def sampling(self, sess, batch_size):
"""fetch model parameters, and generate samples accordingly."""
# get current model parameters
mu, beta = sess.run([self.mu, self.beta])
Wss = sess.run(self.Wss)
bss = sess.run(self.bss)
Wphis = sess.run(self.Wphis)
# construct kernel function and conditional intensity lambda
kernel = GaussianMixtureDiffusionKernel(self.n_comp,
layers=self.layers[1:-1],
beta=beta,
C=self.C,
SIGMA_SHIFT=self.SIGMA_SHIFT,
SIGMA_SCALE=self.SIGMA_SCALE,
MU_SCALE=self.MU_SCALE,
Wss=Wss,
bss=bss,
Wphis=Wphis)
lam = HawkesLam(mu, kernel, maximum=self.maximum)
# sampling points given model parameters
pp = SpatialTemporalPointProcess(lam)
seqs, sizes = pp.generate(T=self.T,
S=self.S,
batch_size=batch_size,
verbose=self.verbose)
return seqs
def test_pretrain_gaussian_mixture_diffusion():
'''
Test Spatio-Temporal Point Process Generator equipped with
pretrained Gaussian mixture diffusion kernel
'''
params = np.load('data/mle_gaussian_mixture_params.npz')
mu = params['mu']
beta = params['beta']
kernel = GaussianMixtureDiffusionKernel(
n_comp=5, layers=[5], C=1., beta=beta,
SIGMA_SHIFT=.05, SIGMA_SCALE=.2, MU_SCALE=.01,
Wss=params['Wss'], bss=params['bss'], Wphis=params['Wphis'])
lam = HawkesLam(mu, kernel, maximum=1e+3)
pp = SpatialTemporalPointProcess(lam)
# # generate points
# points, sizes = pp.generate(
# T=[0., 10.], S=[[-1., 1.], [-1., 1.]],
# batch_size=2, verbose=True)
# print(points.shape)
# print(sizes)
# read or save to local npy file.
points = np.load('data/apd.robbery.permonth.npy')
da = DataAdapter(init_data=points)
points = da.normalize(points)
# np.save('results/gaussian_hpp_Mar_15_layer_5.npy', points)
# plot intensity of the process over the time
plot_spatial_intensity(lam, points[0], S=[[0., 10.], [-1., 1.], [-1., 1.]],
t_slots=1000, grid_size=50, interval=50)
def test_random_gaussian_mixture_diffusion():
'''
Test Spatio-Temporal Point Process Generator equipped with
random Gaussian mixture diffusion kernel
'''
mu = .2
kernel = GaussianMixtureDiffusionKernel(
n_comp=5, layers=[5, 5], C=1., beta=1.,
SIGMA_SHIFT=.2, SIGMA_SCALE=.05, MU_SCALE=.05)
lam = HawkesLam(mu, kernel, maximum=1e+3)
pp = SpatialTemporalPointProcess(lam)
# generate points
points, sizes = pp.generate(
T=[0., 10.], S=[[-1., 1.], [-1., 1.]],
batch_size=2, verbose=True)
print(points.shape)
print(sizes)
# read or save to local npy file.
# points = np.load('results/free_hpp_Mar_15_layer_5.npy')
# np.save('results/gaussian_hpp_Mar_15_layer_5.npy', points)
# plot intensity of the process over the time
plot_spatial_intensity(lam, points[0], S=[[0., 10.], [-1., 1.], [-1., 1.]],
t_slots=1000, grid_size=50, interval=50)
def test_std_diffusion():
'''
Test Spatio-Temporal Point Process Generator equipped with
standard diffusion kernel
'''
# parameters initialization
mu = .1
kernel = StdDiffusionKernel(C=1., beta=1., sigma_x=.1, sigma_y=.1)
lam = HawkesLam(mu, kernel, maximum=1e+3)
pp = SpatialTemporalPointProcess(lam)
# generate points
points, sizes = pp.generate(
T=[0., 10.], S=[[-1., 1.], [-1., 1.]],
batch_size=100, verbose=True)
print(points)
print(sizes)
# read or save to local npy file.
# points = np.load('results/tf_thining_samples.npy')
np.save('results/hpp_Feb_25.npy', points)
# plot intensity of the process over the time
plot_spatial_intensity(lam, points[0], S=[[0., 10.], [-1., 1.], [-1., 1.]],
t_slots=1000, grid_size=50, interval=50)
def test_gaussian_diffusion():
'''
Test Spatio-Temporal Point Process Generator equipped with
Gaussian diffusion kernel
'''
mu = .1
kernel = GaussianDiffusionKernel(
layers=[5, 5], C=1., beta=1.,
SIGMA_SHIFT=.2, SIGMA_SCALE=.05, MU_SCALE=.1, is_centered=True)
lam = HawkesLam(mu, kernel, maximum=1e+3)
pp = SpatialTemporalPointProcess(lam)
print(kernel.Ws)
print(kernel.bs)
# plot kernel parameters over the spatial region.
plot_spatial_kernel("results/kernel.pdf", kernel, S=[[-1., 1.], [-1., 1.]], grid_size=50)
# generate points
points, sizes = pp.generate(
T=[0., 10.], S=[[-1., 1.], [-1., 1.]],
batch_size=2, verbose=True)
print(points)
print(sizes)
# read or save to local npy file.
# points = np.load('results/free_hpp_Mar_15_layer_5.npy')
# np.save('results/gaussian_hpp_Mar_15_layer_5.npy', points)
# plot intensity of the process over the time
plot_spatial_intensity(lam, points[0], S=[[0., 10.], [-1., 1.], [-1., 1.]],
t_slots=1000, grid_size=50, interval=50)
def exp_simulation():
data = np.load(
'../Spatio-Temporal-Point-Process-Simulator/results/spatial-variant-gaussian-b.npy'
)
params = np.load(
'../Spatio-Temporal-Point-Process-Simulator/data/simulation-1-gcomp-b.npz'
)
mu = params['mu']
beta = params['beta']
kernel = GaussianMixtureDiffusionKernel(n_comp=1,
layers=[10],
C=1.,
beta=beta,
SIGMA_SHIFT=.1,
SIGMA_SCALE=.5,
MU_SCALE=.01,
Wss=params['Wss'],
bss=params['bss'],
Wphis=params['Wphis'])
lam = HawkesLam(mu, kernel, maximum=1e+3)
print("mu", mu)
print("beta", beta)
# plot kernel parameters over the spatial region.
utils.plot_spatial_kernel("results/learned-kernel-svgau-b.pdf",
kernel.gdks[0],
S=[[-1., 1.], [-1., 1.]],
grid_size=50)
def _sampling(self, sess, T, S, batch_size):
"""
"""
# get current model parameters
mu, beta, sigma_x, sigma_y = sess.run([self.mu, self.beta, self.sigma_x, self.sigma_y])
print("[%s] mu=%.2f, beta=%.2f, sigma_x=%.2f, sigma_y=%.2f." % (arrow.now(), mu, beta, sigma_x, sigma_y), file=sys.stderr)
# sampling points given model parameters
kernel = DiffusionKernel(beta=beta, sigma_x=sigma_x, sigma_y=sigma_y)
lam = HawkesLam(mu, kernel, maximum=self.maximum)
pp = SpatialTemporalPointProcess(lam)
points, sizes = pp.generate(T=T, S=S, batch_size=batch_size)
return points, sizes
def exp_real_with_1_comp():
# REAL
data = np.load('data/real/SCEDC-1999-2019-24hrs.npy')
data = data[:, 1:, :3]
params = np.load('results/real-24hrs-1-gcomp.npz')
da = utils.DataAdapter(init_data=data)
mu = params['mu']
beta = params['beta']
kernel = GaussianMixtureDiffusionKernel(n_comp=1,
layers=[10],
C=1.,
beta=beta,
SIGMA_SHIFT=.1,
SIGMA_SCALE=.5,
MU_SCALE=.01,
Wss=params['Wss'],
bss=params['bss'],
Wphis=params['Wphis'])
lam = HawkesLam(mu, kernel, maximum=1e+3)
print("mu", mu)
print("beta", beta)
utils.plot_spatial_kernel(
"results/learned-kernel-SCEDC-1999-2019-24hrs.pdf",
kernel.gdks[0],
S=[[-1., 1.], [-1., 1.]],
grid_size=50)
utils.spatial_intensity_on_map(
"results/map-SCEDC-1999-2019-24hrs.html",
da,
lam,
data,
seq_ind=3000,
t=8.0,
# xlim=[-23.226, -22.621],
# ylim=[-43.868, -43.050],
# ngrid=200)
xlim=da.xlim,
ylim=da.ylim,
ngrid=200)
def exp_real_with_2_comp():
# REAL
data = np.load(
'../Spatio-Temporal-Point-Process-Simulator/data/rescale.ambulance.perday.npy'
)
data = data[:, 1:, :3]
params = np.load(
'../Spatio-Temporal-Point-Process-Simulator/data/rescale_ambulance_mle_gaussian_mixture_params.npz'
)
da = utils.DataAdapter(init_data=data)
mu = .1 # params['mu']
beta = params['beta']
kernel = GaussianMixtureDiffusionKernel(n_comp=1,
layers=[5],
C=1.,
beta=beta,
SIGMA_SHIFT=.1,
SIGMA_SCALE=.5,
MU_SCALE=.01,
Wss=params['Wss'],
bss=params['bss'],
Wphis=params['Wphis'])
lam = HawkesLam(mu, kernel, maximum=1e+3)
print("mu", mu)
print("beta", beta)
print(params['Wphis'].shape)
pp = SpatialTemporalPointProcess(lam)
# generate points
points, sizes = pp.generate(T=[0., 10.],
S=[[-1., 1.], [-1., 1.]],
batch_size=100,
verbose=True)
results = da.restore(points)
print(results)
print(sizes)
np.save('results/ambulance-simulation.npy', results)
beta = params['beta']
kernel = GaussianMixtureDiffusionKernel(n_comp=5,
layers=[5],
C=1.,
beta=beta,
SIGMA_SHIFT=.05,
SIGMA_SCALE=.2,
MU_SCALE=.01,
Wss=params['Wss'],
bss=params['bss'],
Wphis=params['Wphis'])
# kernel = GaussianMixtureDiffusionKernel(
# n_comp=20, layers=[5], C=1., beta=.8,
# SIGMA_SHIFT=.05, SIGMA_SCALE=.3, MU_SCALE=.03)
# kernel = StdDiffusionKernel(C=1., beta=.15, sigma_x=.15, sigma_y=.15)
lam = HawkesLam(mu, kernel, maximum=1e+3)
# print(da.normalize(data)[0])
print("mu", mu)
print("beta", beta)
# utils.spatial_intensity_on_map(
# "results/ambulance_intensity_map-2.81.html", da, lam, data, seq_ind=0, t=2.81,
# # # crime range
# # xlim=[33.70, 33.87],
# # ylim=[-84.50, -84.30],
# # # earthquake range
# # xlim=[25.692, 49.687],
# # ylim=[-129.851, -111.094],
# # ambulance range
# xlim=[-23.226, -22.621],
