def dendrogramUSPS():
datafile = np.load('usps.npz');
data = datafile['data_patterns'];
n = data.shape[1];
k = 10;
max_iter = 100;
tries = 20;
kmMinError = np.Inf;
kmR = np.zeros(n);
for i in range(tries):
_,r,_,_ = imp.kmeans(data, k, max_iter);
error = imp.kmeans_crit(data, r);
if error < kmMinError:
kmMinError = error;
kmR = r;
R, kmloss, mergeidx = imp.kmeans_agglo(data, kmR);
imp.agglo_dendro(kmloss, mergeidx);
pl.ylabel('Error');
m = np.ceil(np.sqrt(k));
n = np.ceil(float(k)/m);
maxCluster = max(R[0,:])+1;
counter = 0;
pl.figure();
for j in range(maxCluster):
if (R[0,:] == j).sum() > 0:
mu = data[:,R[0,:]==j].sum(axis=1)/(R[0,:]==j).sum();
pl.subplot(m,n,counter);
pl.imshow(mu.reshape(16,16), cmap=cm.Greys);
counter += 1;
pl.axis('off');
pl.suptitle('All centroids');
for i in range(1,k):
leftMu = data[:,R[i-1,:]==mergeidx[i-1,0]].sum(axis=1)/(R[i-1,:]==mergeidx[i-1,0]).sum();
rightMu = data[:,R[i-1,:] == mergeidx[i-1,1]].sum(axis=1)/(R[i-1,:]==mergeidx[i-1,1]).sum();
newMu = (data[:,R[i-1,:]==mergeidx[i-1,0]].sum(axis=1) + data[:,R[i-1,:]==mergeidx[i-1,1]].sum(axis=1))/((R[i-1,:]==mergeidx[i-1,0]).sum()+ (R[i-1,:]==mergeidx[i-1,1]).sum()) ;
pl.figure();
pl.subplot(1,3,1);
pl.imshow(leftMu.reshape(16,16), cmap=cm.Greys);
pl.axis('off');
pl.title('Left cluster');
pl.subplot(1,3,2);
pl.imshow(rightMu.reshape(16,16), cmap=cm.Greys);
pl.axis('off');
pl.title('Right cluster');
pl.subplot(1,3,3);
pl.imshow(newMu.reshape(16,16), cmap = cm.Greys);
pl.axis('off');
pl.title('Merged cluster');
pl.suptitle('Merge:'+str(i));
pl.show();
def test_agglo(self):
worked = False
for _ in range(10):
mu, r, _ = imp.kmeans(self.X, k=3)
r = r.flatten()
R, kmloss, mergeidx = imp.kmeans_agglo(self.X, r)
mergeidx = np.array(mergeidx, dtype=int)
if set([int(r[3]), int(r[6])]) == set(mergeidx[0, :]):
worked = True
imp.agglo_dendro(kmloss, mergeidx)
break
if not worked:
raise AssertionError('test_agglo: the first merge is not correct.')
def test_agglo():
worked = False
for _ in range(10):
mu, r = imp.kmeans(X, k=3)
r = r.flatten()
R, kmloss, mergeidx = imp.kmeans_agglo(X, r)
mergeidx = np.array(mergeidx, dtype=int)
if set([int(r[3]), int(r[6])]) == set(mergeidx[0, :]):
worked = True
imp.agglo_dendro(kmloss, mergeidx)
break
if not worked:
return ['test_agglo: the first merge is not correct.']
else:
return []
def test_agglo():
worked = False
for _ in range(10):
mu, r = imp.kmeans(X, k=3)
r = r.flatten()
R, kmloss, mergeidx = imp.kmeans_agglo(X, r)
mergeidx = np.array(mergeidx, dtype=int)
if set([int(r[3]), int(r[6])]) == set(mergeidx[0, :]):
worked = True
imp.agglo_dendro(kmloss, mergeidx)
break
if not worked:
return ['test_agglo: the first merge is not correct.']
else:
return []
def test_kmeans(self):
worked1 = False
worked2 = False
for _ in range(10):
mu, r, _ = imp.kmeans(self.X, k=3)
if (r[0] == r[1] == r[2] != r[3] and r[3] == r[4] == r[5] != r[6]
and r[6] == r[7] == r[8]):
worked1 = True
# test one cluster center
if (np.linalg.norm(mu[0] - [10.41666, 0.1666]) < 0.1
or np.linalg.norm(mu[1] - [10.41666, 0.1666]) < 0.1
or np.linalg.norm(mu[2] - [10.41666, 0.1666]) < 0.1):
worked2 = True
if worked1 and worked2:
break
if not worked1:
raise AssertionError('test_kmeans cluster assignments are wrong.')
if not worked2:
raise AssertionError(
'test_kmeans did not find the correct cluster center.')
def test_kmeans():
worked1 = False
worked2 = False
for _ in range(10):
mu, r = imp.kmeans(X, k=3)
if (r[0] == r[1] == r[2] <> r[3] and r[3] == r[4] == r[5] <> r[6]
and r[6] == r[7] == r[8]):
worked1 = True
# test one cluster center
if (np.linalg.norm(mu[:, 0] - [10.41666, 0.1666]) < 0.1
or np.linalg.norm(mu[:, 1] - [10.41666, 0.1666]) < 0.1
or np.linalg.norm(mu[:, 2] - [10.41666, 0.1666]) < 0.1):
worked2 = True
if worked1 and worked2:
break
report = []
if not worked1:
report.append('test_kmeans cluster assignments are wrong.')
if not worked2:
report.append('test_kmeans did not find the correct cluster center.')
return report
def test_kmeans():
worked1 = False
worked2 = False
for _ in range(10):
mu, r = imp.kmeans(X, k=3)
if (r[0]==r[1]==r[2]<>r[3] and r[3]==r[4]==r[5]<>r[6] and r[6]==r[7]==r[8]):
worked1 = True
# test one cluster center
if (np.linalg.norm(mu[:,0] - [10.41666, 0.1666]) < 0.1 or
np.linalg.norm(mu[:,1] - [10.41666, 0.1666]) < 0.1 or
np.linalg.norm(mu[:,2] - [10.41666, 0.1666]) < 0.1):
worked2 = True
if worked1 and worked2:
break
report = []
if not worked1:
report.append('test_kmeans cluster assignments are wrong.')
if not worked2:
report.append('test_kmeans did not find the correct cluster center.')
return report
def analyse2Gaussians():
data = np.load('2_gaussians.npy');
num = 2;
d = data.shape[0];
n = data.shape[1];
max_iter =500;
init_kmeans = True;
tries = 40;
minKMError = np.Inf;
maxEMLL = -np.Inf;
minMu = np.zeros((d,num));
minR = np.zeros(n);
totalIterations = 0;
totalTime = 0;
for _ in range(tries):
mu, r, iterations, t = imp.kmeans(data,num,max_iter);
totalIterations += iterations;
totalTime += t;
error = imp.kmeans_crit(data, r);
if minKMError > error:
minKMError = error;
minMu = mu;
minR = r;
print("Min error:" + str(minKMError));
pl.figure();
pl.scatter(data[0,:],data[1,:],23,minR);
pl.scatter(minMu[0,:],minMu[1,:],23,np.arange(num),marker='x');
strTime = '%.2E' % (totalTime/totalIterations);
pl.title('Kmean error:' + str(minKMError)+ " Avg it:" + ('%.2f' % (float(totalIterations)/tries)) +
" Time per it:" + strTime);
maxMu = np.zeros((d,num));
maxSigma = np.zeros((num,d,d));
maxPi = np.zeros(num);
totalIterations = 0;
totalTime = 0;
for _ in range(tries):
pi, mu, sigma, iterations, t = imp.em_mog(data,num,max_iter,init_kmeans);
totalIterations += iterations;
totalTime += t;
ll = imp.log_likelihood(data, pi, mu, sigma);
if maxEMLL < ll:
maxEMLL = ll;
maxMu = mu;
maxSigma = sigma;
maxPi = pi;
print("Max log likelihood:" + str(maxEMLL));
pl.figure();
imp.plot_em_solution(data, maxMu, maxSigma);
pl.title('EM error:' + str(imp.kmean_loss(data, maxPi, maxMu, maxSigma))+ " Avg it:" + ('%.2f'% (float(totalIterations)/tries))
+ " Time per it:" + ('%.2E' % (totalTime/totalIterations)));
pl.show();
def analyse5Gaussians():
data = np.load('5_gaussians.npy');
numClusters = np.arange(9)+2;
max_iter = 500;
init_kmeans = False;
tries = 30;
n = data.shape[1];
d = data.shape[0];
for num in numClusters:
kmMinError = np.Inf;
kmMu = np.zeros((d,num));
kmR = np.zeros(n);
totalIterations = 0;
totalTime = 0;
for _ in range(tries):
mu, r, iterations, t = imp.kmeans(data, num, max_iter);
error = imp.kmeans_crit(data, r);
totalIterations += iterations
totalTime += t
if(error < kmMinError):
kmMinError = error;
kmMu = mu;
kmR = r;
pl.figure();
pl.scatter(data[0,:],data[1,:],23,kmR);
pl.scatter(kmMu[0,:],kmMu[1,:],23,np.arange(num),marker='x');
error = imp.kmeans_crit(data, kmR);
strTime = '%.2E' % (totalTime/totalIterations);
pl.title('Cluster:' + str(num) + " Error:" + ('%.2f' %error) + " Avg it:" + ('%.2f' % (float(totalIterations)/tries)) + " Time per it:" +
strTime);
pl.show(block=False);
emLL = -np.Inf;
emMu = np.zeros((d,num));
emSigma = np.zeros((num,d,d));
emPi = np.zeros(num);
totalIterations = 0;
totalTime = 0;
for _ in range(tries):
pi, mu, sigma,iterations, t = imp.em_mog(data,num,max_iter,init_kmeans);
totalIterations += iterations
totalTime += t;
ll = imp.log_likelihood(data, pi, mu, sigma);
if ll > emLL:
emLL = ll;
emMu = mu;
emSigma = sigma;
emPi = pi;
pl.figure();
imp.plot_em_solution(data, emMu, emSigma);
error = imp.kmean_loss(data,emPi, emMu, emSigma);
strTime = '%.2E' % (totalTime/totalIterations);
pl.title('Cluster:' + str(num) + " Error:" + ( '%.2f'%error)+ " Avg it:" +('%.2f' %( float(totalIterations)/tries)) +
" Time per it:" +strTime);
pl.show(block=False);
pl.figure();
kmMinError = np.Inf;
for _ in range(tries):
_, r, _, _ = imp.kmeans(data, max(numClusters), max_iter);
error = imp.kmeans_crit(data, r);
if(error < kmMinError):
kmMinError = error;
kmR = r;
_,kmloss, mergeidx = imp.kmeans_agglo(data, kmR);
imp.agglo_dendro(kmloss, mergeidx)
pl.title('Dendrogram');
pl.ylabel('Error');
pl.show();
def analyseUSPS():
datafile = np.load('usps.npz');
data = datafile['data_patterns'];
#n = data.shape[1];
d = data.shape[0];
kmMax_iter = 500;
emMax_iter = 100;
k = 10;
tries = 20;
minKMError = np.Inf;
maxEMLL = -np.Inf;
init_kmeans = False;
kmMu = np.zeros((d,k));
#kmR = np.zeros(n);
emMu = np.zeros((d,k));
emSigma = np.zeros((k,d,d));
emPi = np.zeros(k);
totalIterationsKM = 0;
totalTimeKM = 0;
for _ in range(tries):
mu,r, iterations, t = imp.kmeans(data, k, kmMax_iter);
totalIterationsKM += iterations;
totalTimeKM += t;
error = imp.kmeans_crit(data, r);
if error < minKMError:
minKMError = error;
kmMu = mu;
#kmR = r;
totalIterationsEM = 0;
totalTimeEM = 0;
for _ in range(tries):
pi,mu,sigma, iterations, t = imp.log_em_mog(data, k, emMax_iter, init_kmeans);
totalIterationsEM += iterations;
totalTimeEM += t;
ll = imp.log_log_likelihood(data, pi, mu, sigma);
if ll > maxEMLL:
maxEMLL = ll;
emMu = mu;
emPi = pi
emSigma = sigma;
print('KM error:' + str(minKMError) + ' KM avg it:' + ('%.2f'% (float(totalIterationsKM)/tries))
+ ' KM time per it:'+ ('%.2E' % (totalTimeKM/tries))+ ' EM error:' +str(imp.kmean_loss(data,emPi,emMu,emSigma))
+ ' EM avg it:' + ('%.2f'% (float(totalIterationsEM)/tries))
+ ' EM time per it:'+ ('%.2E' % (totalTimeEM/tries)));
pl.figure();
for i in range(10):
pl.subplot(3,4,i);
pl.imshow(kmMu[:,i].reshape(16,16), cmap=cm.Greys);
pl.axis('off');
pl.suptitle('KMeans');
pl.figure();
for i in range(10):
pl.subplot(3,4,i);
pl.imshow(emMu[:,i].reshape(16,16), cmap=cm.Greys);
pl.axis('off');
pl.suptitle('EM');
pl.show();
