def scm_to_linear_scm(scm, ds, n_samples=100):
linear_scm_model = {}
for node, model in scm.assignment.items():
print("Node: {} Items: {}".format(node, model))
parents = model.parents
if parents:
intercept, coefs = get_coefficients(ds, node, parents)
linear_scm_model[node] = linear_model(parents,
coefs,
offset=intercept,
noise_scale=.1)
# print("mu: {}, std: {}".format(mu, std))
print("PARENTS: {}".format(parents))
print("COEFFS: {}".format(coefs))
else:
mu = np.mean(ds[node])
std = np.std(ds[node])
structural_eq = lambda n_samples: np.random.normal(
loc=mu, scale=std, size=n_samples)
linear_scm_model[node] = structural_eq
return StructuralCausalModel(linear_scm_model)
front_door_example = CausalGraphicalModel(nodes=["x", "y", "z"],
edges=[("x", "z"), ("z", "y")],
latent_edges=[("x", "y")])
chain_csm = StructuralCausalModel({
"a": lambda n_samples: np.ones(n_samples),
"b": lambda a, n_samples: a + 2,
"c": lambda b, n_samples: b * 2
})
big_csm = StructuralCausalModel({
"a":
lambda n_samples: np.random.normal(size=n_samples),
"b":
lambda n_samples: np.random.normal(size=n_samples),
"x":
logistic_model(["a", "b"], [-1, 1]),
"c":
linear_model(["x"], [1]),
"y":
linear_model(["c", "e"], [3, -1]),
"d":
linear_model(["b"], [-1]),
"e":
linear_model(["d"], [2]),
"f":
linear_model(["y"], [0.7]),
"h":
linear_model(["y", "a"], [1.3, 2.1])
})
plt.pause(1.)
plt.clf()
plt.show()
scm = StructuralCausalModel({
"goal_x":
lambda n_samples: np.random.normal(loc=2., scale=2.0, size=n_samples),
"goal_y":
lambda n_samples: np.random.normal(loc=2., scale=2.0, size=n_samples),
"init_x":
lambda n_samples: np.random.normal(loc=8., scale=2.0, size=n_samples),
"init_y":
lambda n_samples: np.random.normal(loc=8., scale=2.0, size=n_samples),
"dist_x":
linear_model(["goal_x", "init_x"], [1, -1], noise_scale=.1),
"dist_y":
linear_model(["goal_y", "init_y"], [1, -1], noise_scale=.1),
"final_x":
linear_model(["init_x", "dist_x"], [1, .2], noise_scale=.1),
"final_y":
linear_model(["init_y", "dist_y"], [1, .2], noise_scale=.1),
})
ds = scm.sample(n_samples=1000)
# ds.to_csv(DATA_FILE)
# print(list(scm.assignment.items()))
# print(ds["init_x"].values)
# print(get_coefficients(ds,"dist_x", ["goal_x", "goal_y", "init_x", "init_y"]))
def online_learning_example(it):
# Initialize weights to random value
weight_gt = [2, -3]
weight_final = np.random.randint(-10, 10, size=2)
# weight_final = np.array([0, 2])
ground_truth_scm = StructuralCausalModel({
"init_x":
lambda n_samples: np.random.normal(loc=8., scale=2.0, size=n_samples),
"init_y":
lambda n_samples: np.random.normal(loc=8., scale=2.0, size=n_samples),
"push_x":
lambda n_samples: np.random.normal(loc=0., scale=1.0, size=n_samples),
"push_y":
lambda n_samples: np.random.normal(loc=0., scale=1.0, size=n_samples),
"final_x":
linear_model(["init_x", "push_x"], weight_gt, noise_scale=.1),
"final_y":
linear_model(["init_y", "push_y"], weight_gt, noise_scale=.1),
})
df_gt = ground_truth_scm.sample(n_samples=it)
sgd = SGD(.001, 1, init_weights=weight_final)
for i in range(it):
gt_init_x = [df_gt.init_x[i]]
gt_init_y = [df_gt.init_y[i]]
gt_push_x = [df_gt.push_x[i]]
gt_push_y = [df_gt.push_y[i]]
gt_final_x = [df_gt.final_x[i]]
gt_final_y = [df_gt.final_y[i]]
intervention_vars = {
"init_x": gt_init_x,
"init_y": gt_init_y,
"push_x": gt_push_x,
"push_y": gt_push_y
}
pred_scm = StructuralCausalModel({
"init_x":
lambda n_samples: np.random.normal(
loc=8., scale=2.0, size=n_samples),
"init_y":
lambda n_samples: np.random.normal(
loc=8., scale=2.0, size=n_samples),
"push_x":
lambda n_samples: np.random.normal(
loc=0., scale=3.0, size=n_samples),
"push_y":
lambda n_samples: np.random.normal(
loc=0., scale=3.0, size=n_samples),
"final_x":
linear_model(["init_x", "push_x"], weight_final, noise_scale=.1),
"final_y":
linear_model(["init_y", "push_y"], weight_final, noise_scale=.1),
})
pred_scm_do = do_multiple(list(intervention_vars), pred_scm)
df_pred = pred_scm_do.sample(n_samples=1, set_values=intervention_vars)
pred_final_x = df_pred.final_x
pred_final_y = df_pred.final_y
plt.plot(df_gt.init_x[i], df_gt.init_y[i], 'bo')
text = "True_weights: {}\n Predicted weights {}".format(
weight_gt, weight_final)
plt.text(df_gt.init_x[i] * (1 + 0.01),
df_gt.init_y[i] * (1 + 0.01),
text,
fontsize=12)
# plt.plot(df_gt.final_x, df_gt.final_y, 'go')
plt.quiver(gt_init_x, gt_init_y, gt_final_x, gt_final_y, color="b")
plt.quiver(gt_init_x, gt_init_y, pred_final_x, pred_final_y, color="r")
weight_final, rmse_x = sgd.fit(gt_final_x, pred_final_x,
[gt_init_x, gt_push_x])
# weight_final_y, rmse_y = sgd.fit(gt_final_y, gt_final_y)
plt.pause(1.)
plt.clf()
plt.show()
simple_confounded_potential_outcomes = CausalGraphicalModel(
["x", "y_0", "y_1", "y", "z"], [("z", "x"), ("z", "y_0"), ("z", "y_1"),
("y_0", "y"), ("y_1", "y"), ("x", "y")])
chain_csm = StructuralCausalModel({
"a": lambda n_samples: np.ones(n_samples),
"b": lambda a, n_samples: a + 2,
"c": lambda b, n_samples: b * 2
})
big_csm = StructuralCausalModel({
"a":
lambda n_samples: np.random.normal(size=n_samples),
"b":
lambda n_samples: np.random.normal(size=n_samples),
"x":
logistic_model(("a", "b"), [-1, 1]),
"c":
linear_model(("x", ), [1]),
"y":
linear_model(("c", "e"), [3, -1]),
"d":
linear_model(("b", ), [-1]),
"e":
linear_model(("d", ), [2]),
"f":
linear_model(("y"), [0.7]),
"h":
linear_model(("y", "a"), [1.3, 2.1])
})