def fit(self, x, y):
n_samples, p = x.shape
linear_names = ["x_{}".format(i) for i in range(p)]
names = linear_names[:]
data = [x[:, i] for i in range(p)]
models = net(Lasso, x, y, max_coarsity=self.max_coarsity).values()
scores = [model.score(x, y) for model in models]
coefs = [model.coef_ for model in models]
importance = get_importance(coefs, scores)
stall_iter = 0
best_names = linear_names[:]
best_model, best_score = _fit_model(x,
y,
best_names,
self.operators,
n_jobs=self.n_jobs)
pop_size = p * (self.mu + 1 + self.q)
for _ in range(self.max_iter):
old_names = sorted(names[:])
stall_iter += 1
new_names = []
new_data = []
for i in range(3 * pop_size):
f, new_name, parents = mutate(names, importance, self.toursize,
self.operators, self.rng)
if size(
new_name
) <= self.max_size and new_name not in new_names and new_name not in names:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
feature = f(*[data[i] for i in parents])
if np.all(np.isfinite(feature)) and all(
abs(np.corrcoef(feature, data[i]))[1,
0] <= self.t
for i in parents):
new_names.append(new_name)
new_data.append(feature)
if len(new_names + names) < pop_size:
break
else:
warnings.warn(
"Failed to produce a new population given the tree-depth {} and correlation threshold {}."
.format(self.max_size, self.t), ConvergenceWarning)
names.extend(new_names)
data.extend(new_data)
models = net(Lasso,
np.array(data).T,
y,
max_coarsity=self.max_coarsity).values()
scores = [model.score(np.array(data).T, y) for model in models]
coefs = [model.coef_ for model in models]
importance = list(get_importance(coefs, scores))
names_to_discard = [
n for n in sorted(names,
key=lambda x: importance[names.index(x)],
reverse=True) if n not in linear_names
][-self.mu * p:]
for n in names_to_discard:
i = names.index(n)
names.pop(i)
data.pop(i)
importance.pop(i)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
model, score = _fit_model(x,
y,
names,
self.operators,
n_jobs=self.n_jobs)
if score > best_score:
best_model = model
best_score = score
stall_iter = 0
elif stall_iter >= self.max_stall_iter:
break
self.model = best_model
return self
exponents = [1]
operators = {}
sym = sf.SymbolicFeatures(exponents=exponents, operators=operators)
features = sym.fit_transform(x)
ests = [Lasso, STRidge]
attrs = ["alpha", "threshold"]
names = ["Lasso", "STRidge"]
for est, attr, name in zip(ests, attrs, names):
models = net(est,
features,
y,
attr,
filter=True,
max_coarsity=5,
r_max=1e5)
m = sorted(models)
scores = np.array([models[k].score(features, y) for k in m])
plt.plot(m, scores, 'o--', label=name)
plt.legend()
plt.xlabel("# coefficient")
plt.ylabel(r"$R^2$")
plt.gca().invert_xaxis()
plt.show()