def main():
### Undirected graph ###
# Initialize model using the Petersen graph
model=gmm.gmm(nx.petersen_graph())
old_graph=model.get_base()
model.set_termination(node_ceiling)
model.set_rule(rand_add)
# Run simualation with tau=4 and Poisson density for motifs
gmm.algorithms.simulate(model,4)
# View results
new_graph=model.get_base()
print(nx.info(new_graph))
# Draw graphs
old_pos=nx.spring_layout(old_graph)
new_pos=nx.spring_layout(new_graph,iterations=2000)
fig1=plt.figure(figsize=(15,7))
fig1.add_subplot(121)
#fig1.text(0.1,0.9,"Base Graph")
nx.draw(old_graph,pos=old_pos,node_size=25,with_labels=False)
fig1.add_subplot(122)
#fig1.text(0.1,0.45,"Simulation Results")
nx.draw(new_graph,pos=new_pos,node_size=20,with_labels=False)
fig1.savefig("undirected_model.png")
### Directed graph ###
# Initialize model using random directed Barabasi-Albert model
directed_base=nx.barabasi_albert_graph(25,2).to_directed()
directed_model=gmm.gmm(directed_base)
directed_model.set_termination(node_ceiling)
directed_model.set_rule(rand_add)
# Run simualation with tau=4 and Poisson density for motifs
gmm.algorithms.simulate(directed_model,4)
# View results
new_directed=directed_model.get_base()
print(nx.info(new_directed))
# Draw directed graphs
old_dir_pos=new_pos=nx.spring_layout(directed_base)
new_dir_pos=new_pos=nx.spring_layout(new_directed,iterations=2000)
fig2=plt.figure(figsize=(7,10))
fig2.add_subplot(211)
fig2.text(0.1,0.9,"Base Directed Graph")
nx.draw(directed_base,pos=old_dir_pos,node_size=25,with_labels=False)
fig2.add_subplot(212)
fig2.text(0.1,0.45, "Simualtion Results")
nx.draw(new_directed,pos=new_dir_pos,node_size=20,with_labels=False)
fig2.savefig("directed_model.png")
# Export files
nx.write_graphml(model.get_base(), "base_model.graphml")
nx.write_graphml(directed_model.get_base(), "directed_model.graphml")
nx.write_graphml(nx.petersen_graph(), "petersen_graph.graphml")
def gmm_test(dt, tf, mux0, P0, YK, Qk, Rk, Nsu=20, flag_informative=True):
global nameBit
# add in this functionality so we can change the propagation function dependent on the nameBit ... may or may not be needed
if not flag_informative:
measure_argument = measurement_uninformative
measure_jacobian = measurement_jac_uninformative
else:
measure_argument = measurement_enkf
measure_jacobian = measurement_jac_gmm
if nameBit == 1:
# create EnKF object
GMM = gmm.gmm(2, Nsu, Qk, Rk, eqom_gmm, jac_gmm, process_influence, measure_argument, measure_jacobian)
elif nameBit == 2:
# create EnKF object
GMM = gmm.gmm(2, Nsu, Qk, Rk, eqom_gmm, jac_gmm, process_influence, measure_argument, measure_jacobian)
elif nameBit == 3:
# create EnKF object
GMM = gmm.gmm(2, Nsu, Qk, Rk, eqom_gmm, jac_gmm, process_influence, measure_argument, measure_jacobian)
nSteps = int(tf / dt) + 1
ts = 0.0
# initialize EnKF
GMM.init_monte(mux0, P0, ts)
xml = np.zeros((nSteps, 2))
pdf = np.zeros((nSteps, GMM.aki.shape[1]))
pdfPts = np.zeros((nSteps, 2, GMM.aki.shape[1]))
alphai = np.zeros((nSteps, GMM.aki.shape[1]))
Pki = np.zeros((nSteps, 2, 2, GMM.aki.shape[1]))
tk = np.arange(0.0, tf, dt)
t1 = time.time()
fig = []
for k in range(0, nSteps):
if k > 0:
# get the new measurement
ym = np.array([YK[k]])
ts = ts + dt
# sync the ENKF, with continuous-time integration
print("Propagate to t = %f" % (ts))
# propagate filter
GMM.propagate_normal(dt)
GMM.update(ym)
# log
alphai[k, :] = GMM.alphai.copy()
xml[k, :] = GMM.get_max_likelihood()
Pki[k, :, :, :] = GMM.Pki.copy()
(pdfPts[k, :, :], pdf[k, :]) = GMM.get_pdf()
if k > 0:
GMM.resample()
t2 = time.time()
print("Elapsed time: %f sec" % (t2 - t1))
return (xml, pdf, pdfPts, alphai, Pki)
def plot_gmm(lib=__library__, alg=__algs__[0], G=1, draw=True, show=True, draw_hist=True, save=False, xlim=None, print_result=True, training=True):
# part = 'training' if training else 'testing'
# if G == 0:
# model = gmm(texts = True, gmm_txt='%s/%s/gmm_%s_%d.npz' % (lib, part, alg, G+1), hist_txt=('%s/%s/hist_%s_%d.npz' % (lib, part, alg, G+1)))
# else:
# model = gmm(texts = True, gmm_txt='%s/%s/gmm_%s_%d.npz' % (lib, part, alg, G), hist_txt=('%s/%s/hist_%s_%d.npz' % (lib, part, alg, G)))
if G == 0:
model = gmm(texts = True, gmm_txt='%s/gmm_%s_%d.npz' % (lib, alg, G+1), hist_txt=('%s/hist_%s_%d.npz' % (lib, alg, G+1)))
else:
model = gmm(texts = True, gmm_txt='%s/gmm_%s_%d.npz' % (lib, alg, G), hist_txt=('%s/hist_%s_%d.npz' % (lib, alg, G)))
x = model.bins
bin_width = x[1] - x[0]
x = np.linspace(model.bins[0], model.bins[-1], num=10000)
y = None
if G != 0:
y = [model.pdf_integral(xi - bin_width/2., xi + bin_width/2.) for xi in x]
abc = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
G_code = abc[G]
if G != 0:
leg = ['Error hist', 'GMM'][::-1]
else:
leg = ['Error hist']
if save:
leg = ['\ss{%s}' % l for l in leg]
# if G == 0: leg = None
# images.save_latex(x, y, leg, '%s_%s_gmm_%s_%s' % (lib, part, alg, G_code), 'Error', 'Relative weight', 0.48, xlim=xlim, model=model, ylim=None)
images.save_latex(x, y, leg, '%s_tmp_gmm_%s_%s' % (lib, alg, G_code), 'Error', 'Relative weight', 0.48, xlim=xlim, model=model, ylim=None)
if draw:
if draw_hist:
pp.figure()
pp.bar(model.bins[:-1], model.count)
pp.hold(True)
if xlim is not None:
pp.xlim(xlim)
# pp.stem(x, y, markerfmt='xr', linefmt='r-')
if G != 0:
pp.plot(x, y, color='r', linewidth=3)
# if draw_hist:
# model.draw_hist(fig=True, limit_hist=False)
# pp.hold()
# pp.title("%s - %d gmm" % (alg, G))
# model.draw_gmm_hist(fig=False)
#
# if draw_hist:
# pp.legend(['Histogram', 'GMM'])
print lib, alg, G
if print_result: model.print_results()
if show:
pp.show()
def plot_gmm(lib, alg=__algs__[0], G=1, draw=True, show=True, draw_hist=True, save=False, xlim=None, print_result=True):
if G == 0:
model = gmm(texts = True, gmm_txt='%s/gmm_%s_%d.npz' % (lib, alg, G+1), hist_txt=('%s/hist_%s_%d.npz' % (lib, alg, G+1)))
else:
model = gmm(texts = True, gmm_txt='%s/gmm_%s_%d.npz' % (lib, alg, G), hist_txt=('%s/hist_%s_%d.npz' % (lib, alg, G)))
x = model.bins
model.count[range(np.where(x == -5)[0][0], np.where(x == 5)[0][0]+1)] = 0
bin_width = x[1] - x[0]
x = np.linspace(model.bins[0], model.bins[-1], num=10000)
y = None
if G != 0:
y = [model.pdf_integral(xi - bin_width/2., xi + bin_width/2.) for xi in x]
abc = ['a', 'b', 'c', 'd', 'e', 'f', 'g']
G_code = abc[G]
if G != 0:
leg = ['Error hist', 'GMM'][::-1]
else:
leg = ['Error hist']
if save:
leg = ['\ss{%s}' % l for l in leg]
# if G == 0: leg = None
images.save_latex(x, y, leg, '%s_gmm_%s_%s_outliers' % (lib, alg, G_code), 'Error', 'Relative weight', 0.48, xlim=xlim, model=model, ylim=None)
if draw:
if draw_hist:
pp.figure()
pp.bar(model.bins[:-1], model.count)
pp.hold(True)
if xlim is not None:
pp.xlim(xlim)
# pp.stem(x, y, markerfmt='xr', linefmt='r-')
if G != 0:
pp.plot(x, y, color='r', linewidth=3)
# if draw_hist:
# model.draw_hist(fig=True, limit_hist=False)
# pp.hold()
# pp.title("%s - %d gmm" % (alg, G))
# model.draw_gmm_hist(fig=False)
#
# if draw_hist:
# pp.legend(['Histogram', 'GMM'])
print lib, alg, G
if print_result: model.print_results()
if show:
pp.show()
def main():
# Load Zachary data, randomly delete nodes, and report
zachary=nx.Graph(nx.read_pajek("karate.net")) # Do not want graph in default MultiGraph format
zachary.name="Original Zachary Data"
print(nx.info(zachary))
zachary_subset=rand_delete(zachary, 15) # Remove half of the structure
zachary_subset.name="Randomly Deleted Zachary Data"
print(nx.info(zachary_subset))
# Create model, and simulate
zachary_model=gmm.gmm(zachary_subset,R=karate_rule,T=node_ceiling_34)
gmm.algorithms.simulate(zachary_model,4,poisson=False,new_name="Simulation from sample") # Use tau=4 because data is so small (it's fun!)
# Report and visualize
print(nx.info(zachary_model.get_base()))
fig=plt.figure(figsize=(30,10))
fig.add_subplot(131)
nx.draw_spring(zachary,with_labels=False,node_size=45,iterations=5000)
plt.text(0.01,-0.1,"Original Karate Club",color="darkblue",size=20)
fig.add_subplot(132)
nx.draw_spring(zachary_subset,with_labels=False,node_size=45,iterations=5000)
plt.text(0.01,-0.1,"Random sample of Karate Club",color="darkblue",size=20)
fig.add_subplot(133)
nx.draw_spring(zachary_model.get_base(),with_labels=False,node_size=45,iterations=5000)
plt.text(0.01,-0.1,"Simulation from random sample",color="darkblue",size=20)
plt.savefig("zachary_simulation.png")
def dotest(k, max_iter):
print('Model Number is ' + str(k))
input("Press Enter to Continue: ")
for i in range(n_classes):
train = trains[i]
d = np.shape(train)[1]
dparams = [None]*d
for j in range(d):
dparams[j] = gmm.gmm(k,[x[j] for x in train], max_iter)
paramses[i] = dparams
for i, params in enumerate(paramses):
print("Class " + str(i))
for j, param in enumerate(params):
print(" Dimension " + str(j))
for l, m in enumerate(param):
print(' Model ' + str(l))
print(m)
input('Params calculated, Press Enter to stat true rate: ')
testlen = len(testA) + len(testB)
rate1 = mylib.statTrueRate(testA, paramses, classifier, 0 ) / testlen
rate2 = mylib.statTrueRate(testB, paramses, classifier, 1 ) / testlen
print(rate1 + rate2)
def doit(dic,priors,classes,K,diag):
err = {'train':list(), 'test':list()}
for k in K:
print '*'*15,'K =',str(k),'*'*15
nums, means, covs, nll = {},{},{},{}
# Build GMM models
for dif in dic['train']:
data = pack(dic['train'][dif])
for i in xrange(6):
_nums,_means,_covs,_nll = gmm.gmm(data, weights=None, K=k, hard=True, diagcov=diag)
if(i != 0):
if(_nll > nll[dif]):
continue
nums[dif],means[dif],covs[dif],nll[dif] = _nums,_means,_covs,_nll
criteria = [snll for dif in dic['train']]
kwparams = [{'nums':nums[dif], 'means':means[dif], 'covs':covs[dif], 'prior':priors[dif]} for dif in dic['train']]
# Evaluate
for x in dic:
labels, labels_est = [], []
for dif in dic[x]:
points = dic[x][dif]
labels += [dif for i in xrange(len(points))]
labels_est += optcriterion(points, classes, criteria, kwparams=kwparams, _max=False);
e = 100.0*sum( np.array(labels) != np.array(labels_est) ) / len(labels)
err[x].append( e )
print 'Confusion marix for' , x , 'data','(K={:},diagcov={:})'.format(k,diag)
utils.confusion(labels, labels_est, True)
print '% Error: ', e,'\n'
if(len(K) > 1):
pl.plot(K,err['train'],'--', label= 'Train'+(' (diagcov=True)' if diag else ''))
pl.plot(K,err['test'], label= 'Test'+(' (diagcov=True)' if diag else ''))
def doit(dic, priors, classes, K, diag):
err = {'train': list(), 'test': list()}
for k in K:
print '*' * 15, 'K =', str(k), '*' * 15
nums, means, covs, nll = {}, {}, {}, {}
# Build GMM models
for dif in dic['train']:
data = pack(dic['train'][dif])
for i in xrange(6):
_nums, _means, _covs, _nll = gmm.gmm(data,
weights=None,
K=k,
hard=True,
diagcov=diag)
if (i != 0):
if (_nll > nll[dif]):
continue
nums[dif], means[dif], covs[dif], nll[
dif] = _nums, _means, _covs, _nll
criteria = [snll for dif in dic['train']]
kwparams = [{
'nums': nums[dif],
'means': means[dif],
'covs': covs[dif],
'prior': priors[dif]
} for dif in dic['train']]
# Evaluate
for x in dic:
labels, labels_est = [], []
for dif in dic[x]:
points = dic[x][dif]
labels += [dif for i in xrange(len(points))]
labels_est += optcriterion(points,
classes,
criteria,
kwparams=kwparams,
_max=False)
e = 100.0 * sum(
np.array(labels) != np.array(labels_est)) / len(labels)
err[x].append(e)
print 'Confusion marix for', x, 'data', '(K={:},diagcov={:})'.format(
k, diag)
utils.confusion(labels, labels_est, True)
print '% Error: ', e, '\n'
if (len(K) > 1):
pl.plot(K,
err['train'],
'--',
label='Train' + (' (diagcov=True)' if diag else ''))
pl.plot(K,
err['test'],
label='Test' + (' (diagcov=True)' if diag else ''))
def train_model():
""" Read all MFCC files at ./MFCC/ and use the coefficients to train
models for each user. The model used is a Gaussian Mixture Model
"""
models = []
names = []
for files in os.listdir(mfcc_path):
print 'Training ' + files.split('.')[0] + "'s model"
mfcc = np.loadtxt(mfcc_path + files)
models.append(gmm.gmm(mfcc))
names.append(files.split('.')[0])
return models, names
def train_model():
""" Read all MFCC files at ./MFCC/ and use the coefficients to train
models for each user. The model used is a Gaussian Mixture Model
"""
models = []
names = []
for files in os.listdir(mfcc_path):
print 'Training ' + files.split('.')[0] + "'s model"
mfcc = np.loadtxt(mfcc_path + files)
models.append(gmm.gmm(mfcc))
names.append(files.split('.')[0])
return models, names
def __init__(self, N, K): self.N = N self.K = K self.gmm = [] for i in range(N): self.gmm.append(G.gmm(K, name='Class'+str(i))) self.con_mat = np.array([]) self.recall = [] self.meanrecall = 0.0 self.precision = [] self.meanprecision = 0.0 self.fmeasure = [] self.meanfmeasure = 0.0 self.accuracy = 0.0
def main():
# Load Zachary data, randomly delete nodes, and report
zachary = nx.Graph(nx.read_pajek(
"karate.net")) # Do not want graph in default MultiGraph format
zachary.name = "Original Zachary Data"
print(nx.info(zachary))
zachary_subset = rand_delete(zachary, 15) # Remove half of the structure
zachary_subset.name = "Randomly Deleted Zachary Data"
print(nx.info(zachary_subset))
# Create model, and simulate
zachary_model = gmm.gmm(zachary_subset, R=karate_rule, T=node_ceiling_34)
gmm.algorithms.simulate(zachary_model,
4,
poisson=False,
new_name="Simulation from sample"
) # Use tau=4 because data is so small (it's fun!)
# Report and visualize
print(nx.info(zachary_model.get_base()))
fig = plt.figure(figsize=(30, 10))
fig.add_subplot(131)
nx.draw_spring(zachary, with_labels=False, node_size=45, iterations=5000)
plt.text(0.01, -0.1, "Original Karate Club", color="darkblue", size=20)
fig.add_subplot(132)
nx.draw_spring(zachary_subset,
with_labels=False,
node_size=45,
iterations=5000)
plt.text(0.01,
-0.1,
"Random sample of Karate Club",
color="darkblue",
size=20)
fig.add_subplot(133)
nx.draw_spring(zachary_model.get_base(),
with_labels=False,
node_size=45,
iterations=5000)
plt.text(0.01,
-0.1,
"Simulation from random sample",
color="darkblue",
size=20)
plt.savefig("zachary_simulation.png")
def binomial_simulation(graph_set, seed=None, verbose=False):
"""
Given a set of binomial random graphs, simulate using above growth rule
"""
simulated_graphs = list()
# Iterate over graph set to produce several simulations
for i in xrange(0, len(graph_set) - 1):
# Required to change the growth rule dynamically,
# so we define it inside the simulation's for-loop
termination_size = graph_set[i + 1].number_of_nodes()
def node_ceiling(G):
if G.number_of_nodes() >= termination_size:
return False
else:
return True
# Setup GMM
erdos_renyi_model = gmm.gmm(graph_set[i])
erdos_renyi_model.set_termination(node_ceiling)
erdos_renyi_model.set_rule(binomial_growth)
sim_name = "GMM SIMULATION-" + erdos_renyi_model.get_base().name
gmm.algorithms.simulate(erdos_renyi_model,
tau=3,
poisson=True,
seed=seed,
new_name=sim_name)
# Run simulation
simualted_erdos_renyi = erdos_renyi_model.get_base()
# Print the results to stdout
if verbose:
print(nx.info(graph_set[i + 1]))
print(nx.info(simualted_erdos_renyi))
print("")
simulated_graphs.append(simualted_erdos_renyi)
# Return a list of simulated graphs
return (simulated_graphs)
def binomial_simulation(graph_set, seed=None, verbose=False):
"""
Given a set of binomial random graphs, simulate using above growth rule
"""
simulated_graphs=list()
# Iterate over graph set to produce several simulations
for i in xrange(0,len(graph_set)-1):
# Required to change the growth rule dynamically,
# so we define it inside the simulation's for-loop
termination_size=graph_set[i+1].number_of_nodes()
def node_ceiling(G):
if G.number_of_nodes()>=termination_size:
return False
else:
return True
# Setup GMM
erdos_renyi_model=gmm.gmm(graph_set[i])
erdos_renyi_model.set_termination(node_ceiling)
erdos_renyi_model.set_rule(binomial_growth)
sim_name="GMM SIMULATION-"+erdos_renyi_model.get_base().name
gmm.algorithms.simulate(erdos_renyi_model,tau=3, poisson=True, seed=seed, new_name=sim_name)
# Run simulation
simualted_erdos_renyi=erdos_renyi_model.get_base()
# Print the results to stdout
if verbose:
print(nx.info(graph_set[i+1]))
print(nx.info(simualted_erdos_renyi))
print("")
simulated_graphs.append(simualted_erdos_renyi)
# Return a list of simulated graphs
return(simulated_graphs)
def setUp(self):
"""Tests that all models from setup pass inspection."""
# Node ceiling termination rule for 100 nodes
def node_ceiling(G):
if G.number_of_nodes() >= 100:
return True
else:
return False
# Simple random growth rule: connects random node from base to random node from new
def rand_add(base, new):
from numpy.random import randint
new = nx.convert_node_labels_to_integers(new, first_label=max(base.nodes()) + 1)
new_base = nx.compose(base, new)
base_connector = randint(base.number_of_nodes())
new_connector = randint(min(new.nodes()), max(new.nodes()))
new_base.add_edge(base_connector, new_connector)
return new_base
# This models should pass without exception
self.full_model = gmm.gmm(self.five_cycle, T=node_ceiling, R=rand_add)
def setUp(self):
"""Tests that all models from setup pass inspection."""
# Node ceiling termination rule for 100 nodes
def node_ceiling(G):
if G.number_of_nodes() >= 100:
return True
else:
return False
# Simple random growth rule: connects random node from base to random node from new
def rand_add(base, new):
from numpy.random import randint
new = nx.convert_node_labels_to_integers(
new, first_label=max(base.nodes()) + 1)
new_base = nx.compose(base, new)
base_connector = randint(base.number_of_nodes())
new_connector = randint(min(new.nodes()), max(new.nodes()))
new_base.add_edge(base_connector, new_connector)
return new_base
# This models should pass without exception
self.full_model = gmm.gmm(self.five_cycle, T=node_ceiling, R=rand_add)
def dotest(k, max_iter):
print('Model Number is ' + str(k))
input("Press Enter to Continue: ")
for i in range(n_classes):
train = trains[i]
d = np.shape(train)[1]
dparams = [None] * d
for j in range(d):
dparams[j] = gmm.gmm(k, [x[j] for x in train], max_iter)
paramses[i] = dparams
for i, params in enumerate(paramses):
print("Class " + str(i))
for j, param in enumerate(params):
print(" Dimension " + str(j))
for l, m in enumerate(param):
print(' Model ' + str(l))
print(m)
input('Params calculated, Press Enter to stat true rate: ')
testlen = len(testA) + len(testB)
rate1 = mylib.statTrueRate(testA, paramses, classifier, 0) / testlen
rate2 = mylib.statTrueRate(testB, paramses, classifier, 1) / testlen
print(rate1 + rate2)
def watts_strogatz_simulation(test_graph, k, p, seed=None, verbose=False):
"""
Given a set of Watts-Strogatz "small-world" random graphs, simulate using above growth rule
"""
# Required to change the growth rule dynamically, so
# we define it inside the simulation function
def watts_strogatz_growth(base, new):
"""
Select k random nodes from new_nodes and connect each node to the nodes in base_nodes
with probability p.
"""
# To keep new nodes from over-writing current ones rename the new nodes starting
# from the last node in base
new = nx.convert_node_labels_to_integers(
new, first_label=max(base.nodes()) + 1)
new_nodes = new.nodes()
base_nodes = base.nodes()
new_base = nx.compose(base, new)
# Shuffle base nodes for random selection
random.shuffle(base_nodes)
# We take only the first k nodes from the shuffled set of new nodes. Then, with probability p,
# connect those nodes to all nodes in base_nodes.
for n in base_nodes[0:k]:
edge_test = zip(random.uniform(size=len(new_nodes)), new_nodes)
for d, m in edge_test:
if (d <= p):
new_base.add_edge(m, n)
# Makes graph fully connceted in WS way
while nx.number_connected_components(new_base) > 1:
mc_nodes = nx.connected_component_subgraphs(new_base)[0].nodes()
cc_nodes = nx.connected_component_subgraphs(new_base)[1].nodes()
new_edges = list()
for i in range(k):
random.shuffle(mc_nodes)
random.shuffle(cc_nodes)
new_edges.append((mc_nodes[0], cc_nodes[0]))
new_base.add_edges_from(new_edges)
new_base.name = ""
return (new_base)
# Node ceiling of 1,00 nodes, ensures fully connceted graph
def node_ceiling(G):
if G.number_of_nodes() >= 100:
return False
else:
return True
# Setup GMM
watts_strogatz_model = gmm.gmm(test_graph)
watts_strogatz_model.set_termination(node_ceiling)
watts_strogatz_model.set_rule(watts_strogatz_growth)
sim_name = "GMM(p[" + str(p) + "])_" + watts_strogatz_model.get_base().name
gmm.algorithms.simulate(watts_strogatz_model,
tau=4,
poisson=True,
seed=seed,
new_name=sim_name)
# Run simulation
simualted_watts_strogatz = watts_strogatz_model.get_base()
# Print the results to stdout
if verbose:
print(nx.info(simualted_watts_strogatz))
print("")
# Return a list of simulated graphs
return (simualted_watts_strogatz)
def main(argin='./',adaptFlag = False):
# output file
# initialize EKF
#Qkin = np.array([[20.0]])#continuous-time integration value
Qkin = np.array([[20.0]])#Euler integration value
Rkin = np.array([ [0.0001] ])
GMM = gmm.gmm(2,25,Qkin,Rkin,stateDerivativeGMM,stateJacobian,stateProcessInfluence,measurementFunction,measurementJacobian)
dt = 0.01
tfin = 10.0
nSteps = int(tfin/dt)
tsim = 0.0
muk = np.array([1.0,0.0])
Pk = np.array([[0.1,0.0],[0.0,1.0]])
xk = np.random.multivariate_normal(muk,Pk)
yk = simMeasurementFunction(xk,tsim)
# initial covariance
GMM.init_monte(xk,Pk)
## true state
xt = np.zeros((nSteps,2))
## discretized PDF value
XK = np.zeros((nSteps,2,GMM.aki.shape[1]))
pk = np.zeros((nSteps,GMM.aki.shape[1]))
alphai = np.zeros((nSteps,GMM.aki.shape[1]))
Pki = np.zeros((nSteps,2,2,GMM.aki.shape[1]))
yt = np.zeros(nSteps)
tplot = np.zeros(nSteps)
t1 = time.time()
for k in range(nSteps):
# log
tplot[k] = tsim
xt[k,:] = xk.copy()
(XK[k,:,:],pk[k,:]) = GMM.get_pdf()
alphai[k,:] = GMM.alphai.copy()
Pki[k,:,:,:] = GMM.Pki.copy()
# propagate filter
GMM.propagate_normal(dt)
# simulate
y = sp.odeint(stateDerivative,xk,np.array([tsim,tsim+dt]),args=([],) )
xk = y[-1,:].copy()
# update time
tsim = tsim + dt
# measurement
ymeas = simMeasurementFunction(xk,tsim)
# store measurement
yt[k] = ymeas[0]
# update EKF
GMM.update(ymeas)
print("%f,%f" % (tsim,ymeas[0]))
# resample?
GMM.resample()
t2 = time.time()
print('Completed simulation in %f seconds' % (t2-t1))
print(XK.shape,tplot.shape)
# len(tplot) x Ns matrix of times
tMesh = np.kron(np.ones((GMM.aki.shape[1],1)),tplot).transpose()
# find the max of the PDF for the maximum liklihood estimate
xml = np.zeros((nSteps,2))
Pkk = np.zeros((nSteps,2))
for k in range(nSteps):
idmax = np.argmax(pk[k,:])
xml[k,:] = XK[k,:,idmax].transpose()
# compute the covariance
mu = np.zeros(2)
for j in range(GMM.aki.shape[1]):
mu = mu + alphai[k,j]*XK[k,:,j]
Pxx = np.zeros((2,2))
for j in range(GMM.aki.shape[1]):
Pxx = Pxx + alphai[k,j]*(Pki[k,:,:,j] + np.outer(XK[k,:,j]-mu,XK[k,:,j]-mu))
#print("%f,%f|%f,%f,%f,%f" % (mu[0],mu[1],Pxx[0,0],Pxx[0,1],Pxx[1,0],Pxx[1,1]))
Pkk[k,0] = Pxx[0,0]
Pkk[k,1] = Pxx[1,1]
fig = plt.figure()
ax = []
for k in range(4):
if k < 2:
nam = 'x' + str(k+1)
else:
nam = 'e' + str(k-1)
ax.append( fig.add_subplot(2,2,k+1,ylabel=nam) )
if k < 2:
if k == 0:
# plot the discrete PDF as a function of time
mex = tMesh.reshape((len(tplot)*GMM.aki.shape[1],))
mey = XK[:,0,:].reshape((len(tplot)*GMM.aki.shape[1],))
mez = pk.reshape((len(tplot)*GMM.aki.shape[1],))
elif k == 1:
# plot the discrete PDF as a function of time
mex = tMesh.reshape((len(tplot)*GMM.aki.shape[1],))
mey = XK[:,1,:].reshape((len(tplot)*GMM.aki.shape[1],))
mez = pk.reshape((len(tplot)*GMM.aki.shape[1],))
idx = mez.argsort()
mexx,meyy,mezz = mex[idx],mey[idx],mez[idx]
cc = ax[k].scatter(mexx,meyy,c=mezz,s=20,edgecolor='')
fig.colorbar(cc,ax=ax[k])
# plot the truth
ax[k].plot(tplot,xt[:,k],'b-')
elif k < 4:
ax[k].plot(tplot,xt[:,k-2]-xml[:,k-2],'b-')
ax[k].plot(tplot,3.0*np.sqrt(Pkk[:,k-2]),'r--')
ax[k].plot(tplot,-3.0*np.sqrt(Pkk[:,k-2]),'r--')
ax[k].grid()
fig.show()
raw_input("Return to exit")
print("Completed test_enky.py")
return
pos = np.zeros((img.shape[0], img.shape[1], 2))
for i in range(pos.shape[0]):
for j in range(pos.shape[1]):
pos[i, j, :] = [i, j]
pos = pos.reshape((-1, 2))
data = np.hstack((pos, np.reshape(img, (-1, 3))))
# k_means
res = k_means(data, 3, iter_times=20, dist_func=dist)
tag = res[:, -1]
tag = np.reshape(tag, c.shape)
plt.figure(2)
plt.imshow(tag)
plt.axis('off')
plt.show()
# gmm
res = gmm(data, 3, iter_times=20)
tag = res[:, -1]
tag = np.reshape(tag, c.shape)
plt.figure(3)
plt.imshow(tag)
plt.axis('off')
plt.show()
# dbscan
res = dbscan(data, 10, 100, dist_func=dist)
tag = res[:, -1]
tag = np.reshape(tag, c.shape)
plt.figure(4)
plt.imshow(tag)
plt.axis('off')
plt.show()
def barabasi_albert_simulation(test_graph, m, seed=None, verbose=False):
"""
Given a set of Barabasi-Albert "preferential attachment" random graphs, simulate using above growth rule
"""
# Required to change the growth rule dynamically, so
# we define it inside the simulation function
def barabasi_albert_growth(base, new):
"""
Select m random nodes from new_nodes and connect each node to the nodes in base_nodes
as a function of the degree of the node in base_nodes. The basic "preferential
attachment" model.
"""
# To keep new nodes from over-writing current ones rename the new nodes starting
# from the last node in base
new = nx.convert_node_labels_to_integers(
new, first_label=max(base.nodes()) + 1)
new_nodes = new.nodes()
base_nodes = base.nodes()
new_base = nx.compose(base, new)
# Shuffle new_nodes
random.shuffle(new_nodes)
# Create edge test based on degree centrality
base_degree = nx.degree_centrality(base).items()
# Create "preferential attachment" structure by connecting m
# nodes from new structure to nodes in base as a function of
# base nodes' degree centrality
for i in xrange(m):
edge_made = False
while edge_made is False:
# Randomly select a node in base and add connection
# based on its degree centrality
p = random.uniform()
j = random.randint(0, len(base_nodes))
if p <= base_degree[j][1]:
k = random.randint(
len(new_nodes)) # Randomly select a new node
new_base.add_edge(new_nodes[k], base_degree[j][0])
edge_made = True
return (new_base)
# Simple node ceiling of 1,000 nodes
def node_ceiling(G):
if G.number_of_nodes() >= 1000:
return False
else:
return True
# Setup GMM
barabasi_albert_model = gmm.gmm(test_graph)
barabasi_albert_model.set_termination(node_ceiling)
barabasi_albert_model.set_rule(barabasi_albert_growth)
sim_name = "GMM SIMULATION-" + barabasi_albert_model.get_base().name
gmm.algorithms.simulate(barabasi_albert_model,
tau=3,
poisson=True,
seed=seed,
new_name=sim_name)
# Run simulation
simualted_barabasi_albert = barabasi_albert_model.get_base()
# Print the results to stdout
if verbose:
print(nx.info(simualted_barabasi_albert))
print("")
# Return a list of simulated graphs
return (simualted_barabasi_albert)
def execute(lib=__library__):
start = datetime.datetime.now()
print "START TIME", str(start)
# fetch shape of images
img = images.fetch_disp("bm", 1, __library__)
sh = (int(np.shape(img)[0]), int(np.shape(img)[1]))
shape = (1080, 1920)
# check if ground is available
img = images.fetch_ground(1, __library__)
ground_avail = img is not None
if ground_avail:
# shape = list(np.array(shape) * 2)
shape = (shape[0], shape[1])
else:
# print shape
shape = (shape[0], shape[1] * 2)
print "starting..."
global __begin__
global __end__
for alg in __algs__:
if __timer__:
tm = timer(__begin__, __end__)
mx = 0
overall_diff = np.array([], dtype=__dtype__)
file_num = 0
file_count = 0
max_count = 10
found_count = 0
full_count = 0
# cv2.imshow('1', images.fetch_disp(alg, 1, __library__))
# cv2.imshow('1b', (images.fetch_ground(1, __library__).astype('float') - images.fetch_disp(alg, 1, __library__)[:,:,0].astype('float')).astype('uint8'))
# cv2.imshow('2', images.fetch_disp(alg, 2, __library__))
# cv2.imshow('2b', (images.fetch_ground(2, __library__).astype('float') - images.fetch_disp(alg, 2, __library__)[:,:,0].astype('float')).astype('uint8'))
cv2.waitKey()
if __save_data_only__:
print 'fetching data'
print __library__
for i in range(__begin__, __end__ + 1):
orig = images.fetch_orig(i, __library__)
ground = images.fetch_ground(i, __library__)
disp = images.fetch_disp(alg, i, __library__)
# cv2.imshow('disp', disp)
# cv2.imshow('diff', (ground.astype('float') - disp[:,:,0].astype('float')).astype('uint8'))
# cv2.waitKey()
if orig is None: print 'orig is None'
if ground is None: print 'ground is None'
if disp is None: print 'disp is None'
if disp is not None and ground_avail:
if len(disp.shape) == 3: disp = disp[:,:,0]
diff = ((ground.astype(__dtype__)) - (disp.astype(__dtype__))).astype(__dtype__)
full_count += np.product(diff.shape)
# exclude not calculated
flat = diff[np.where(disp != 0)]
flat_ground = ground[np.where(disp != 0)]
flat = flat[np.where(flat_ground != 0)]
found_count += len(np.nonzero(flat.flatten())[0])
# exclude incalculable values
# diff[ground > 16*6] = 0
diff = np.array(diff, dtype=__dtype__)
overall_diff = np.concatenate((overall_diff, flat))
if file_count == max_count:
print 'saving', '%s/data/data_%s_%d.npz' % (lib, alg, file_num), i
np.savez('%s/data/data_%s_%d.npz' % (lib, alg, file_num), data=overall_diff)
overall_diff = np.array([], dtype=__dtype__)
file_count = 1
file_num += 1
else:
file_count += 1
if __timer__:
tm.progress(i)
print 'found', found_count, 'of', full_count
if full_count != 0:
print float(found_count) / full_count * 100.0
# np.savez('%s/data/data_%s_%d.npz' % (lib, alg, file_num), data=overall_diff)
if __save_data_only__: return
data_files = list()
total = __end__ - __begin__ + 1
# range(int(math.ceil(total / float(max_count))))
for g in __G__:
print
print 'calculating for G =', g
for alg in __algs__:
print 'Calculating for alg', alg
data_files = ['%s/data/data_%s_%d.npz' % (lib, alg, i) for i in range(0, int(math.floor((__end__ - __begin__) / 10.)))]
print data_files
# data_files = ['%s/data/data_%s_%d.npz' % (lib, alg, i) for i in range(int(__begin__ / 10.), int(math.ceil(__end__ / 10.)))]
#if __library__ == 'tsukuba': # data_files=data_files[:-1]
print 'creating model'
model = gmm(data_files, debug=__dbg__, history=True, timer=__timer__, limit=limit)
model.fit_model(g)
model.save_results(library=__library__, alg=alg, hist=True)
# model.draw_results()
print "END TIME ", str(datetime.datetime.now())
print "TIME PASSED", str(datetime.datetime.now() - start)
print
# pp.show()
print "done"
plt.scatter(points[:, 0], points[:, 1])
plt.xlabel('x')
plt.ylabel('y')
plt.title(
"k-means clusters(with euclidean distance) over data{}".format(k))
print("method: kmeans - data{} - euclidean distance".format(k))
etr = Estimater()
etr.get_internal_index(results, euclidean_distance)
results = np.hstack((results, data[:, -1:]))
etr.get_external_index(results)
plt.show()
# gmm
for k in range(1, 4):
data = np.load("data/{}.npy".format(k))
results = gmm(data[:, :-1], k=3)
plt.subplot(1, 3, k)
for i in range(10):
points = results[results[:, -1] == i]
plt.scatter(points[:, 0], points[:, 1])
plt.xlabel('x')
plt.ylabel('y')
plt.title("gmm clusters over data{}".format(k))
print("method: gmm - data{}".format(k))
etr = Estimater()
etr.get_internal_index(results, euclidean_distance)
results = np.hstack((results, data[:, -1:]))
etr.get_external_index(results)
plt.show()
# dbscan
return self.base
else:
return self.rule(self.base,new)
def am_gmm(self):
"""Simple function to test if object is a gmm"""
return self.am_gmm
if __name__ == '__main__':
# Create most basic GMM object with five node cycle graph as base.
import gmm
G=nx.cycle_graph(5)
model=gmm.gmm(G)
# Using five node cycle graph, create gmm object with node ceiling
# termination rule of 100 nodes
def degree_ceiling(G):
if G.number_of_nodes()>=100:
return True
else:
return False
model=gmm.gmm(G,degree_ceiling)
# Finally, add a simple random growth rule
def rand_add(base, new):
def setUp(self):
# Most basic GMM; no growth or termination rules.
self.base_model = gmm.gmm(self.five_cycle)
self.base_directed = gmm.gmm(
copy.deepcopy(self.five_cycle).to_directed())
lda = dimred.LDA(data, classes, labels, center=True, scale=True)
projData = lda.transform(data)
x = projData[0, :]
y = projData[1, :]
points = splitClasses(data, classes, labels)
gaussians = [density.Gaussian(data=p) for p in points]
means = np.array([g.mean() for g in gaussians]).T
covs = np.array([g.cov() for g in gaussians])
nums = np.array([p.shape[1] for p in points])
weights, nll = gmm.calcresps(data, nums, means, covs, hard=hard)
nums, means, covs, nll = gmm.gmm(data,
weights,
K=3,
hard=hard,
diagcov=False)
numsd, meansd, covsd, nlld = gmm.gmm(data,
weights,
K=3,
hard=hard,
diagcov=True)
print 'Question 2'
print 'NLL for diagcov=False: ', nll
print 'NLL for diagcov=True: ', nlld
print 'The diagcov=False seems to be the better choice.'
print 'Because we have enough data per class we can use a full estimate '
print 'of the covariance matrix without over-fitting. '
print 'These extra parameters allow us to make a better model which is '
def setUp(self):
# Most basic GMM; no growth or termination rules.
self.base_model=gmm.gmm(self.five_cycle)
self.base_directed=gmm.gmm(copy.deepcopy(self.five_cycle).to_directed())
def gmm_demo():
df = load_data()
df = extract_features(df)
features_list = list(df.columns.values)[1:]
gmm(df, features_list)
ib = []
for row in data:
Ib = eval(row[6])
ib.append(Ib)
ib = np.array(ib)
def ibase():
return ib
C = 3
(mu_est, sigma_est, p_est, counter, difference) = gmm(ib,C,1e-5)
print('___ Ibase Parameters ___')
print('------Means--------')
print('mu_1=%1.4f'%mu_est[0])
print('mu_2=%1.4f'%mu_est[1])
print('mu_3=%1.4f'%mu_est[2])
print('------Variance--------')
print('sigma_1=%1.4f'%sigma_est[0])
print('sigma_2=%1.4f'%sigma_est[1])
print('sigma_3=%1.4f'%sigma_est[2])
print('------Weights-------')
print('W_1=%1.4f'%p_est[0])
print('W_2=%1.4f'%p_est[1])
if "xy" in opts.color and opts.method != "watershed":
temp = img.copy()
img = np.zeros((img.shape[0], img.shape[1], img.shape[2] + 2))
img[:, :, 0:temp.shape[2]] = temp
img[:, :, temp.shape[2]] = np.array(range(img.shape[0])).reshape(
img.shape[0], 1)
img[:, :, temp.shape[2] + 1] = np.array(range(img.shape[1])).reshape(
1, img.shape[1])
# execute the requested clustering method
if "watershed" in opts.method:
clustering = watershed(img, opts.k)
if "kmeans" in opts.method:
clustering = kmeans(img, opts.k)
if "gmm" in opts.method:
clustering = gmm(img, opts.k)
if "hierarchical" in opts.method:
clustering = hierarchical(img, opts.k)
# read the truth
truth = sio.loadmat(opts.img_file.replace('jpg', 'mat'))
truth = truth['groundTruth'][0, 4][0][0]['Segmentation']
# plot and save the results with a nice title
title = opts.img_file.split("/")[-1].replace(".jpg", "") + "_k=" + str(
opts.k) + "_" + opts.method
showSaveResults(imgoriginal, clustering, truth, title)
# calculate and report mutual information
print("Mutual information (more is better): " +
str(mutual_info_score(truth.flatten(), clustering.flatten())))
print(
def gmm_demo():
df = load_data()
df = extract_features(df)
features_list = list(df.columns.values)[1:]
gmm(df, features_list)
if __name__ == '__main__':
from pandas import read_csv
import gmm
#Load datasets using pandas interface
data1 = read_csv("BankNoteAuthentication.csv")
data2 = read_csv("WineQuality-WhiteWine.csv")
data1.dropna(axis="columns", how="any", inplace=True)
data1.dropna(axis="columns", how="any", inplace=True)
#Data set 1 clustered
trainee1 = data1[["skewness", "curtosis"]]
model1 = gmm.gmm(clusters=5, iter=25, randSeed=42)
normalized1 = model1.normalizeSet(trainee1)
model1.trainModel(normalized1)
model1.draw(normalized1,
model1.u,
model1.sig,
xAxis=trainee1.columns.values[0],
yAxis=trainee1.columns.values[1])
#Data set 2 clustered
trainee2 = data2[["total sulfur dioxide", "chlorides"]]
model2 = gmm.gmm(clusters=2, iter=50, randSeed=42)
normalized2 = model2.normalizeSet(trainee2)
def __init__(self):
self.gmm = gmm()
def main():
# Ensure all data has been downloaded and processed
#utils.download_trips_dataset()
#for y in utils.YEARS:
# utils.load_trips_dataframe(y)
# process_trips(trips_df)
np.random.seed(1)
rc('font', family='serif')
station_info = utils.load_station_info()
start_time_matrix, station_idx, time_idx, time_at_idx = utils.load_start_time_matrix(
)
stop_time_matrix, _, _ = utils.load_stop_time_matrix()
start_time_matrix = start_time_matrix.astype(np.int16)
stop_time_matrix = stop_time_matrix.astype(np.int16)
inverse_station = {v: k for k, v in station_idx.items()}
flow_matrix = stop_time_matrix - start_time_matrix
print("Total data points (excluding inactive stations): {:,}".format(
np.sum(utils.construct_active_stations_by_bucket(start_time_matrix))))
# Figure 1 Total Volume
plot_total_start_trips(start_time_matrix, time_idx)
# Figure 2 Avg Week
select_stations = [195, 511]
plot_avg_week_for_stations(start_time_matrix, station_idx, station_info,
select_stations,
"Number of trips started at station over week",
"avg_week_start_time.pdf")
plot_avg_week_for_stations(stop_time_matrix, station_idx, station_info,
select_stations,
"Number of trips stopped at station over week",
"avg_week_stop_time.pdf")
plot_avg_week_for_stations(
flow_matrix,
station_idx,
station_info,
select_stations, #360, 195, 146, 432, 161, 497, 517],
"Net change in bikes at station over week",
"avg_week_flow_time.pdf",
"Number of Arriving Trips")
# # Cluster stations
print("Clustering stations")
avg_weekly_flow = utils.get_station_agg_trips_over_week(
flow_matrix, np.mean)
cluster_assignments, means, ppc = gmm.gmm(avg_weekly_flow,
K=3,
posterior_predictive_check=True)
# Figure 4 Cluster Means
# Plot weekly graph for mean of each cluster
plot_cluster_means(means, time_at_idx)
# Figure 6 Station Map
# Plot clustered stations on map
plot_map(cluster_assignments, station_info, station_idx, inverse_station)
# Figure 7 Posterior Predictive Check
# Plot the posterior predictive check
random_indices = randint(0, flow_matrix.shape[0], size=150)
plot_posterior_predictive_check(
flow_matrix[random_indices], ppc[0][random_indices],
"Posterior Predictive Check on average flow", "post_pred_check.pdf")
# Predictions
# Figure 8 Accuracy
plot_average_predictor_error(start_time_matrix, flow_matrix)
plot_seasonal_average_predictor_error(start_time_matrix, flow_matrix)
plot_cluster_predictor_error(start_time_matrix, flow_matrix,
cluster_assignments, means)
# Figure 9 Seasonal Prediction
plot_predicted_total_start_trips(start_time_matrix)
# Figure 10 Sample across three seasons
plot_cluster_predictor_sample_error(start_time_matrix, flow_matrix,
cluster_assignments, means)
# Figures not used in the paper
# Some interesting stations: 3412, 3324, 3285, 3286, 3153, 360, 195, 2023, 3095, 432, 511, 438
plot_avg_week_for_stations(
flow_matrix,
station_idx,
station_info, [360, 195, 497, 146, 161],
"Net change in bikes at station over week (normalized)",
"normalized_avg_week_flow_time.pdf",
normalize=True)
plot_avg_week_for_stations(
flow_matrix,
station_idx,
station_info, [360, 195, 497, 146, 161],
"Net change in bikes at station over week (normalized, rounded)",
"normalized_round_avg_week_flow_time.pdf",
normalize=True,
round=True)
plot_avg_week_for_stations(
flow_matrix,
station_idx,
station_info,
None,
"Net change in bikes at station over week (normalized)",
"normalized_all_avg_week_flow_time.pdf",
normalize=True)
plot_avg_week_for_stations(flow_matrix, station_idx, station_info, None,
"Net change in bikes at station over week ",
"all_avg_week_flow_time.pdf")
plot_avg_week_for_stations(
flow_matrix,
station_idx,
station_info,
None,
"Net change in bikes at station over week (normalized, rounded)",
"normalized_round_all_avg_week_flow_time.pdf",
normalize=True,
round=True)
# Plot clustered stations in 2d
plot_clustered_stations_2d(avg_weekly_flow, cluster_assignments, means,
inverse_station)
# Predictions
plot_predicted_flow_baseline(flow_matrix)
def watts_strogatz_simulation(test_graph, k, p, seed=None, verbose=False):
"""
Given a set of Watts-Strogatz "small-world" random graphs, simulate using above growth rule
"""
# Required to change the growth rule dynamically, so
# we define it inside the simulation function
def watts_strogatz_growth(base, new):
"""
Select k random nodes from new_nodes and connect each node to the nodes in base_nodes
with probability p.
"""
# To keep new nodes from over-writing current ones rename the new nodes starting
# from the last node in base
new=nx.convert_node_labels_to_integers(new,first_label=max(base.nodes())+1)
new_nodes=new.nodes()
base_nodes=base.nodes()
new_base=nx.compose(base,new)
# Shuffle base nodes for random selection
random.shuffle(base_nodes)
# We take only the first k nodes from the shuffled set of new nodes. Then, with probability p,
# connect those nodes to all nodes in base_nodes.
for n in base_nodes[0:k]:
edge_test=zip(random.uniform(size=len(new_nodes)), new_nodes)
for d,m in edge_test:
if (d <= p):
new_base.add_edge(m,n)
# Makes graph fully connceted in WS way
while nx.number_connected_components(new_base)>1:
mc_nodes=nx.connected_component_subgraphs(new_base)[0].nodes()
cc_nodes=nx.connected_component_subgraphs(new_base)[1].nodes()
new_edges=list()
for i in range(k):
random.shuffle(mc_nodes)
random.shuffle(cc_nodes)
new_edges.append((mc_nodes[0], cc_nodes[0]))
new_base.add_edges_from(new_edges)
new_base.name=""
return(new_base)
# Node ceiling of 1,00 nodes, ensures fully connceted graph
def node_ceiling(G):
if G.number_of_nodes()>=100:
return False
else:
return True
# Setup GMM
watts_strogatz_model=gmm.gmm(test_graph)
watts_strogatz_model.set_termination(node_ceiling)
watts_strogatz_model.set_rule(watts_strogatz_growth)
sim_name="GMM(p["+str(p)+"])_"+watts_strogatz_model.get_base().name
gmm.algorithms.simulate(watts_strogatz_model,tau=4, poisson=True, seed=seed, new_name=sim_name)
# Run simulation
simualted_watts_strogatz=watts_strogatz_model.get_base()
# Print the results to stdout
if verbose:
print(nx.info(simualted_watts_strogatz))
print("")
# Return a list of simulated graphs
return(simualted_watts_strogatz)
data, classes, labels = readData(os.path.join('..','data','wine.data'))
lda = dimred.LDA(data, classes, labels, center=True, scale=True)
projData = lda.transform(data)
x = projData[0,:]
y = projData[1,:]
points = splitClasses(data,classes,labels)
gaussians = [density.Gaussian(data=p) for p in points]
means = np.array([g.mean() for g in gaussians]).T
covs = np.array([g.cov() for g in gaussians])
nums = np.array([ p.shape[1] for p in points ])
weights,nll = gmm.calcresps(data, nums, means, covs, hard=hard)
nums, means, covs, nll = gmm.gmm(data, weights, K=3, hard=hard, diagcov=False)
numsd,meansd,covsd,nlld = gmm.gmm(data, weights, K=3, hard=hard, diagcov=True)
print 'Question 2'
print 'NLL for diagcov=False: ', nll
print 'NLL for diagcov=True: ', nlld
print 'The diagcov=False seems to be the better choice.'
print 'Because we have enough data per class we can use a full estimate '
print 'of the covariance matrix without over-fitting. '
print 'These extra parameters allow us to make a better model which is '
print 'evident in the differance in negative-log-likelihood(NNL).\n'
weights,nll = gmm.calcresps(data, nums, means, covs, hard=hard)
pl.figure()
labeled_images = preprocess.load_preprocessed(path + "/Processed", FEATURES_FILE) labeled_images_seg = preprocess.load_preprocessed(seg_path, FEATURES_SEG_FILE) cluster_counts = accuracy_rate.counts_per_cluster(labeled_images) cluster_counts_seg = accuracy_rate.counts_per_cluster(labeled_images_seg) img_features = pca.pca_features(FEATURES_FILE) features_df = pd.DataFrame(img_features) labeled_features = labeled_images.join(features_df) img_features_seg = pca.pca_features(FEATURES_SEG_FILE) features_seg_df = pd.DataFrame(img_features_seg) labeled_features_seg = labeled_images.join(features_seg_df) gmm_labels, gmm_score = gmm.gmm(img_features) gmm_labels_seg, gmm_score_seg = gmm.gmm(img_features_seg) kmeans_labels, kmeans_score = kmeans.kmeans(img_features) kmeans_labels_seg, kmeans_score_seg = kmeans.kmeans(img_features_seg) labeled = accuracy_rate.assigned_images(labeled_images, gmm_labels, kmeans_labels, 0) accuracy_rate.output_files(labeled, 0) voted_labels = accuracy_rate.match_labels(labeled, cluster_counts, 0) labeled_seg = accuracy_rate.assigned_images(labeled_images_seg, gmm_labels_seg, kmeans_labels_seg, SEG_NUM) accuracy_rate.ouput_files(labeled_seg, SEG_NUM) voted_labels_seg = accuracy_rate.match_labels(labeled_seg, cluster_counts_seg, SEG_NUM) gmm_rate, kmeans_rate = get_accuracy_rate(labeled, voted_labels) gmm_rate_seg, kmeans_rate_seg = get_accuracy_rate(labeled_seg, voted_labels_seg)
