesh in uppers_meshed
]
lower_flat = znp.stack(lowers_meshed_flat, axis=-1)
upper_flat = znp.stack(uppers_meshed_flat, axis=-1)
counts_flat = znp.reshape(sample.values(), (-1, ))
counts_flat = tf.cast(counts_flat,
znp.int32) # TODO: what if we have fractions?
lower_flat_repeated = tf.repeat(lower_flat, counts_flat, axis=0)
upper_flat_repeated = tf.repeat(upper_flat, counts_flat, axis=0)
sample_unbinned = tf.random.uniform(
(znp.sum(counts_flat), ndim),
minval=lower_flat_repeated,
maxval=upper_flat_repeated,
dtype=self.dtype,
)
return sample_unbinned
def _counts(self, x, norm):
pdf = self.pdfs[0]
edges = [znp.array(edge) for edge in self.axes.edges]
edges_flat = [znp.reshape(edge, [-1]) for edge in edges]
lowers = [edge[:-1] for edge in edges_flat]
uppers = [edge[1:] for edge in edges_flat]
lowers_meshed = znp.meshgrid(*lowers, indexing="ij")
uppers_meshed = znp.meshgrid(*uppers, indexing="ij")
shape = tf.shape(lowers_meshed[0])
lowers_meshed_flat = [
znp.reshape(lower_mesh, [-1]) for lower_mesh in lowers_meshed
]
uppers_meshed_flat = [
znp.reshape(upper_mesh, [-1]) for upper_mesh in uppers_meshed
]
lower_flat = znp.stack(lowers_meshed_flat, axis=-1)
upper_flat = znp.stack(uppers_meshed_flat, axis=-1)
options = {"type": "bins"}
if pdf.is_extended:
@z.function
def integrate_one(limits):
l, u = tf.unstack(limits)
limits_space = zfit.Space(obs=self.obs, limits=[l, u])
return pdf.ext_integrate(limits_space,
norm=False,
options=options)
missing_yield = False
else:
@z.function
def integrate_one(limits):
l, u = tf.unstack(limits)
limits_space = zfit.Space(obs=self.obs, limits=[l, u])
return pdf.integrate(limits_space, norm=False, options=options)
missing_yield = True
limits = znp.stack([lower_flat, upper_flat], axis=1)
try:
values = tf.vectorized_map(integrate_one, limits)[:, 0]
except ValueError:
values = tf.map_fn(integrate_one, limits)
values = znp.reshape(values, shape)
if missing_yield:
values *= self.get_yield()
if norm:
values /= pdf.normalization(norm)
return values
def to_unbinned(self):
meshed_center = znp.meshgrid(*self.axes.centers, indexing="ij")
flat_centers = [
znp.reshape(center, (-1, )) for center in meshed_center
]
centers = znp.stack(flat_centers, axis=-1)
flat_weights = znp.reshape(self.values(), (-1, )) # TODO: flow?
space = self.space.copy(binning=None)
from zfit import Data
return Data.from_tensor(obs=space,
tensor=centers,
weights=flat_weights)
def spline_interpolator(alpha, alphas, densities):
alphas = alphas[None, :, None]
shape = tf.shape(densities[0])
densities_flat = [znp.reshape(density, [-1]) for density in densities]
densities_flat = znp.stack(densities_flat, axis=0)
alpha_shaped = znp.reshape(alpha, [1, -1, 1])
y_flat = tfa.image.interpolate_spline(
train_points=alphas,
train_values=densities_flat[None, ...],
query_points=alpha_shaped,
order=2,
)
y_flat = y_flat[0, 0]
y = tf.reshape(y_flat, shape)
return y
def to_minimize_func(values):
nonlocal current_loss, nan_counter
do_print = self.verbosity > 8
is_nan = False
gradient = None
value = None
try:
gradient, value = update_params_value_grad(
loss, params, values)
except tf.errors.InvalidArgumentError:
err = "NaNs"
is_nan = True
except:
err = "unknonw error"
raise
finally:
if value is None:
value = f"invalid, {err}"
if gradient is None:
gradient = [f"invalid, {err}"] * len(params)
if do_print:
print_gradient(
params,
run(values),
[float(run(g)) for g in gradient],
loss=run(value),
)
loss_evaluated = run(value)
is_nan = is_nan or np.isnan(loss_evaluated)
if is_nan:
nan_counter += 1
info_values = {}
info_values["loss"] = run(value)
info_values["old_loss"] = current_loss
info_values["nan_counter"] = nan_counter
value = self.strategy.minimize_nan(loss=loss,
params=params,
minimizer=self,
values=info_values)
else:
nan_counter = 0
current_loss = value
gradient = znp.stack(gradient)
return value, gradient
def test_unbinned_data2D():
n = 751
gauss, gauss_binned, obs, obs_binned = create_gauss2d_binned(n, 50)
data = znp.random.uniform([-5, 50], [10, 600], size=(1000, 2))
y_binned = gauss_binned.pdf(data)
y_true = gauss.pdf(data)
max_error = np.max(y_true) / 10
np.testing.assert_allclose(y_true, y_binned, atol=max_error)
centers = obs_binned.binning.centers
X, Y = znp.meshgrid(*centers, indexing="ij")
centers = znp.stack([znp.reshape(t, (-1,)) for t in (X, Y)], axis=-1)
ycenter_binned = gauss_binned.pdf(centers)
ycenter_true = gauss.pdf(centers)
np.testing.assert_allclose(ycenter_binned, ycenter_true, atol=max_error / 10)
# for the extended case
y_binned_ext = gauss_binned.ext_pdf(data)
y_true_ext = gauss.ext_pdf(data)
max_error_ext = np.max(y_true_ext) / 10
np.testing.assert_allclose(y_true_ext, y_binned_ext, atol=max_error_ext)
ycenter_binned_ext = gauss_binned.ext_pdf(centers)
ycenter_true_ext = gauss.ext_pdf(centers)
np.testing.assert_allclose(
ycenter_binned_ext, ycenter_true_ext, atol=max_error_ext / 10
)
x_outside = znp.array([[-7.0, 55], [3.0, 13], [2, 150], [12, 30], [14, 1000]])
y_outside = gauss_binned.pdf(x_outside)
assert y_outside[0] == 0
assert y_outside[1] == 0
assert y_outside[2] > 0
assert y_outside[3] == 0
assert y_outside[4] == 0
y_outside_ext = gauss_binned.ext_pdf(x_outside)
assert y_outside_ext[0] == 0
assert y_outside_ext[1] == 0
assert y_outside_ext[2] > 0
assert y_outside_ext[3] == 0
assert y_outside_ext[4] == 0
def unbinned_to_binindex(data, space, flow=False):
if flow:
warnings.warn(
"Flow currently not fully supported. Values outside the edges are all 0."
)
values = [znp.reshape(data.value(ob), (-1, )) for ob in space.obs]
edges = [znp.reshape(edge, (-1, )) for edge in space.binning.edges]
bins = [
tfp.stats.find_bins(x=val, edges=edge)
for val, edge in zip(values, edges)
]
stacked_bins = znp.stack(bins, axis=-1)
if flow:
stacked_bins += 1
bin_is_nan = tf.math.is_nan(stacked_bins)
zeros = znp.zeros_like(stacked_bins)
binindices = znp.where(bin_is_nan, zeros, stacked_bins)
stacked_bins = znp.asarray(binindices, dtype=znp.int32)
return stacked_bins
def _ext_pdf(self, x, norm):
if not self._automatically_extended:
raise SpecificFunctionNotImplemented
pdf = self.pdfs[0]
density = pdf.ext_pdf(x.space, norm=norm)
density_flat = znp.reshape(density, (-1, ))
centers_list = znp.meshgrid(*pdf.space.binning.centers, indexing="ij")
centers_list_flat = [
znp.reshape(cent, (-1, )) for cent in centers_list
]
centers = znp.stack(centers_list_flat, axis=-1)
# [None, :, None] # TODO: only 1 dim now
probs = tfa.image.interpolate_spline(
train_points=centers[None, ...],
train_values=density_flat[None, :, None],
query_points=x.value()[None, ...],
order=self.order,
)
return probs[0, ..., 0]
def _minimize(self, loss, params):
from .. import run
minimizer_fn = tfp.optimizer.bfgs_minimize
params = tuple(params)
do_print = self.verbosity > 8
current_loss = None
nan_counter = 0
# @z.function
def update_params_value_grad(loss, params, values):
for param, value in zip(params, tf.unstack(values, axis=0)):
param.set_value(value)
value, gradients = loss.value_gradient(params=params)
return gradients, value
def to_minimize_func(values):
nonlocal current_loss, nan_counter
do_print = self.verbosity > 8
is_nan = False
gradient = None
value = None
try:
gradient, value = update_params_value_grad(
loss, params, values)
except tf.errors.InvalidArgumentError:
err = "NaNs"
is_nan = True
except:
err = "unknonw error"
raise
finally:
if value is None:
value = f"invalid, {err}"
if gradient is None:
gradient = [f"invalid, {err}"] * len(params)
if do_print:
print_gradient(
params,
run(values),
[float(run(g)) for g in gradient],
loss=run(value),
)
loss_evaluated = run(value)
is_nan = is_nan or np.isnan(loss_evaluated)
if is_nan:
nan_counter += 1
info_values = {}
info_values["loss"] = run(value)
info_values["old_loss"] = current_loss
info_values["nan_counter"] = nan_counter
value = self.strategy.minimize_nan(loss=loss,
params=params,
minimizer=self,
values=info_values)
else:
nan_counter = 0
current_loss = value
gradient = znp.stack(gradient)
return value, gradient
initial_inv_hessian_est = tf.linalg.tensor_diag(
[p.step_size for p in params])
minimizer_kwargs = dict(
initial_position=znp.stack(params),
x_tol=self.tol,
# f_relative_tolerance=self.tolerance * 1e-5, # TODO: use edm for stopping criteria
initial_inverse_hessian_estimate=initial_inv_hessian_est,
parallel_iterations=1,
max_iterations=self.max_calls,
)
minimizer_kwargs.update(self.options)
result = minimizer_fn(to_minimize_func, **minimizer_kwargs)
# save result
params_result = run(result.position)
assign_values(params, values=params_result)
info = {
"n_eval": run(result.num_objective_evaluations),
"n_iter": run(result.num_iterations),
"grad": run(result.objective_gradient),
"original": result,
}
edm = -999
fmin = run(result.objective_value)
status = -999
converged = run(result.converged)
params = OrderedDict((p, val) for p, val in zip(params, params_result))
result = FitResult(
params=params,
edm=edm,
fmin=fmin,
info=info,
loss=loss,
status=status,
converged=converged,
minimizer=self.copy(),
)
return result
