def jump(self, xi, t):
n = xi.size(0)
# spatial multivariate gaussian
m_gauss = MultivariateNormal(self.mu_jump * torch.ones(n),
self.std_jump * torch.eye(n))
# poisson process, probability of arrival at time t
exp_d = Exponential(self.lambd)
# independent events, mult probabilities
p = torch.exp(m_gauss.log_prob(xi)) * (
1 - torch.exp(exp_d.log_prob(self.last_jump)))
# one sample from bernoulli trial
event = Bernoulli(p).sample([1])
if event:
coord_before = xi
xi = self.jump_event(xi, t) # flatten resulting sampled location
coord_after = xi
# saving jump coordinate info
self.log_jump(t, coord_before, coord_after)
self.last_jump = 0
# if no jump, increase counter for bern trial
else:
self.last_jump += 1
return xi
def test_Algorithms(self):
priors = Exponential(2.), Exponential(2.)
# ===== Test for multiple models ===== #
hidden1d = AffineModel((f0, g0), (f, g), priors, (self.linear.noise0, self.linear.noise))
oned = LinearGaussianObservations(hidden1d, 1., Exponential(1.))
hidden2d = AffineModel((f0mvn, g0mvn), (fmvn, gmvn), priors, (self.mvn.noise0, self.mvn.noise))
twod = LinearGaussianObservations(hidden2d, self.a, Exponential(1.))
# ====== Run inference ===== #
for trumod, model in [(self.model, oned), (self.mvnmodel, twod)]:
x, y = trumod.sample(550)
algs = [
(NESS, {'particles': 1000, 'filter_': SISR(model.copy(), 200)}),
(SMC2, {'particles': 1000, 'filter_': SISR(model.copy(), 200)}),
(NESSMC2, {'particles': 1000, 'filter_': SISR(model.copy(), 200)}),
(IteratedFilteringV2, {'particles': 1000, 'filter_': SISR(model.copy(), 1000)})
]
for alg, props in algs:
alg = alg(**props).initialize()
alg = alg.fit(y)
parameter = alg.filter.ssm.hidden.theta[-1]
kde = parameter.get_kde(transformed=False)
tru_val = trumod.hidden.theta[-1]
densval = kde.logpdf(tru_val.numpy().reshape(-1, 1))
priorval = parameter.dist.log_prob(tru_val)
assert bool(densval > priorval.numpy())
def test_exponential_shape_tensor_param(self):
expon = Exponential(torch.Tensor([1, 1]))
self.assertEqual(expon._batch_shape, torch.Size((2,)))
self.assertEqual(expon._event_shape, torch.Size(()))
self.assertEqual(expon.sample().size(), torch.Size((2,)))
self.assertEqual(expon.sample((3, 2)).size(), torch.Size((3, 2, 2)))
self.assertEqual(expon.log_prob(self.tensor_sample_1).size(), torch.Size((3, 2)))
self.assertRaises(ValueError, expon.log_prob, self.tensor_sample_2)
def test_exponential_shape_scalar_param(self):
expon = Exponential(1.)
self.assertEqual(expon._batch_shape, torch.Size())
self.assertEqual(expon._event_shape, torch.Size())
self.assertEqual(expon.sample().size(), torch.Size((1,)))
self.assertEqual(expon.sample((3, 2)).size(), torch.Size((3, 2)))
self.assertRaises(ValueError, expon.log_prob, self.scalar_sample)
self.assertEqual(expon.log_prob(self.tensor_sample_1).size(), torch.Size((3, 2)))
self.assertEqual(expon.log_prob(self.tensor_sample_2).size(), torch.Size((3, 2, 3)))
def test_ParameterInDistribution(self):
shape = 10, 100
a = 1e-2 * torch.ones((shape[0], 1))
dt = 1e-2
dist = Normal(loc=0., scale=Parameter(Exponential(10.)))
init = Normal(a, 1.)
sde = AffineEulerMaruyama((f_sde, g_sde), (a, 0.15),
init,
dist,
dt=dt,
num_steps=10)
sde.sample_params(shape)
# ===== Initialize ===== #
x = sde.i_sample(shape)
# ===== Propagate ===== #
num = 1000
samps = [x]
for t in range(num):
samps.append(sde.propagate(samps[-1]))
samps = torch.stack(samps)
self.assertEqual(samps.size(), torch.Size([num + 1, *shape]))
# ===== Sample path ===== #
path = sde.sample_path(num + 1, shape)
self.assertEqual(samps.shape, path.shape)
def test_Inference(self):
# ===== Distributions ===== #
dist = Normal(0., 1.)
mvn = Independent(Normal(torch.zeros(2), torch.ones(2)), 1)
# ===== Define model ===== #
linear = AffineProcess((f, g), (1., 0.25), dist, dist)
model = LinearGaussianObservations(linear, scale=0.1)
mv_linear = AffineProcess((fmvn, gmvn), (0.5, 0.25), mvn, mvn)
mvnmodel = LinearGaussianObservations(mv_linear, torch.eye(2), scale=0.1)
# ===== Test for multiple models ===== #
priors = Exponential(1.), LogNormal(0., 1.)
hidden1d = AffineProcess((f, g), priors, dist, dist)
oned = LinearGaussianObservations(hidden1d, 1., scale=0.1)
hidden2d = AffineProcess((fmvn, gmvn), priors, mvn, mvn)
twod = LinearGaussianObservations(hidden2d, torch.eye(2), scale=0.1 * torch.ones(2))
particles = 1000
# ====== Run inference ===== #
for trumod, model in [(model, oned), (mvnmodel, twod)]:
x, y = trumod.sample_path(1000)
algs = [
(NESS, {'particles': particles, 'filter_': APF(model.copy(), 200)}),
(NESS, {'particles': particles, 'filter_': UKF(model.copy())}),
(SMC2, {'particles': particles, 'filter_': APF(model.copy(), 200)}),
(SMC2FW, {'particles': particles, 'filter_': APF(model.copy(), 200)}),
(NESSMC2, {'particles': particles, 'filter_': APF(model.copy(), 200)})
]
for alg, props in algs:
alg = alg(**props).initialize()
alg = alg.fit(y)
w = normalize(alg._w_rec if hasattr(alg, '_w_rec') else torch.ones(particles))
tru_params = trumod.hidden.theta._cont + trumod.observable.theta._cont
inf_params = alg.filter.ssm.hidden.theta._cont + alg.filter.ssm.observable.theta._cont
for trup, p in zip(tru_params, inf_params):
if not p.trainable:
continue
kde = p.get_kde(weights=w)
transed = p.bijection.inv(trup)
densval = kde.logpdf(transed.numpy().reshape(-1, 1))
priorval = p.distr.log_prob(trup)
assert (densval > priorval.numpy()).all()
def _make_ma_mdp(self):
joint_action_shape = self.joint_action_shape
n_states = self.n_states
n_agents = len(joint_action_shape)
rand = self.rand
# Reward perturbation
perturbation = mu.unsqueeze(
self.reward_perturbation, -1,
n_states + 3 - self.reward_perturbation.dim())
# Generate transition probability tensor
trans_prob = th.rand(n_states,
*joint_action_shape,
n_states,
generator=rand)
# Acyclic (episodic) MDP
if self.acyclic:
states_idx, next_states_idx = th.tril_indices(n_states)
trans_prob[states_idx, ..., next_states_idx] = 0
# Normalize transition probability matrix
trans_prob /= trans_prob.sum(dim=-1, keepdim=True)
trans_prob[th.isnan(trans_prob)] = 0
# Generate random reward (following method ensures enough variance in rewards)
# 1) Generate rewards "core" for state, joint actions and agents
rewards = th.randn(n_states,
*joint_action_shape,
1,
n_agents,
generator=rand)
# 2) Multiply "core" by scales to generate different rewards for next state
scales_dist = Exponential(th.tensor(1.))
with mu.use_rand(rand):
rewards *= scales_dist.sample(
(n_states, *joint_action_shape, n_states, n_agents))
# 3) Correlate rewards
rewards = rewards @ self.reward_correlation
## Transition probability
self._trans_prob = trans_prob
## Rewards for state-joint actions
self._rewards = rewards
def forward(self, theta, n_samp=1):
n_exp = theta.shape[0]
n_samp = torch.Size([n_samp, 1])
unit = torch.ones(n_exp).to(self.device)
with torch.autograd.no_grad():
d0 = Normal(theta[:, 0], unit)
z0 = d0.sample(n_samp).permute(2, 0, 1)
d1 = Normal(3 * unit, torch.exp(theta[:, 1] / 3))
z1 = d1.sample(n_samp).permute(2, 0, 1)
d2_1 = Normal(-2 * unit, unit)
d2_2 = Normal(2 * unit, .5 * unit)
d2_b = Bernoulli(.5 * unit)
z2_b = d2_b.sample(n_samp).float()
# Gaussian Mixture
z2 = ((z2_b * d2_1.sample(n_samp) +
(1 - z2_b) * d2_2.sample(n_samp)).permute(2, 0, 1))
d3 = Uniform(-5 * unit, theta[:, 2])
z3 = d3.sample(n_samp).permute(2, 0, 1)
d4 = Exponential(.5 * unit)
z4 = d4.sample(n_samp).permute(2, 0, 1)
z = torch.cat((z0, z1, z2, z3, z4), 2)
X = torch.matmul(self.R,
z.view(-1, 5).unsqueeze(2)).view(n_exp, -1, 5)
return X
def test_Parameter(self):
# ===== Start stuff ===== #
param = Parameter(Normal(0., 1.))
param.sample_(1000)
assert param.shape == torch.Size([1000])
# ===== Construct view ===== #
view = param.view(1000, 1)
# ===== Change values ===== #
param.values = torch.empty_like(param).normal_()
assert (view[:, 0] == param).all()
# ===== Have in tuple ===== #
vals = (param.view(1000, 1, 1), )
param.values = torch.empty_like(param).normal_()
assert (vals[0][:, 0, 0] == param).all()
# ===== Set t_values ===== #
view = param.view(1000, 1)
param.t_values = torch.empty_like(param).normal_()
assert (view[:, 0] == param.t_values).all()
# ===== Check we cannot set different shape ===== #
with self.assertRaises(ValueError):
param.values = torch.empty(1).normal_()
# ===== Check that we cannot set out of bounds values for parameter ===== #
positive = Parameter(Exponential(1.))
positive.sample_(1)
with self.assertRaises(ValueError):
positive.values = -torch.empty_like(positive).normal_().abs()
# ===== Check that we can set transformed values ===== #
values = torch.empty_like(positive).normal_()
positive.t_values = values
assert (positive == positive.bijection(values)).all()
if (t > 10
and abs(self.loss_hist[-1] / self.loss_hist[-2] - 1) < tol
and verbose):
print('Convergence detected!')
break
if __name__ == '__main__':
torch.manual_seed(0)
theta_true = 5.3
# NOTE: see n=3, n=10, n=30, n=1000
# For small n, the MAP estimate with jacobian is far from truth. As n
# increases, the difference is small, because the posterior gets more
# peaked.
n = 5
y = Exponential(theta_true).sample((n, ))
model = Model(y)
model.fit(lr=.01)
model_with_jacobian = Model(y, use_jacobian=True)
model_with_jacobian.fit(lr=.01)
true_map_est = ((n + model.prior_shape - 1) /
(model.prior_rate + model.y_sum))
map_est = model.log_theta.exp()
map_est_with_jacobian = model_with_jacobian.log_theta.exp()
print('True MAP est: {}'.format(true_map_est))
print('MAP est (w/o jacobian): {}'.format(map_est))
def test_exponential_sample(self):
self._set_rng_seed(1)
for rate in [1e-5, 1.0, 10.]:
self._check_sampler_sampler(Exponential(rate),
scipy.stats.expon(scale=1. / rate),
'Exponential(rate={})'.format(rate))
def test_exponential(self):
rate = Variable(torch.randn(5, 5).abs(), requires_grad=True)
rate_1d = Variable(torch.randn(1).abs(), requires_grad=True)
self.assertEqual(Exponential(rate).sample().size(), (5, 5))
self.assertEqual(Exponential(rate).sample((7, )).size(), (7, 5, 5))
self.assertEqual(Exponential(rate_1d).sample((1, )).size(), (1, 1))
self.assertEqual(Exponential(rate_1d).sample().size(), (1, ))
self.assertEqual(Exponential(0.2).sample((1, )).size(), (1, ))
self.assertEqual(Exponential(50.0).sample((1, )).size(), (1, ))
self._gradcheck_log_prob(Exponential, (rate, ))
state = torch.get_rng_state()
eps = rate.new(rate.size()).exponential_()
torch.set_rng_state(state)
z = Exponential(rate).rsample()
z.backward(torch.ones_like(z))
self.assertEqual(rate.grad, -eps / rate**2)
rate.grad.zero_()
self.assertEqual(z.size(), (5, 5))
def ref_log_prob(idx, x, log_prob):
m = rate.data.view(-1)[idx]
expected = math.log(m) - m * x
self.assertAlmostEqual(log_prob, expected, places=3)
self._check_log_prob(Exponential(rate), ref_log_prob)
def likelihood(self, x, z_obj):
"""Evaluate likelihood of x under model.
Args:
x (torch.Tensor), (n, T, c, w, h) The given sequence of observations.
z_obj (torch.Tensor), (nTO, 4): Samples from z distribution.
Returns:
log_p_xz (torch.Tensor), (nT): Image likelihood.
"""
# reshape x to merge batch and sequences to pseudo batch since spn
# will work on image basis shape (n4, c, w, h)
x_img = x.flatten(end_dim=1)
# 1. Background Likelihood
# reshape to (n4, o, 4) and extract marginalisation information
z_img = z_obj.view(-1, self.c.num_obj, 4)
marginalise_patch, marginalise_bg, overlap_ratios = self.masks_from_z(
z_img)
# flatten, st. shape is (n4, cwh) for both
img_flat = x_img.flatten(start_dim=1)
marg_flat = marginalise_bg.flatten(start_dim=1)
# get likelihood of background under bg_spn, output from (n4, 1) to (n4)
bg_loglik = self.bg_spn.forward(img_flat, marg_flat)[:, 0]
# 2. Patch Likelihoods
# extract patches (n4o, c, w, h) from transformer
patches = self.patches_from_z(x_img, z_obj)
# flatten into (n4o, c w_out h_out)
patches_flat = patches.flatten(start_dim=1)
marginalise_flat = marginalise_patch.flatten(start_dim=1)
# (n4o)
patches_loglik = self.obj_spn.forward(patches_flat,
marginalise_flat)[:, 0]
# scale patch_likelihoods by size of patch to obtain
# well calibrated likelihoods
patches_loglik = patches_loglik * z_obj[:, 0] * z_obj[:, 1]
# shape n4o to n4,o
patches_loglik = patches_loglik.view(-1, self.c.num_obj)
# 3. Add Exponential overlap_penalty
overlap_prior = Exponential(self.c.overlap_beta)
overlap_log_liks = overlap_prior.log_prob(overlap_ratios)
# 4. Assemble final img likelihood E_q(z|x)[log p(x, z)]
# expectation is approximated by a single sample
patches_loglik = patches_loglik.sum(1)
overlap_log_liks = overlap_log_liks.sum(1)
scores = [bg_loglik, patches_loglik, overlap_log_liks]
scores = torch.stack(scores, -1)
# shape (n4)
log_p_xz = scores.sum(-1)
if ((self.step_counter % self.c.print_every == 0)
or (self.step_counter % self.c.plot_every == 0)):
if self.c.debug:
self.prop_dict['bg'] = bg_loglik.mean().detach()
self.prop_dict['patch'] = patches_loglik.mean().detach()
self.prop_dict['overlap'] = overlap_log_liks.mean().detach()
if self.c.debug and self.c.debug_extend_plots:
self.prop_dict['overlap_ratios'] = overlap_ratios.detach()
self.prop_dict['patches'] = patches.detach()
self.prop_dict['marginalise_flat'] = marginalise_flat.detach()
self.prop_dict['patches_loglik'] = patches_loglik.detach()
self.prop_dict['marginalise_bg'] = marginalise_bg.detach()
self.prop_dict['bg_loglik'] = bg_loglik.detach()
return log_p_xz, self.prop_dict
assert isinstance(transform_to_inspect, target_transform)
samples = prior.sample((2, ))
transformed_samples = transform(samples)
assert torch.allclose(samples, transform.inv(transformed_samples))
@pytest.mark.parametrize(
"prior, enable_transform",
(
(BoxUniform(zeros(5), ones(5)), True),
(BoxUniform(zeros(1), ones(1)), True),
(BoxUniform(zeros(5), ones(5)), False),
(MultivariateNormal(zeros(5), eye(5)), True),
(Exponential(rate=ones(1)), True),
(LogNormal(zeros(1), ones(1)), True),
(
MultipleIndependent(
[Exponential(rate=ones(1)),
BoxUniform(zeros(5), ones(5))]),
True,
),
),
)
def test_mcmc_transform(prior, enable_transform):
"""
Test whether the transform for MCMC returns the log_abs_det in the correct shape.
"""
num_samples = 1000
def sample(self, size):
m = Exponential(torch.tensor([1.0]))
return m.sample(size)
def test_exponential_sample(self):
set_rng_seed(1) # see Note [Randomized statistical tests]
for rate in [1e-5, 1.0, 10.]:
self._check_sampler_sampler(Exponential(rate),
scipy.stats.expon(scale=1. / rate),
'Exponential(rate={})'.format(rate))