def preview_stage(self, X, centroids, labels):
n = X.shape[-1]
if n <= 4:
plt.figure(figsize=(10, 6))
for i, (x, y) in enumerate(itertools.combinations(range(n), 2)):
plt.subplot(2, 3, i + 1)
for k in range(self.K):
color = plt.cm.get_cmap('nipy_spectral')(k * 255 / self.K)
plt.scatter(
X[labels == k, x],
X[labels == k, y],
marker='x',
c=color,
)
plt.scatter(
centroids[:, x],
centroids[:, y],
marker='*',
c='yellow',
s=200,
)
plt.autoscale()
plt.show()
else:
draw_filters.draw_filters(centroids, filter_shape=(8, 8))
y, score = predict(x_train_noise, y_train)
print "[train]", score.data
scores_train.append(float(score.data))
y, score = predict(x_valid_noise, y_valid)
print "[valid]", score.data
scores_valid.append(float(score.data))
# show error rate of train and valid
plt.figure()
plt.plot(np.arange(len(scores_train)), np.array(scores_train))
plt.plot(np.arange(len(scores_valid)), np.array(scores_valid))
plt.legend(['train', 'valid'])
plt.show()
draw_filters.draw_filters(model.l1.W)
plt.draw()
except KeyboardInterrupt:
pass
# show error rate of train and valid
plt.figure()
plt.plot(np.arange(len(scores_train)), np.array(scores_train))
plt.plot(np.arange(len(scores_valid)), np.array(scores_valid))
plt.legend(['train', 'valid'])
plt.show()
x_test = generate_noisy_data(x_test, stddev=stddev)
result = np.empty(D)
y, score = predict(x_test, y_test)
