def test_find_cm_on_examples():
for example in examples:
cover = find_cm(example.UDG)
correct_cover = example.MCM
assert cover.shape == correct_cover.shape
for clique in correct_cover:
assert (~np.any(cover - clique, axis=1)).any()
def test_reduction_rule_1_on_3cycle_plus_isolated():
graph = np.zeros((4, 4), dtype=int)
graph[1:4, 1:4] = 1
graph[0, 0] = 1
cover = find_cm(graph)
assert cover.shape == (1, 4)
assert ~np.any(cover - [0, 1, 1, 1], axis=1)
def predict(self, dataset: DatasetInterface):
# make sample data
num_latent, num_observed = triangle.MCM.shape
decoder = MeDILCausalModel(biadj_mat=triangle.MCM)
sampler = gaussian_mixture_sampler(num_latent)
input_sample, output_sample = decoder.sample(sampler, num_samples=1000)
np.save("measurement_data", output_sample)
# step 1: estimate UDG
p_vals, null_corr, dep_graph = hypothesis_test(output_sample.T, num_resamples=100)
# dep_graph is adjacency matrix of the estimated UDG
# step 2: learn graphical MCM
learned_biadj_mat = find_cm(dep_graph)
# step 3: learn functional MCM
num_latent, num_observed = learned_biadj_mat.shape
decoder = MeDILCausalModel(biadj_mat=learned_biadj_mat)
sampler = uniform_sampler(num_latent)
minMCM = GAN("measurement_data.npy", decoder, latent_sampler=sampler, batch_size=100)
trainer = Trainer(max_epochs=100)
trainer.fit(minMCM)
# confirm given and learned causal structures match
vis.show_dag(triangle.MCM)
vis.show_dag(learned_biadj_mat)
# compare plots of disttrance correlation values for given and learned MCMs
generated_sample = decoder.sample(sampler, 1000)[1].detach().numpy()
generated_dcor_mat = distance_correlation(generated_sample.T)
vis.show_obs_dcor_mat(null_corr, print_val=True)
vis.show_obs_dcor_mat(generated_dcor_mat, print_val=True)
# get params for learned functional MCM; replace '0' with 'i' to get params for any M_i
result = decoder.observed["0"].causal_function
# make sample data
num_latent, num_observed = triangle.MCM.shape
decoder = MeDILCausalModel(biadj_mat=triangle.MCM)
sampler = gaussian_mixture_sampler(num_latent)
input_sample, output_sample = decoder.sample(sampler, num_samples=1000)
np.save("measurement_data", output_sample)
# step 1: estimate UDG
p_vals, null_corr, dep_graph = hypothesis_test(output_sample.T,
num_resamples=100)
# dep_graph is adjacency matrix of the estimated UDG
# step 2: learn graphical MCM
learned_biadj_mat = find_cm(dep_graph)
# step 3: learn functional MCM
num_latent, num_observed = learned_biadj_mat.shape
decoder = MeDILCausalModel(biadj_mat=learned_biadj_mat)
sampler = uniform_sampler(num_latent)
minMCM = GAN("measurement_data.npy",
decoder,
latent_sampler=sampler,
batch_size=100)
trainer = Trainer(max_epochs=100)
trainer.fit(minMCM)
# confirm given and learned causal structures match
def test_find_cm_on_3_cycle():
cycle_3 = np.ones((3, 3), dtype=int)
cover = find_cm(cycle_3)
assert cover.shape == (1, 3)
assert ~np.any(cover - [1, 1, 1], axis=1)