plt.title('W4[' + str(coeffi) + ',' + str(coeffj) + ']')
plt.grid(True)
plt.suptitle("1d projections of the appoximated 23d energy landscape;" + \
" x: values, y: probabilities\nPeaks correspond to lower total cost")
plt.show()
# Now, let's start with clustering
# Sometimes we want to skip the elbow search, debug purposes
elbow = True
if (elbow):
# Find the optimal number of clusters
mcmc.elbow_search(X, 3, 70)
ncent = int(input("Enter the number of centroids: "))
# Store the clusters, i.e. the points on which the distrobution concentrates
centroids, freq = mcmc.detailed_clustering(X, ncent, auxiliary_tot_cost)
# Plot the clustered empirical distribution
for i in range(23):
bb = 50 # Bins
plt.subplot(5, 5, i + 1)
# If 0 or 1, we are with parameter b2
if (i < 2):
plt.hist(centroids[:, i], bb, density=True, color='green')
plt.title('b2[' + str(i + 1) + ']')
elif (i < 5):
plt.hist(centroids[:, i], bb, density=True, color='darkorange')
plt.title('b3[' + str(i - 1) + ']')
elif (i < 7):
plt.hist(centroids[:, i], bb, density=True, color='magenta')
plt.title('b4[' + str(i - 4) + ']')
X.append(np.array(x[0:-1].split(' ')).astype("float64"))
X = np.asanyarray(X)
samples_file.close()
print("Read", len(X), "samples of dimension", len(X[0]))
# Plot the sampling distribution
d = len(X[0])
for i in range(d):
plt.subplot(d, 1, i + 1)
plt.hist(X[:, i], 50, density=True)
plt.suptitle("Complete distribution")
plt.show()
# Find the optimal number of clustering
mcmc.elbow_search(X, 1, 10)
ncent = int(input("Enter the number of centroids: "))
# Store the clusters, which will be candidate modes
centroids, freq = mcmc.detailed_clustering(X, ncent, U)
# Perform gradient descent on each centroid to identify the modes
print("\nSearch for the modes: ")
for i in range(ncent):
print("Gradient descent on candidate mode number", i)
if (mcmc.simple_descent(centroids[i], U, gradU)):
print("MODE FOUND: centroid number ", i)
for i in range(d):
plt.subplot(d, 1, i + 1)
plt.scatter(centroids[:, i], freq, marker="*")
plt.suptitle("Clustered distribution")
plt.show()