def run_experiment(index, dataset_name, name, constraint_getter, master_tree, X, y, out_dir, n_iters=1000, add_constraint=200, add_score=200, add_likelihood=200, should_continue=False): N, D = X.shape df = Inverse(c=1) if dataset_name == 'iris': lm = GaussianLikelihoodModel(sigma=np.eye(D) / 9.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() elif dataset_name == 'zoo': lm = GaussianLikelihoodModel(sigma=np.diag(np.diag(np.cov(X.T))) / 4.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() else: lm = GaussianLikelihoodModel(sigma=np.diag(np.diag(np.cov(X.T))) / 2.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() if should_continue: with open(out_dir / name / 'scores-%u.pkl' % index, 'r') as fp: scores = pickle.load(fp) with open(out_dir / name / 'costs-%u.pkl' % index, 'r') as fp: costs = pickle.load(fp) with open(out_dir / name / 'final-tree-%u.pkl' % index, 'r') as fp: tree = DirichletDiffusionTree(df=df, likelihood_model=lm) tree.set_state(pickle.load(fp)) sampler = MetropolisHastingsSampler(tree, X) else: scores = [] costs = [] tree = DirichletDiffusionTree(df=df, likelihood_model=lm) sampler = MetropolisHastingsSampler(tree, X) sampler.initialize_assignments() if dataset_name == 'zoo': sampler.tree = sampler.tree.induced_subtree(master_tree.points()) current_run = [] for i in tqdm(xrange(n_iters + 1)): sampler.sample() current_run.append(sampler.tree) if i % add_score == 0: scores.append(dist(master_tree, sampler.tree)) if i % add_likelihood == 0: costs.append(sampler.tree.marg_log_likelihood()) if i != 0 and i % add_constraint == 0: if constraint_getter is not None: constraint = constraint_getter.get_constraint(current_run) if constraint is not None: sampler.add_constraint(constraint) current_run = [] # plot_tree(sampler.tree, y) (out_dir / name).mkdir_p() with open(out_dir / name / 'scores-%u.pkl' % index, 'w') as fp: pickle.dump(scores, fp) print len(costs) with open(out_dir / name / 'costs-%u.pkl' % index, 'w') as fp: pickle.dump(costs, fp) # with open(out_dir / name / 'trees-%u.pkl' % index, 'r') as fp: # previous_trees = pickle.load(fp) # with open(out_dir / name / 'trees-%u.pkl' % index, 'w') as fp: # pickle.dump(previous_trees + [t.get_state() for t in trees], fp) with open(out_dir / name / 'final-tree-%u.pkl' % index, 'w') as fp: pickle.dump(sampler.tree.get_state(), fp) return costs, scores, sampler
lm = GaussianLikelihoodModel(sigma=np.cov(X.T) / 2.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() model = DirichletDiffusionTree(df=df, likelihood_model=lm, constraints=train_constraints) sampler = MetropolisHastingsSampler(model, X) sampler.initialize_assignments() n_iters = 2000000 scores = [] for i in tqdm(xrange(n_iters)): sampler.sample() if i % 100 == 0: wat = sampler.tree.score_constraints(train_constraints) assert wat == len(train_constraints), (wat, len(train_constraints)) scores.append( float(sampler.tree.score_constraints(test_constraints)) / len(test_constraints)) fontsize = 18 plt.figure() plt.xlim([0, n_iters / 100]) plt.ylim([0, 1]) plt.xlabel("Iterations", fontsize=fontsize) plt.ylabel("Constraint Score", fontsize=fontsize) plt.plot(scores) plt.legend(loc='best', fontsize=12)
from trees.util import plot_tree, plot_tree_2d from trees.ddt import DirichletDiffusionTree, Inverse, GaussianLikelihoodModel from trees.mcmc import MetropolisHastingsSampler from tqdm import tqdm if __name__ == "__main__": D = 2 N = 100 X = np.random.multivariate_normal(mean=np.zeros(D), cov=np.eye(D), size=N).astype(np.float32) df = Inverse(c=1) lm = GaussianLikelihoodModel(sigma=np.eye(D) / 4.0, mu0=np.zeros(D), sigma0=np.eye(D)) ddt = DirichletDiffusionTree(df=df, likelihood_model=lm) mh = MetropolisHastingsSampler(ddt, X) mh.initialize_assignments() for _ in tqdm(xrange(1000)): mh.sample() plt.figure() plt.plot(mh.likelihoods) plt.figure() plot_tree(mh.tree) plt.figure() plot_tree_2d(mh.tree, X) plt.show()
test_constraints = test_constraints[:10000] df = Inverse(c=1) lm = GaussianLikelihoodModel(sigma=np.cov(X.T) / 2.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() model = DirichletDiffusionTree(df=df, likelihood_model=lm, constraints=train_constraints) sampler = MetropolisHastingsSampler(model, X) sampler.initialize_assignments() n_iters = 2000000 scores = [] for i in tqdm(xrange(n_iters)): sampler.sample() if i % 100 == 0: wat = sampler.tree.score_constraints(train_constraints) assert wat == len(train_constraints), (wat, len(train_constraints)) scores.append(float(sampler.tree.score_constraints(test_constraints)) / len(test_constraints)) fontsize = 18 plt.figure() plt.xlim([0, n_iters/ 100]) plt.ylim([0, 1]) plt.xlabel("Iterations", fontsize=fontsize) plt.ylabel("Constraint Score", fontsize=fontsize) plt.plot(scores) plt.legend(loc='best', fontsize=12) plt.savefig("wat.png")
from trees.ddt import DirichletDiffusionTree, Inverse, GaussianLikelihoodModel from trees.mcmc import MetropolisHastingsSampler from tqdm import tqdm if __name__ == "__main__": D = 2 N = 100 X = np.random.multivariate_normal(mean=np.zeros(D), cov=np.eye(D), size=N).astype(np.float32) df = Inverse(c=1) lm = GaussianLikelihoodModel(sigma=np.eye(D) / 4.0, mu0=np.zeros(D), sigma0=np.eye(D)) ddt = DirichletDiffusionTree(df=df, likelihood_model=lm) mh = MetropolisHastingsSampler(ddt, X) mh.initialize_assignments() for _ in tqdm(range(1000)): mh.sample() plt.figure() plt.plot(mh.likelihoods) #plt.figure() #plot_tree(mh.tree) #plt.figure() #plot_tree_2d(mh.tree, X) plt.show()
def run_experiment(index, dataset_name, name, constraint_getter, master_tree, X, y, out_dir, n_iters=1000, add_constraint=200, add_score=200, add_likelihood=200, should_continue=False): N, D = X.shape df = Inverse(c=1) if dataset_name == 'iris': lm = GaussianLikelihoodModel(sigma=np.eye(D) / 9.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() elif dataset_name == 'zoo': lm = GaussianLikelihoodModel(sigma=np.diag(np.diag(np.cov(X.T))) / 4.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() else: lm = GaussianLikelihoodModel(sigma=np.diag(np.diag(np.cov(X.T))) / 2.0, sigma0=np.eye(D) / 2.0, mu0=X.mean(axis=0)).compile() if should_continue: with open(out_dir / name / 'scores-%u.pkl' % index, 'r') as fp: scores = pickle.load(fp) with open(out_dir / name / 'costs-%u.pkl' % index, 'r') as fp: costs = pickle.load(fp) with open(out_dir / name / 'final-tree-%u.pkl' % index, 'r') as fp: tree = DirichletDiffusionTree(df=df, likelihood_model=lm) tree.set_state(pickle.load(fp)) sampler = MetropolisHastingsSampler(tree, X) else: scores = [] costs = [] tree = DirichletDiffusionTree(df=df, likelihood_model=lm) sampler = MetropolisHastingsSampler(tree, X) sampler.initialize_assignments() if dataset_name == 'zoo': sampler.tree = sampler.tree.induced_subtree(master_tree.points()) current_run = [] for i in tqdm(xrange(n_iters + 1)): sampler.sample() current_run.append(sampler.tree) if i % add_score == 0: scores.append(dist(master_tree, sampler.tree)) if i % add_likelihood == 0: costs.append(sampler.tree.marg_log_likelihood()) if i != 0 and i % add_constraint == 0: if constraint_getter is not None: constraint = constraint_getter.get_constraint(current_run) if constraint is not None: sampler.add_constraint(constraint) current_run = [] # plot_tree(sampler.tree, y) (out_dir / name).mkdir_p() with open(out_dir / name / 'scores-%u.pkl' % index, 'w') as fp: pickle.dump(scores, fp) print len(costs) with open(out_dir / name / 'costs-%u.pkl' % index, 'w') as fp: pickle.dump(costs, fp) # with open(out_dir / name / 'trees-%u.pkl' % index, 'r') as fp: # previous_trees = pickle.load(fp) # with open(out_dir / name / 'trees-%u.pkl' % index, 'w') as fp: # pickle.dump(previous_trees + [t.get_state() for t in trees], fp) with open(out_dir / name / 'final-tree-%u.pkl' % index, 'w') as fp: pickle.dump(sampler.tree.get_state(), fp) return costs, scores, sampler