def test_():
true_parameters = np.array([1.0, 5.0, 0.2])
num_simulations = 1000
parameters = np.tile(true_parameters, num_simulations).reshape(-1, 3)
observations = Stats().calc(Model().sim(parameters))
print(parameters.shape, observations.shape)
utils.plot_hist_marginals(observations,
ground_truth=get_ground_truth_observation(),
show_xticks=True)
plt.show()
def _test():
# prior = MG1Uniform(low=torch.zeros(3), high=torch.Tensor([10, 10, 1 / 3]))
# uniform = distributions.Uniform(
# low=torch.zeros(3), high=torch.Tensor([10, 10, 1 / 3])
# )
# x = torch.Tensor([10, 20, 1 / 3]).reshape(1, -1)
# print(uniform.log_prob(x))
# print(prior.log_prob(x))
d = LotkaVolterraOscillating()
samples = d.sample((1000, ))
utils.plot_hist_marginals(utils.tensor2numpy(samples), lims=[-6, 3])
plt.show()
def test_():
task = "nonlinear-gaussian"
simulator, prior = simulators.get_simulator_and_prior(task)
parameter_dim, observation_dim = (
simulator.parameter_dim,
simulator.observation_dim,
)
true_observation = simulator.get_ground_truth_observation()
neural_posterior = utils.get_neural_posterior("maf", parameter_dim,
observation_dim, simulator)
apt = APT(
simulator=simulator,
true_observation=true_observation,
prior=prior,
neural_posterior=neural_posterior,
num_atoms=-1,
use_combined_loss=False,
train_with_mcmc=False,
mcmc_method="slice-np",
summary_net=None,
retrain_from_scratch_each_round=False,
discard_prior_samples=False,
)
num_rounds, num_simulations_per_round = 20, 1000
apt.run_inference(num_rounds=num_rounds,
num_simulations_per_round=num_simulations_per_round)
samples = apt.sample_posterior(2500)
samples = utils.tensor2numpy(samples)
figure = utils.plot_hist_marginals(
data=samples,
ground_truth=utils.tensor2numpy(
simulator.get_ground_truth_parameters()).reshape(-1),
lims=simulator.parameter_plotting_limits,
)
figure.savefig(
os.path.join(utils.get_output_root(), "corner-posterior-apt.pdf"))
samples = apt.sample_posterior_mcmc(num_samples=1000)
samples = utils.tensor2numpy(samples)
figure = utils.plot_hist_marginals(
data=samples,
ground_truth=utils.tensor2numpy(
simulator.get_ground_truth_parameters()).reshape(-1),
lims=simulator.parameter_plotting_limits,
)
figure.savefig(
os.path.join(utils.get_output_root(), "corner-posterior-apt-mcmc.pdf"))
def test_():
# if torch.cuda.is_available():
# device = torch.device("cuda")
# torch.set_default_tensor_type("torch.cuda.FloatTensor")
# else:
# device = torch.device("cpu")
# torch.set_default_tensor_type("torch.FloatTensor")
loc = torch.Tensor([0, 0])
covariance_matrix = torch.Tensor([[1, 0.99], [0.99, 1]])
likelihood = distributions.MultivariateNormal(
loc=loc, covariance_matrix=covariance_matrix)
bound = 1.5
low, high = -bound * torch.ones(2), bound * torch.ones(2)
prior = distributions.Uniform(low=low, high=high)
# def potential_function(inputs_dict):
# parameters = next(iter(inputs_dict.values()))
# return -(likelihood.log_prob(parameters) + prior.log_prob(parameters).sum())
prior = distributions.Uniform(low=-5 * torch.ones(4),
high=2 * torch.ones(4))
from lfi.nsf import distributions as distributions_
likelihood = distributions_.LotkaVolterraOscillating()
potential_function = PotentialFunction(likelihood, prior)
# kernel = Slice(potential_function=potential_function)
from pyro.infer.mcmc import HMC, NUTS
# kernel = HMC(potential_fn=potential_function)
kernel = NUTS(potential_fn=potential_function)
num_chains = 3
sampler = MCMC(
kernel=kernel,
num_samples=10000 // num_chains,
warmup_steps=200,
initial_params={"": torch.zeros(num_chains, 4)},
num_chains=num_chains,
)
sampler.run()
samples = next(iter(sampler.get_samples().values()))
utils.plot_hist_marginals(utils.tensor2numpy(samples),
ground_truth=utils.tensor2numpy(loc),
lims=[-6, 3])
# plt.show()
plt.savefig("/home/conor/Dropbox/phd/projects/lfi/out/mcmc.pdf")
plt.close()
def test_():
from scipy import stats
from lfi import utils
from matplotlib import pyplot as plt
mean = np.zeros(2)
cov = np.array([[1, 0.9], [0.9, 1]])
distribution = stats.multivariate_normal(mean=mean, cov=cov)
prior = stats.uniform(loc=-1 * np.ones(2), scale=2 * np.ones(2))
lp_f = lambda y: distribution.logpdf(y) + prior.logpdf(y).sum()
x = np.zeros(2)
sampler = SliceSampler(x=x, lp_f=lp_f)
samples = sampler.gen(1000)
utils.plot_hist_marginals(samples, lims=[-4, 4])
plt.savefig(os.path.join(utils.get_output_root(), "slice-test.pdf"))
def sample_true_posterior():
prior = distributions.Uniform(low=-3 * torch.ones(5),
high=3 * torch.ones(5))
# print(log_prob)
potential_function = (lambda parameters: simulator.log_prob(
observations=true_observation, parameters=parameters) + prior.log_prob(
torch.Tensor(parameters)).sum().item())
sampler = SliceSampler(x=true_parameters, lp_f=potential_function, thin=10)
sampler.gen(200)
samples = sampler.gen(2500)
# figure = corner.corner(
# samples,
# truths=true_parameters,
# truth_color='C1',
# bins=25,
# color='black',
# labels=[r'$ \theta_{1} $', r'$ \theta_{2} $', r'$ \theta_{3} $',
# r'$ \theta_{4} $', r'$ \theta_{5} $'],
# show_titles=True,
# hist_kwargs={'color': 'grey', 'fill': True},
# title_fmt='.2f',
# plot_contours=True,
# quantiles=[0.5]
# )
# plt.tight_layout()
figure = utils.plot_hist_marginals(samples,
ground_truth=true_parameters,
lims=[-4, 4])
np.save(
os.path.join(utils.get_output_root(),
"./true-posterior-samples-gaussian.npy"),
samples,
)
plt.show()
def main():
task = "mg1"
simulator, prior = simulators.get_simulator_and_prior(task)
parameter_dim, observation_dim = (
simulator.parameter_dim,
simulator.observation_dim,
)
true_observation = simulator.get_ground_truth_observation()
neural_likelihood = utils.get_neural_likelihood(
"maf", parameter_dim, observation_dim
)
snl = SNL(
simulator=simulator,
true_observation=true_observation,
prior=prior,
neural_likelihood=neural_likelihood,
mcmc_method="slice-np",
)
num_rounds, num_simulations_per_round = 10, 1000
snl.run_inference(
num_rounds=num_rounds, num_simulations_per_round=num_simulations_per_round
)
samples = snl.sample_posterior(1000)
samples = utils.tensor2numpy(samples)
figure = utils.plot_hist_marginals(
data=samples,
ground_truth=utils.tensor2numpy(
simulator.get_ground_truth_parameters()
).reshape(-1),
lims=simulator.parameter_plotting_limits,
)
figure.savefig("./corner-posterior-snl.pdf")
def test_():
num_simulations = 250
true_parameters = np.log([0.01, 0.5, 1.0, 0.01])
parameters, observations = sim_data(
gen_params=lambda num_simulations, rng: np.tile(
true_parameters, num_simulations
).reshape(-1, 4),
sim_model=lambda parameters, rng: Stats().calc(
Model().sim(parameters, rng=rng)
),
n_samples=num_simulations,
)
print(parameters.shape, observations.shape)
# utils.plot_hist_marginals(parameters, ground_truth=np.log([0.01, 0.5, 1.0, 0.01]))
utils.plot_hist_marginals(
observations, show_xticks=True, ground_truth=get_ground_truth_observation()
)
plt.show()
def test_():
features = 3
context_features = 5
num_mixture_components = 5
model = MixtureOfGaussiansMADE(
features=features,
hidden_features=32,
context_features=context_features,
num_mixture_components=num_mixture_components,
num_blocks=2,
use_residual_blocks=True,
random_mask=False,
activation=F.relu,
dropout_probability=0,
use_batch_norm=False,
epsilon=1e-2,
custom_initialization=True,
)
context = torch.randn(2, context_features)
samples = model.sample(1000, context=context)
utils.plot_hist_marginals(utils.tensor2numpy(samples.squeeze(0)),
lims=[-10, 10])
plt.show()
def test_():
task = "lotka-volterra"
simulator, prior = simulators.get_simulator_and_prior(task)
parameter_dim, observation_dim = (
simulator.parameter_dim,
simulator.observation_dim,
)
true_observation = simulator.get_ground_truth_observation()
classifier = utils.get_classifier("mlp", parameter_dim, observation_dim)
ratio_estimator = SRE(
simulator=simulator,
true_observation=true_observation,
classifier=classifier,
prior=prior,
num_atoms=-1,
mcmc_method="slice-np",
retrain_from_scratch_each_round=False,
)
num_rounds, num_simulations_per_round = 10, 1000
ratio_estimator.run_inference(
num_rounds=num_rounds,
num_simulations_per_round=num_simulations_per_round)
samples = ratio_estimator.sample_posterior(num_samples=2500)
samples = utils.tensor2numpy(samples)
figure = utils.plot_hist_marginals(
data=samples,
ground_truth=utils.tensor2numpy(
simulator.get_ground_truth_parameters()).reshape(-1),
lims=[-4, 4],
)
figure.savefig(
os.path.join(utils.get_output_root(), "corner-posterior-ratio.pdf"))
mmds = ratio_estimator.summary["mmds"]
if mmds:
figure, axes = plt.subplots(1, 1)
axes.plot(
np.arange(0, num_rounds * num_simulations_per_round,
num_simulations_per_round),
np.array(mmds),
"-o",
linewidth=2,
)
figure.savefig(os.path.join(utils.get_output_root(), "mmd-ratio.pdf"))
def _summarize(self, round_):
# Update summaries.
try:
mmd = utils.unbiased_mmd_squared(
self._parameter_bank[-1],
self._simulator.get_ground_truth_posterior_samples(
num_samples=1000),
)
self._summary["mmds"].append(mmd.item())
except:
pass
# Median |x - x0| for most recent round.
median_observation_distance = torch.median(
torch.sqrt(
torch.sum(
(self._summary_net(self._observation_bank[-1]) -
self._summary_net(self._true_observation).reshape(1,
-1))**2,
dim=-1,
)))
self._summary["median-observation-distances"].append(
median_observation_distance.item())
# KDE estimate of negative log prob true parameters using
# parameters from most recent round.
negative_log_prob_true_parameters = -utils.gaussian_kde_log_eval(
samples=self._parameter_bank[-1],
query=self._simulator.get_ground_truth_parameters().reshape(1, -1),
)
self._summary["negative-log-probs-true-parameters"].append(
negative_log_prob_true_parameters.item())
# Rejection sampling acceptance rate
rejection_sampling_acceptance_rate = self._estimate_acceptance_rate()
self._summary["rejection-sampling-acceptance-rates"].append(
rejection_sampling_acceptance_rate)
# Plot most recently sampled parameters.
parameters = utils.tensor2numpy(self._parameter_bank[-1])
figure = utils.plot_hist_marginals(
data=parameters,
ground_truth=utils.tensor2numpy(
self._simulator.get_ground_truth_parameters()).reshape(-1),
lims=self._simulator.parameter_plotting_limits,
)
# Write quantities using SummaryWriter.
self._summary_writer.add_figure(tag="posterior-samples",
figure=figure,
global_step=round_ + 1)
self._summary_writer.add_scalar(
tag="epochs-trained",
scalar_value=self._summary["epochs"][-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="best-validation-log-prob",
scalar_value=self._summary["best-validation-log-probs"][-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="median-observation-distance",
scalar_value=self._summary["median-observation-distances"][-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="negative-log-prob-true-parameters",
scalar_value=self._summary["negative-log-probs-true-parameters"]
[-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="rejection-sampling-acceptance-rate",
scalar_value=self._summary["rejection-sampling-acceptance-rates"]
[-1],
global_step=round_ + 1,
)
if self._summary["mmds"]:
self._summary_writer.add_scalar(
tag="mmd",
scalar_value=self._summary["mmds"][-1],
global_step=round_ + 1,
)
self._summary_writer.flush()
def _summarize(self, round_):
# Update summaries.
try:
mmd = utils.unbiased_mmd_squared(
self._parameter_bank[-1],
self._simulator.get_ground_truth_posterior_samples(
num_samples=1000),
)
self._summary["mmds"].append(mmd.item())
except:
pass
median_observation_distance = torch.median(
torch.sqrt(
torch.sum(
(self._observation_bank[-1] -
self._true_observation.reshape(1, -1))**2,
dim=-1,
)))
self._summary["median-observation-distances"].append(
median_observation_distance.item())
negative_log_prob_true_parameters = -utils.gaussian_kde_log_eval(
samples=self._parameter_bank[-1],
query=self._simulator.get_ground_truth_parameters().reshape(1, -1),
)
self._summary["negative-log-probs-true-parameters"].append(
negative_log_prob_true_parameters.item())
# Plot most recently sampled parameters in TensorBoard.
parameters = utils.tensor2numpy(self._parameter_bank[-1])
figure = utils.plot_hist_marginals(
data=parameters,
ground_truth=utils.tensor2numpy(
self._simulator.get_ground_truth_parameters()).reshape(-1),
lims=self._simulator.parameter_plotting_limits,
)
self._summary_writer.add_figure(tag="posterior-samples",
figure=figure,
global_step=round_ + 1)
self._summary_writer.add_scalar(
tag="epochs-trained",
scalar_value=self._summary["epochs"][-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="best-validation-log-prob",
scalar_value=self._summary["best-validation-log-probs"][-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="median-observation-distance",
scalar_value=self._summary["median-observation-distances"][-1],
global_step=round_ + 1,
)
self._summary_writer.add_scalar(
tag="negative-log-prob-true-parameters",
scalar_value=self._summary["negative-log-probs-true-parameters"]
[-1],
global_step=round_ + 1,
)
if self._summary["mmds"]:
self._summary_writer.add_scalar(
tag="mmd",
scalar_value=self._summary["mmds"][-1],
global_step=round_ + 1,
)
self._summary_writer.flush()
