def generate_synth_data(V=10, D=10, T=5, N=100, alpha_beta=10., alpha_theta=10., plot=False, train_frac=0.5):
# true_lda = StandardLDA(T, V, alpha_beta=alpha_beta, alpha_theta=alpha_theta)
true_lda = LogisticNormalCorrelatedLDA(T, V, alpha_beta=alpha_beta)
print("Sigma: ", true_lda.Sigma)
# true_lda = StickbreakingCorrelatedLDA(T, V, alpha_beta=alpha_beta)
data = np.zeros((D,V),dtype=int)
for d in xrange(D):
doc = true_lda.generate(N=N, keep=True)
data[d,:] = doc.w
if plot:
plt.figure()
plt.imshow(data, interpolation="none")
plt.xlabel("Vocabulary")
plt.ylabel("Documents")
plt.colorbar()
plt.show()
# Split each document into two
train_data = np.zeros_like(data)
test_data = np.zeros_like(data)
for d,w in enumerate(data):
# Get vector where i is repeated w[i] times
wcnt = ibincount(w)
# Subsample wcnt
train_inds = np.random.rand(wcnt.size) < train_frac
train_data[d] = np.bincount(wcnt[train_inds], minlength=V)
test_data[d] = np.bincount(wcnt[~train_inds], minlength=V)
assert np.allclose(train_data[d] + test_data[d], w)
return true_lda, train_data, test_data
def generate_dataset(bias=1.):
# Create the model with these parameters
network_hypers = {
'C': 1,
'kappa': 1.,
'c': np.zeros(K, dtype=np.int),
'p': 1 * np.ones((1, 1)),
'v': 100.
}
bkgd_hypers = {"alpha": 3., "beta": 3. / bias}
dt_model = pyhawkes.models.\
DiscreteTimeNetworkHawkesModelSpikeAndSlab(K=K, dt=dt, dt_max=dt_max, B=B,
bkgd_hypers=bkgd_hypers,
network_hypers=network_hypers)
# dt_model.bias_model.lambda0 = bias * np.ones(K)
assert dt_model.check_stability()
S_dt, _ = dt_model.generate(T=int(np.ceil(T / dt)), keep=False)
print "sampled dataset with ", S_dt.sum(), "events"
# Convert S_dt to continuous time
S_ct = dt * np.concatenate([ibincount(S) for S in S_dt.T]).astype(float)
S_ct += dt * np.random.rand(*S_ct.shape)
assert np.all(S_ct < T)
C_ct = np.concatenate([k * np.ones(S.sum())
for k, S in enumerate(S_dt.T)]).astype(int)
# Sort the data
perm = np.argsort(S_ct)
S_ct = S_ct[perm]
C_ct = C_ct[perm]
return S_dt, S_ct, C_ct
def generate_dataset(bias=1.):
# Create the model with these parameters
network_hypers = {'C': 1, 'kappa': 1., 'c': np.zeros(K, dtype=np.int), 'p': 1*np.ones((1,1)), 'v': 100.}
bkgd_hypers = {"alpha": 3., "beta": 3./bias}
dt_model = pyhawkes.models.\
DiscreteTimeNetworkHawkesModelSpikeAndSlab(K=K, dt=dt, dt_max=dt_max, B=B,
bkgd_hypers=bkgd_hypers,
network_hypers=network_hypers)
# dt_model.bias_model.lambda0 = bias * np.ones(K)
assert dt_model.check_stability()
S_dt,_ = dt_model.generate(T=int(np.ceil(T/dt)), keep=False)
print "sampled dataset with ", S_dt.sum(), "events"
# Convert S_dt to continuous time
S_ct = dt * np.concatenate([ibincount(S) for S in S_dt.T]).astype(float)
S_ct += dt * np.random.rand(*S_ct.shape)
assert np.all(S_ct < T)
C_ct = np.concatenate([k*np.ones(S.sum()) for k,S in enumerate(S_dt.T)]).astype(int)
# Sort the data
perm = np.argsort(S_ct)
S_ct = S_ct[perm]
C_ct = C_ct[perm]
return S_dt, S_ct, C_ct
def convert_discrete_to_continuous(S, dt):
# Convert S to continuous time
from pybasicbayes.util.general import ibincount
T = S.shape[0] * dt
S_ct = dt * np.concatenate([ibincount(Sk) for Sk in S.T]).astype(float)
S_ct += dt * np.random.rand(*S_ct.shape)
assert np.all(S_ct < T)
C_ct = np.concatenate([k*np.ones(Sk.sum()) for k,Sk in enumerate(S.T)]).astype(int)
# Sort the data
perm = np.argsort(S_ct)
S_ct = S_ct[perm]
C_ct = C_ct[perm]
return S_ct, C_ct, T
def convert_discrete_to_continuous(S, dt):
# Convert S to continuous time
from pybasicbayes.util.general import ibincount
T = S.shape[0] * dt
S_ct = dt * np.concatenate([ibincount(Sk) for Sk in S.T]).astype(float)
S_ct += dt * np.random.rand(*S_ct.shape)
assert np.all(S_ct < T)
C_ct = np.concatenate([k*np.ones(Sk.sum()) for k,Sk in enumerate(S.T)]).astype(int)
# Sort the data
perm = np.argsort(S_ct)
S_ct = S_ct[perm]
C_ct = C_ct[perm]
return S_ct, C_ct, T
def downsample_data(X, n):
"""
Downsample each row of X such that it sums to n by randomly removing entries
"""
from pybasicbayes.util.general import ibincount
assert X.ndim == 2
D,K = X.shape
Xsub = X.copy().astype(np.int)
for d in xrange(D):
xi = ibincount(Xsub[d])
Xsub[d] = np.bincount(np.random.choice(xi, size=n, replace=False), minlength=K)
assert Xsub[d].sum() == n
return Xsub.astype(np.float)
def downsample_data(X, n):
"""
Downsample each row of X such that it sums to n by randomly removing entries
"""
from pybasicbayes.util.general import ibincount
assert X.ndim == 2
D,K = X.shape
Xsub = X.copy().astype(np.int)
for d in range(D):
xi = ibincount(Xsub[d])
Xsub[d] = np.bincount(np.random.choice(xi, size=n, replace=False), minlength=K)
assert Xsub[d].sum() == n
return Xsub.astype(np.float)
dt_max = 10.
T = 100.
network_hypers = {'kappa': 1., 'p': 1., 'v': 10.}
dt_model = pyhawkes.models.\
DiscreteTimeNetworkHawkesModelSpikeAndSlab(K=K, dt=dt, dt_max=dt_max, B=B,
network_hypers=network_hypers)
assert dt_model.check_stability()
S_dt, _ = dt_model.generate(T=int(np.ceil(T / dt)), keep=False)
print("sampled dataset with ", S_dt.sum(), "events")
print("DT LL: ", dt_model.heldout_log_likelihood(S_dt))
# Convert S_test to continuous time
S_ct = dt * np.concatenate([ibincount(S) for S in S_dt.T]).astype(float)
S_ct += dt * np.random.rand(*S_ct.shape)
assert np.all(S_ct < T)
C_ct = np.concatenate([k * np.ones(S.sum())
for k, S in enumerate(S_dt.T)]).astype(int)
# Sort the data
perm = np.argsort(S_ct)
S_ct = S_ct[perm]
C_ct = C_ct[perm]
ct_model = pyhawkes.models.ContinuousTimeNetworkHawkesModel(
K, dt_max=1., network_hypers=network_hypers)
ct_model.add_data(S_ct, C_ct, T)
# ct.resample_model()
dt_max = 10.
T = 100.
network_hypers = {'kappa': 1., 'p': 1., 'v': 10.}
dt_model = pyhawkes.models.\
DiscreteTimeNetworkHawkesModelSpikeAndSlab(K=K, dt=dt, dt_max=dt_max, B=B,
network_hypers=network_hypers)
assert dt_model.check_stability()
S_dt,_ = dt_model.generate(T=int(np.ceil(T/dt)), keep=False)
print("sampled dataset with ", S_dt.sum(), "events")
print("DT LL: ", dt_model.heldout_log_likelihood(S_dt))
# Convert S_test to continuous time
S_ct = dt * np.concatenate([ibincount(S) for S in S_dt.T]).astype(float)
S_ct += dt * np.random.rand(*S_ct.shape)
assert np.all(S_ct < T)
C_ct = np.concatenate([k*np.ones(S.sum()) for k,S in enumerate(S_dt.T)]).astype(int)
# Sort the data
perm = np.argsort(S_ct)
S_ct = S_ct[perm]
C_ct = C_ct[perm]
ct_model = pyhawkes.models.ContinuousTimeNetworkHawkesModel(K, dt_max=1.,
network_hypers=network_hypers)
ct_model.add_data(S_ct, C_ct, T)
# ct.resample_model()
# Hard code parameters
