np.random.seed(31337)
X_train, y_train, X_test, y_test, scaler = cf.get_normalised_data()
basic_svm = SGDClassifier(loss="hinge", penalty="l2", l1_ratio=0.0, random_state=31337, n_jobs=5)
# From the Scipy docs: to sample a random variable Y such that Y=exp(X) where X~N(mu,sigma), use
# scipy.stats.lognormal(s=sigma, scale=np.exp(mu))
random_search = RandomizedSearchCV(basic_svm,
param_distributions={'alpha': sp_lognormal(s=2, scale=np.exp(-4))},
n_iter=20, verbose=1)
random_search.fit(X_train, y_train)
print("Chosen: ", random_search.best_params_["alpha"])
print("Best CV score: ", random_search.best_score_)
chosen_svm = random_search.best_estimator_
os.makedirs("output/svc", exist_ok=True)
labels = cf.get_label_names()
for i in range(10):
# Don't forget to rescale the hyperplanes to get human-readable versions---the l2 penalty makes
# them close to the origin, so they look indistinguishable.
this_hyperplane = 127*(chosen_svm.coef_[i]/np.max(np.abs(chosen_svm.coef_[i]))) + scaler.mean_
cf.plot_image(this_hyperplane, "output/svc/archetype " + labels[i] + ".png")
preds = chosen_svm.predict(X_test)
print("Test score: %.3f" % chosen_svm.score(X_test, y_test))
alpha_grid = [ 5.0, 1.0, 0.5, 1e-1, 5e-2, 1e-2, 1e-3, 1e-4, 1e-5, 1e-6 ]
cv_results = []
for alpha in alpha_grid:
print(alpha)
mean_score, std_score, svm = evaluate_svm(alpha)
cv_results.append({"alpha": alpha, "mean": mean_score, "std": std_score, "svm": svm})
best_score = np.max([x['mean'] for x in cv_results])
best_score_index = np.where([x['mean'] >= best_score for x in cv_results])[0][0]
onesd_at_best = cv_results[best_score_index]['std']
score_to_beat = best_score - onesd_at_best
for i in range(best_score_index-1, -1, -1):
if cv_results[i]['mean'] > score_to_beat:
best_score_index = i
chosen_alpha = cv_results[best_score_index]["alpha"]
print("Chosen: ", chosen_alpha)
print("Best CV score: ", cv_results[best_score_index]["mean"])
chosen_svm = cv_results[best_score_index]["svm"]
os.makedirs("output/svc", exist_ok=True)
for i in range(10):
this_hyperplane = 127*(chosen_svm.coef_[i]/np.max(np.abs(chosen_svm.coef_[i]))) + scaler.mean_
cf.plot_image(this_hyperplane, "output/svc/archetype " + str(i) + ".png")
preds = chosen_svm.predict(X_test)
print("Test score: %.3f" % chosen_svm.score(X_test, y_test))
