def _iterate_run(self, pred_dt_plus, true_dt_plus, true_dt_minus, **kwargs):
# append source_data kwarg to force the use of new training data each time
kwargs['source_data'] = CartesianSpaceTimeData(self.training)
return super(SeppValidationFixedModel, self)._iterate_run(pred_dt_plus,
true_dt_plus=true_dt_plus,
true_dt_minus=true_dt_minus,
**kwargs)
def test_cross_linkage(self):
data_source = CartesianSpaceTimeData(np.random.randn(5000, 3))
data_target = CartesianSpaceTimeData(np.random.randn(1000, 3))
max_t = max_d = 0.5
linkage_fun_sep = utils.linkage_func_separable(max_t, max_d)
i, j = utils.linkages(data_source,
linkage_fun_sep,
data_target=data_target)
self.assertTrue(np.all(i < 5000))
self.assertTrue(np.all(j < 1000))
# manually test restrictions
# all time differences positive
self.assertTrue(np.all(data_target[j, 0] > data_source[i, 0]))
# all time diffs less than max_t
self.assertTrue(np.all(data_target[j, 0] - data_source[i, 0] <= max_t))
# all distances <= max_d
d = np.sqrt((data_target[j, 1] - data_source[i, 1])**2 +
(data_target[j, 2] - data_source[i, 2])**2)
self.assertTrue(np.all(d <= max_d))
def pairwise_distance_histogram(data,
max_t,
max_d,
remove_coincident_pairs=False,
nbin=40,
vmax=None,
fmax=None):
import seaborn as sns
data = CartesianSpaceTimeData(data)
linkage_fun = linkage_func_separable(max_t, max_d)
i, j = linkages(data,
linkage_fun,
remove_coincident_pairs=remove_coincident_pairs)
interpoint = data[j] - data[i]
df = pandas.DataFrame(interpoint[:, 1:], columns=('x (m)', 'y (m)'))
if vmax is not None:
joint_kws = dict(vmax=vmax)
else:
joint_kws = dict()
sns.set_context("paper", font_scale=2.)
with sns.axes_style("white"):
grid = sns.jointplot(x='x (m)',
y='y (m)',
data=df,
kind='hex',
color='k',
stat_func=None,
space=0,
marginal_kws=dict(bins=nbin),
joint_kws=joint_kws,
size=8)
if fmax is not None:
pc = [
t for t in grid.ax_joint.get_children()
if isinstance(t, collections.PolyCollection)
][0]
cdata = pc.get_array()
plt.close(grid.fig)
cdata.sort()
vmax = cdata[int(len(cdata) * fmax)]
grid = sns.jointplot(x='x (m)',
y='y (m)',
data=df,
kind='hex',
color='k',
stat_func=None,
space=0,
marginal_kws=dict(bins=nbin),
joint_kws=dict(vmax=vmax),
size=8)
return grid
def txy_to_cartesian_data_array(t, x, y):
ndim = t.ndim
if x.ndim != ndim or y.ndim != ndim:
raise AttributeError("Ndim does not match")
shape = t.shape
if x.shape != shape or y.shape != shape:
raise AttributeError("Shape does not match")
return CartesianSpaceTimeData(
np.concatenate(
(t[..., np.newaxis], x[..., np.newaxis], y[..., np.newaxis]),
axis=ndim))
def test_self_linkage(self):
data1 = CartesianSpaceTimeData(np.random.randn(5000, 3))
max_t = max_d = 0.5
linkage_fun_sep = utils.linkage_func_separable(max_t, max_d)
i, j = utils.linkages(data1, linkage_fun_sep)
# manually test restrictions
# all time differences positive
self.assertTrue(np.all(data1[j, 0] > data1[i, 0]))
# all time diffs less than max_t
self.assertTrue(np.all(data1[j, 0] - data1[i, 0] <= max_t))
# all distances <= max_d
d = np.sqrt((data1[j, 1] - data1[i, 1])**2 +
(data1[j, 2] - data1[i, 2])**2)
self.assertTrue(np.all(d <= max_d))
def simulation():
decay_const = 1.
intensity = 1000.
x_decay_mean = 0.5
inv_func_t = lambda x: -1.0 / decay_const * np.log(
(intensity - x) / intensity)
sim_times = simulate.nonstationary_poisson(inv_func=inv_func_t, t_max=1)
x = np.random.exponential(scale=x_decay_mean, size=sim_times.size)
y = np.random.random(sim_times.size)
out = x > 1.
while np.any(out):
xn = np.random.exponential(scale=2, size=out.sum())
x[out] = xn
out = x > 1.
return CartesianSpaceTimeData.from_args(sim_times, x, y)
def pairwise_distance_histogram_manual(data,
max_t,
max_d,
ax=None,
remove_coincident_pairs=False,
nbin=40,
vmax=None,
fmax=None,
cmap='binary',
colorbar=True,
mask=True):
# nbin must be EVEN to ensure no weird origin effects
if np.mod(nbin, 2):
nbin += 1
data = CartesianSpaceTimeData(data)
linkage_fun = linkage_func_separable(max_t, max_d)
i, j = linkages(data,
linkage_fun,
remove_coincident_pairs=remove_coincident_pairs)
interpoint = data[j] - data[i]
xe = np.linspace(-max_d, max_d, nbin)
H, xedges, yedges = np.histogram2d(interpoint[:, 1], interpoint[:, 2],
[xe, xe])
# H needs to be rotated and flipped
H = np.rot90(H)
H = np.flipud(H)
# Mask zeros if requested
if mask:
H = np.ma.masked_where(H == 0, H) # Mask pixels with a value of zero
# Plot 2D histogram using pcolor
if ax is None:
fig = plt.figure()
ax = fig.add_subplot(111)
if fmax is not None:
dd_sorted = sorted(np.ma.getdata(H).flat)
vmax = dd_sorted[int(len(dd_sorted) * fmax)]
quadmesh = ax.pcolormesh(xedges, yedges, H, vmax=vmax, cmap=cmap)
if colorbar:
cbar = plt.colorbar(quadmesh, ax=ax)
cbar.ax.set_ylabel('Counts')
def estimate_intensity(self):
self.intensity = {}
lookup_ind = []
lookup_data = []
for (i, itxy) in zip(self.ii, self.i_data):
if i in lookup_ind:
continue
lookup_ind.append(i)
lookup_data.append(itxy)
for (j, jtxy) in zip(self.jj, self.j_data):
if j in lookup_ind:
continue
lookup_ind.append(j)
lookup_data.append(jtxy)
# intensity = self.kde.pdf(CartesianSpaceTimeData(np.array(lookup_data)), normed=False)
intensity = self.kde.partial_marginal_pdf(CartesianSpaceTimeData(np.array(lookup_data)).space, normed=False) / self.T
self.intensity = dict(
[(ix, v) for ix, v in zip(lookup_ind, intensity)]
)
cid_all = np.arange(res_all.ndata)
else:
max_t = 90
max_d = 500
du = 50
dv = 5
number_nn = [15, 100]
simplification_tol = 20 # for speedup in intersection etc.
full_poly = cad.get_camden_region()
poly = full_poly.simplify(simplification_tol)
bdy = poly.boundary
res_all, t0, cid_all = cad.get_crimes_by_type(nicl_type=3)
res_all = CartesianSpaceTimeData(res_all)
cid_all = np.array(sorted(cid_all))
S = poly.area
T = np.ptp(res_all.time)
n = res_all.ndata
# estimate intensity with KDE
scott_spatial_bandwidth = res_all.space.data.std(
axis=0, ddof=1) * res_all.ndata**(-1. / float(2 + 4))
my_temporal_bandwidth = 0.1
k = kde_models.FixedBandwidthKdeSeparable(
res_all,
bandwidths=list(scott_spatial_bandwidth) + [my_temporal_bandwidth])
# k = kde_models.VariableBandwidthNnKdeSeparable(res_all, number_nn=number_nn)