def test_neals_funnel_smoke(): dim = 10 guide = AutoIAFNormal(neals_funnel) svi = SVI(neals_funnel, guide, Adam(1e-10), Trace_ELBO()) svi_state = svi.init(random.PRNGKey(0), dim) def body_fn(i, val): svi_state, loss = svi.update(val, dim) return svi_state svi_state = lax.fori_loop(0, 1000, body_fn, svi_state) params = svi.get_params(svi_state) neutra = NeuTraReparam(guide, params) model = neutra.reparam(neals_funnel) nuts = NUTS(model) mcmc = MCMC(nuts, num_warmup=50, num_samples=50) mcmc.run(random.PRNGKey(1), dim) samples = mcmc.get_samples() transformed_samples = neutra.transform_sample(samples['auto_shared_latent']) assert 'x' in transformed_samples assert 'y' in transformed_samples
def main(args): print("Start vanilla HMC...") nuts_kernel = NUTS(dual_moon_model) mcmc = MCMC( nuts_kernel, args.num_warmup, args.num_samples, num_chains=args.num_chains, progress_bar=False if "NUMPYRO_SPHINXBUILD" in os.environ else True) mcmc.run(random.PRNGKey(0)) mcmc.print_summary() vanilla_samples = mcmc.get_samples()['x'].copy() guide = AutoBNAFNormal( dual_moon_model, hidden_factors=[args.hidden_factor, args.hidden_factor]) svi = SVI(dual_moon_model, guide, optim.Adam(0.003), Trace_ELBO()) print("Start training guide...") svi_result = svi.run(random.PRNGKey(1), args.num_iters) print("Finish training guide. Extract samples...") guide_samples = guide.sample_posterior( random.PRNGKey(2), svi_result.params, sample_shape=(args.num_samples, ))['x'].copy() print("\nStart NeuTra HMC...") neutra = NeuTraReparam(guide, svi_result.params) neutra_model = neutra.reparam(dual_moon_model) nuts_kernel = NUTS(neutra_model) mcmc = MCMC( nuts_kernel, args.num_warmup, args.num_samples, num_chains=args.num_chains, progress_bar=False if "NUMPYRO_SPHINXBUILD" in os.environ else True) mcmc.run(random.PRNGKey(3)) mcmc.print_summary() zs = mcmc.get_samples(group_by_chain=True)["auto_shared_latent"] print("Transform samples into unwarped space...") samples = neutra.transform_sample(zs) print_summary(samples) zs = zs.reshape(-1, 2) samples = samples['x'].reshape(-1, 2).copy() # make plots # guide samples (for plotting) guide_base_samples = dist.Normal(jnp.zeros(2), 1.).sample(random.PRNGKey(4), (1000, )) guide_trans_samples = neutra.transform_sample(guide_base_samples)['x'] x1 = jnp.linspace(-3, 3, 100) x2 = jnp.linspace(-3, 3, 100) X1, X2 = jnp.meshgrid(x1, x2) P = jnp.exp(DualMoonDistribution().log_prob(jnp.stack([X1, X2], axis=-1))) fig = plt.figure(figsize=(12, 8), constrained_layout=True) gs = GridSpec(2, 3, figure=fig) ax1 = fig.add_subplot(gs[0, 0]) ax2 = fig.add_subplot(gs[1, 0]) ax3 = fig.add_subplot(gs[0, 1]) ax4 = fig.add_subplot(gs[1, 1]) ax5 = fig.add_subplot(gs[0, 2]) ax6 = fig.add_subplot(gs[1, 2]) ax1.plot(svi_result.losses[1000:]) ax1.set_title('Autoguide training loss\n(after 1000 steps)') ax2.contourf(X1, X2, P, cmap='OrRd') sns.kdeplot(guide_samples[:, 0], guide_samples[:, 1], n_levels=30, ax=ax2) ax2.set(xlim=[-3, 3], ylim=[-3, 3], xlabel='x0', ylabel='x1', title='Posterior using\nAutoBNAFNormal guide') sns.scatterplot(guide_base_samples[:, 0], guide_base_samples[:, 1], ax=ax3, hue=guide_trans_samples[:, 0] < 0.) ax3.set( xlim=[-3, 3], ylim=[-3, 3], xlabel='x0', ylabel='x1', title='AutoBNAFNormal base samples\n(True=left moon; False=right moon)' ) ax4.contourf(X1, X2, P, cmap='OrRd') sns.kdeplot(vanilla_samples[:, 0], vanilla_samples[:, 1], n_levels=30, ax=ax4) ax4.plot(vanilla_samples[-50:, 0], vanilla_samples[-50:, 1], 'bo-', alpha=0.5) ax4.set(xlim=[-3, 3], ylim=[-3, 3], xlabel='x0', ylabel='x1', title='Posterior using\nvanilla HMC sampler') sns.scatterplot(zs[:, 0], zs[:, 1], ax=ax5, hue=samples[:, 0] < 0., s=30, alpha=0.5, edgecolor="none") ax5.set(xlim=[-5, 5], ylim=[-5, 5], xlabel='x0', ylabel='x1', title='Samples from the\nwarped posterior - p(z)') ax6.contourf(X1, X2, P, cmap='OrRd') sns.kdeplot(samples[:, 0], samples[:, 1], n_levels=30, ax=ax6) ax6.plot(samples[-50:, 0], samples[-50:, 1], 'bo-', alpha=0.2) ax6.set(xlim=[-3, 3], ylim=[-3, 3], xlabel='x0', ylabel='x1', title='Posterior using\nNeuTra HMC sampler') plt.savefig("neutra.pdf")