new_Y_train = []
for i in range(len(Y_train)):
# if int(Y_train[i]) == 0 or int(Y_train[i]) == 1:
if int(Y_train[i]) in selected_digits:
new_X_train.append(X_train[i])
new_Y_train.append(Y_train[i])
# Use all dataset:
# new_X_train = X_train
# new_Y_train = Y_train
# Used for visualization only
pca = PCA(n_components=2, whiten=False)
reduced_X_train = pca.fit_transform(new_X_train)
unique_labels = get_unique_list(new_Y_train)
n_clusters = len(unique_labels)
print(" >> Labels = ", n_clusters)
print(" >> TOTAL Cluster (n) = ", n_clusters)
# kmeans = KMeans(n_clusters=n_clusters)
kmeans = MyKMeans(n_clusters=n_clusters)
'''
Do dimension Reduction first, and analyze the result
Default: DISABLED; you may enable this.
RESULT: In clustering, PCA does not affect the accuracy!
'''
# pca_x_train = PCA(n_components=64, whiten=False)
# new_X_train = pca_x_train.fit_transform(new_X_train)
kmeans.fit(new_X_train)
# simulate different number of clusters
n_clusters = [2, 3, 4, 5, 6, 7, 8, 9, 10] # from n=2 ~ n=10 (max)
# n_clusters = [2, 3] # from n=2 ~ n=10 (max)
# Start simulation ...
kmeans, reduced_X_train, y_kmeans = None, None, None
for i in range(len(n_clusters)):
selected_digits = get_selected_digits(n_clusters[i])
X_train, Y_train = filter_dataset(selected_digits, init_X_train,
init_Y_train)
# Used for visualization only
pca = PCA(n_components=2, whiten=False)
reduced_X_train = pca.fit_transform(X_train)
unique_labels = get_unique_list(Y_train)
kmeans = MyKMeans(
n_clusters=len(unique_labels)
) # n_clusters = total number of unique digits (labels)
# Start KMeans: Sklearn
highest_acc = 0.0
for j in range(K):
kmeans.fit(X_train)
y_kmeans = kmeans.predict(X_train)
accuracy = kmeans.eval_acc(y_kmeans, Y_train) * 100
# acc_scores.append(accuracy)
highest_acc = accuracy if accuracy > highest_acc else highest_acc
str_hacc = str(round(highest_acc, 2))
print(" >>> highest_acc of n_clusters[%s] = %s " %