def calculate(self, inputs, verbose=False):
n_evs = inputs[:len(self.signals)]
binned_signals = inputs[len(self.signals):-1]
x_kj = inputs[-1]
if x_kj is not self.last_x_kj:
if x_kj.shape == tuple(
[len(edges) - 1 for edges in self.bin_edges]):
print(
'NOTE: pre-binned data detected; assuming binning is correct'
)
self.counts = cp.asarray(x_kj)
else:
self.counts, _ = cp.histogramdd(cp.asarray(x_kj),
bins=self.bin_edges)
self.last_x_kj = x_kj
if verbose:
print('Evaluate:', ', '.join(['%0.3f' % s for s in n_evs]))
expected = cp.sum(cp.asarray(
[n * bin_ints for n, bin_ints in zip(n_evs, binned_signals)]),
axis=0)
expected = expected.reshape(self.counts.shape)
mask = expected != 0
res = cp.sum(n_evs) - cp.sum(
self.counts[mask] * cp.log(expected[mask]))
if verbose:
print('NLL:', res)
if np.isnan(res):
if self.nan_behavior == 'unlikely':
return 1e200
else:
raise Exception('NaN likelihood!')
return res if np == cp else res.get()
def bin_mc(self, t_ij, w_i):
'''
'''
counts, _ = cp.histogramdd(cp.asarray(t_ij),
bins=self.bin_edges,
weights=cp.asarray(w_i))
return (counts.flatten() / self.bin_vol / cp.sum(counts)).reshape(
counts.shape)
def load_mc(self, t_ij):
for j, (l, h) in enumerate(
zip(self.observables.lows, self.observables.highs)):
in_bounds = np.logical_and(t_ij[:, j] > l, t_ij[:, j] < h)
t_ij = t_ij[in_bounds]
self.t_ij = cp.asarray(t_ij)
self.w_i = cp.ones(t_ij.shape[0])
if self.bootstrap_binning is not None:
counts, _ = cp.histogramdd(cp.asarray(self.t_ij),
bins=self.bin_edges,
weights=cp.asarray(self.w_i))
self.counts = (cp.asarray(counts).flatten() / self.bin_vol /
cp.sum(cp.asarray(counts))).reshape(counts.shape)
self.sigma_j = cp.std(self.t_ij, axis=0)
self.h_ij = self._adapt_bandwidth()
for j, (l, h, refl) in enumerate(
zip(self.observables.lows, self.observables.highs,
self.reflect_axes)):
if not refl:
continue
if type(refl) == tuple:
low, high = refl
mask = self.t_ij[:, j] < low
t_ij_reflected_low = cp.copy(self.t_ij[mask, :])
h_ij_reflected_low = self.h_ij[mask, :]
w_i_reflected_low = self.w_i[mask, :]
t_ij_reflected_low[:, j] = 2 * l - t_ij_reflected_low[:, j]
mask = self.t_ij[:, j] > high
t_ij_reflected_high = cp.copy(self.t_ij[mask, :])
h_ij_reflected_high = self.h_ij[mask, :]
w_i_reflected_high = self.w_i[mask, :]
t_ij_reflected_high[:, j] = 2 * h - t_ij_reflected_high[:, j]
else:
t_ij_reflected_low = cp.copy(self.t_ij)
h_ij_reflected_low = self.h_ij
w_i_reflected_low = self.w_i
t_ij_reflected_low[:, j] = 2 * l - self.t_ij[:, j]
t_ij_reflected_high = cp.copy(self.t_ij)
h_ij_reflected_high = self.h_ij
w_i_reflected_high = self.w_i
t_ij_reflected_high[:, j] = 2 * h - self.t_ij[:, j]
self.t_ij = cp.concatenate(
[self.t_ij, t_ij_reflected_low, t_ij_reflected_high])
self.h_ij = cp.concatenate(
[self.h_ij, h_ij_reflected_low, h_ij_reflected_high])
self.w_i = cp.concatenate(
[self.w_i, w_i_reflected_low, w_i_reflected_high])
self.t_ij = cp.ascontiguousarray(self.t_ij)
self.h_ij = cp.ascontiguousarray(self.h_ij)
self.w_i = cp.ascontiguousarray(self.w_i)
def test_histogramdd_disallow_arraylike_bins(self):
x = testing.shaped_random((16, 2), cupy, scale=100)
bins = [[0, 10, 20, 50, 90]] * 2 # too many dimensions
with pytest.raises(ValueError):
y, bin_edges = cupy.histogramdd(x, bins=bins)
def time_fine_binning(self):
np.histogramdd(self.d, (10000, 10000), ((0, 100), (0, 100)))
def time_small_coverage(self):
np.histogramdd(self.d, (200, 200), ((50, 51), (50, 51)))
def time_full_coverage(self):
np.histogramdd(self.d, (200, 200), ((0, 100), (0, 100)))