def fit(X, k, do_plot=False):
N, D = X.shape
y = np.ones(N)
medians = np.zeros((k, D))
# random initialization
for kk in range(k):
i = np.random.randint(N)
medians[kk] = X[i]
dist = np.zeros((N, k))
while True:
y_old = y
# Compute L1 distance to each median
for n in range(N):
current_obj = X[n, ]
for i in range(k):
current_median = medians[i, ]
distance = np.abs(current_obj[0] - current_median[0]) + np.abs(
current_obj[1] - current_median[1])
# print distance
dist[n, i] = distance
dist[np.isnan(dist)] = np.inf
y = np.argmin(dist, axis=1)
# Update medians
for kk in range(k):
# medians[kk] = X[y==kk].median(axis=0)
cluster = X[y == kk]
median_x = np.median(cluster, axis=0)
# print median_x
medians[kk] = median_x
changes = np.sum(y != y_old)
print('Running K-medians, changes in cluster assignment = {}'.format(
changes))
# Stop if no point changed cluster
if changes == 0:
break
model = dict()
model['medians'] = medians
model['predict'] = predict
model['error'] = error
if do_plot and D == 2:
utils.plot_2dclustering(X, y)
print("Displaying figure...")
plt.show()
return model
def closure_1_3_1():
k = 4
best_model = None
min_error = np.inf
for i in range(50):
model = Kmeans(k)
model.fit(X)
error = model.error(X)
if error < min_error:
min_error = error
best_model = model
plt.figure()
utils.plot_2dclustering(X, best_model.predict(X))
fname = os.path.join("..", "figs",
"kmeans_outliers_best_model.png")
plt.savefig(fname)
print("\nFigure saved as '%s'" % fname)
def fit(X, k, do_plot=False):
N, D = X.shape
y = np.ones(N)
medians = np.zeros((k, D))
for kk in range(k):
i = np.random.randint(N)
medians[kk] = X[i]
while True:
y_old = y
# Compute distance to each median
for n in range(N):
dist1 = np.absolute(np.sum(X[n, :]) - np.sum(medians, axis=1))
y[n] = np.argmin(dist1)
medians = np.zeros((k, D))
# Update medians
for kk in range(k):
medians[kk] = np.median(X[y == kk], axis=0)
changes = np.sum(y != y_old)
print('Running K-medians, changes in cluster assignment = {}'.format(
changes))
# Stop if no point changed cluster
if changes == 0:
break
if do_plot and D == 2:
utils.plot_2dclustering(X, y)
print("Displaying figure...")
plt.show()
model = dict()
model['medians'] = medians
model['predict'] = predict
model['error'] = error
return model
def fit(X, k, do_plot=False):
N, D = X.shape
y = np.ones(N)
means = np.zeros((k, D))
for kk in range(k):
i = np.random.randint(N)
means[kk] = X[i]
while True:
y_old = y
# Compute euclidean distance to each mean
dist2 = utils.euclidean_dist_squared(X, means)
dist2[np.isnan(dist2)] = np.inf
y = np.argmin(dist2, axis=1)
means = np.zeros((k, D))
# Update means
for kk in range(k):
means[kk] = X[y == kk].mean(axis=0)
changes = np.sum(y != y_old)
print('Running K-means, changes in cluster assignment = {}'.format(changes))
# Stop if no point changed cluster
if changes == 0:
break
if do_plot and D == 2:
utils.plot_2dclustering(X, y)
print("Displaying figure...")
plt.show()
model = dict()
model['means'] = means
model['predict'] = predict
model['error'] = error
return model
'--question',
required=True,
choices=[
'1', '1.1', '1.2', '1.3', '1.4', '2', '2.2', '4',
'4.1', '4.3'
])
io_args = parser.parse_args()
question = io_args.question
if question == '1':
X = utils.load_dataset('clusterData')['X']
model = Kmeans(k=4)
model.fit(X)
utils.plot_2dclustering(X, model.predict(X))
fname = os.path.join("..", "figs", "kmeans_basic.png")
plt.savefig(fname)
print("\nFigure saved as '%s'" % fname)
if question == '1.1':
X = utils.load_dataset('clusterData')['X']
# part 1: implement kmeans.error
# part 2: get clustering with lowest error out of 50 random initialization
best_model = None
min_error = np.inf
for i in range(50):
model = Kmeans(k=4)
print("Testing error =", te_err)
if question == '3.1':
X = utils.load_dataset('clusterData')['X']
model = kmeans.fit(X, k=4)
low = model['error'](model, X)
for i in range(49):
new_model = kmeans.fit(X, k=4)
err = new_model['error'](new_model, X)
if err < low:
model = new_model
low = err
utils.plot_2dclustering(X, model['predict'](model, X))
print("Displaying figure...")
plt.title("K-Means on clusterData")
plt.show()
# part 1: implement kmeans.error
# part 2: get clustering with lowest error out of 50 random initialization
if question == '3.2':
X = utils.load_dataset('clusterData')['X']
# part 3: plot min error across 50 random inits, as k is varied from 1 to 10
low = np.zeros(10)
for k in range(1, 11):