except:
pass
os.mkdir(out_dir)
# Output parameters...
low = -5.0
high = 9.0
width = 400
height = 200
scale = 1.2 * max(
map(lambda i: gt_weight[i] * gt[i].prob(gt[i].getMean()), xrange(len(gt))))
# Iterate, slowlly building up the number of samples used and outputting the fit for each...
out = [8, 16, 32, 64, 128, 256, 512, 1024, 2048]
model = DPGMM(dims, 8)
for i, point in enumerate(samples):
model.add(point)
if (i + 1) in out:
print '%i datapoints:' % (i + 1)
# First fit the model...
model.setPrior()
p = ProgBar()
it = model.solve()
del p
print 'Updated fitting in %i iterations' % it
# Some information...
#print 'a:'
#print model.alpha
x = numpy.linspace(x_low, x_high, 1024)
y_true = numpy.zeros(1024)
for i in range(len(gt)):
y_true += gt_weight[i] * gt[i].prob(x[:, None])
plt.figure(figsize=(16, 8))
plt.xlim(x_low, x_high)
plt.ylim(0.0, y_high)
plt.plot(x, y_true, c='g')
plt.savefig(os.path.join(out_dir, '0000.png'), bbox_inches='tight')
# Iterate, slowlly building up the number of samples used and outputting the fit for each...
out = [8, 16, 32, 64, 128, 256, 512, 1024, 2048]
model = DPGMM(dims, 8)
model.setConcGamma(1 / 8., 1 / 8.)
for i, point in enumerate(samples):
model.add(point)
if (i + 1) in out:
print '%i datapoints:' % (i + 1)
# First fit the model...
model.setPrior()
p = ProgBar()
it = model.solve()
del p
print 'Updated fitting in %i iterations' % it
except: pass os.mkdir(out_dir) # Output parameters... low = -2.0 high = 14.0 width = 800 height = 400 scale = 1.5 * max(map(lambda i: gt_weight[i]*gt[i].prob(gt[i].getMean()), xrange(len(gt)))) # Fill in the model... model = DPGMM(dims) for point in samples: model.add(point) model.setPrior() # Iterate over the number of sticks, increasing till it stops getting better... prev = None while True: print 'Stick count = %i'%model.getStickCap() p = ProgBar() it = model.solve() del p print 'Updated fitting in %i iterations'%it # Now plot the estimated distribution against the actual distribution... img = numpy.ones((height,width,3))
which = numpy.random.multinomial(1, mix).argmax()
covar = sd[which] * numpy.identity(3)
s = numpy.random.multivariate_normal(mean[which, :], covar)
train.append(s)
test = []
for _ in xrange(testCount):
which = numpy.random.multinomial(1, mix).argmax()
covar = sd[which] * numpy.identity(3)
s = numpy.random.multivariate_normal(mean[which, :], covar)
test.append((s, which))
# Train a model...
print "Trainning model..."
model = DPGMM(3)
for feat in train:
model.add(feat)
model.setPrior() # This sets the models prior using the data that has already been added.
model.setConcGamma(1.0, 0.25) # Make the model a little less conservative about creating new categories..
p = ProgBar()
iters = model.solveGrow()
del p
print "Solved model with %i iterations" % iters
# Classify the test set...
probs = model.stickProb(numpy.array(map(lambda t: t[0], test)))
catGuess = probs.argmax(axis=1)
# Output parameters...
low = -2.0
high = 14.0
width = 800
height = 400
scale = 1.5 * max(map(lambda i: gt_weight[i]*gt[i].prob(gt[i].getMean()), xrange(len(gt))))
# Iterate, slowlly building up the number of samples used and outputting the fit for each...
out = [8,16,32,64,128,256,512,1024,2048]
model = DPGMM(dims, 8)
for i,point in enumerate(samples):
model.add(point)
if (i+1) in out:
print '%i datapoints:'%(i+1)
# First fit the model...
model.setPrior()
p = ProgBar()
it = model.solve()
del p
print 'Updated fitting in %i iterations'%it
# Now plot the estimated distribution against the actual distribution...
img = numpy.ones((height,width,3))
draw = model.sampleMixture()
for _ in xrange(trainCount):
which = numpy.random.multinomial(1, mix).argmax()
covar = sd[which] * numpy.identity(3)
s = numpy.random.multivariate_normal(mean[which, :], covar)
train.append(s)
test = []
for _ in xrange(testCount):
which = numpy.random.multinomial(1, mix).argmax()
covar = sd[which] * numpy.identity(3)
s = numpy.random.multivariate_normal(mean[which, :], covar)
test.append((s, which))
# Train a model...
print 'Trainning model...'
model = DPGMM(3)
for feat in train:
model.add(feat)
model.setPrior(
) # This sets the models prior using the data that has already been added.
model.setConcGamma(
1.0, 0.25
) # Make the model a little less conservative about creating new categories..
p = ProgBar()
iters = model.solveGrow()
del p
print 'Solved model with %i iterations' % iters
# Classify the test set...
plt.figure(figsize=(16,8))
plt.xlim(x_low, x_high)
plt.ylim(0.0, y_high)
plt.plot(x, y_true, c='g')
plt.savefig(os.path.join(out_dir, '0000.png'), bbox_inches='tight')
# Iterate, slowlly building up the number of samples used and outputting the fit for each...
out = [8,16,32,64,128,256,512,1024,2048]
model = DPGMM(dims, 8)
model.setConcGamma(1/8., 1/8.)
for i,point in enumerate(samples):
model.add(point)
if (i+1) in out:
print '%i datapoints:'%(i+1)
# First fit the model...
model.setPrior()
p = ProgBar()
it = model.solve()
del p
print 'Updated fitting in %i iterations'%it
dims=2
numpy.random.seed(1);
gt = Gaussian(dims)
gt.setMean([1.0,0.0])
gt.setCovariance([[1.0,0.8],[0.8,1.0]])
sample_count = 30000
sample=[]
for _ in xrange(sample_count):
sample.append(gt.sample())
f=open('data.txt','w')
for x in sample:
f.write('%lf,%lf\n'%(x[0],x[1]))
f.close()
model = DPGMM(dims, 1)
for i,data in enumerate(sample):
model.add(data)
start = time.time()
model.setPrior()
elapsed_time = time.time() - start
num=model.solve()
print elapsed_time
print num
print "%f"%(model.prob([2.0,1.0]))
#for i in range(10):
# x=i*0.4-2.0
#print "%f,%f"%(x,model.prob([x]))
cov = matrix([[1,1],[1,2]])
x1,y1 = random.multivariate_normal(mu1,cov,500).T
samples = []
for i in range(len(x1)):
samples.append([x1[i],y1[i]])
low = -8.0
high = 8.0
width = 100
height = 200
scale = 1.1 * norm_pdf_multivariate(mu1,mu1,cov)
out = [8,16,32,64,128,256,512,1024,2048]
model = DPGMM(dims, 6)
for i,point in enumerate(samples):
model.add(point)
model.setPrior()
#p = ProgBar()
it = model.solve()
#del p
img = ones((height,width,3))
draw = model.sampleMixture()
for px in xrange(width):
x = float(px)/float(width) * (high-low) + low
for py in xrange(width):
y = float(py)/float(width) * (high-low) + low
# print "%f,%f"%(x,y)
print "%f\t%f\t%f"%(x,y,model.prob([x,y]))
sys.exit()
# Output parameters...
low = -2.0
high = 14.0
width = 800
height = 400
scale = 1.5 * max(map(lambda i: gt_weight[i]*gt[i].prob(gt[i].getMean()), xrange(len(gt))))
# Iterate a number of sample counts...
out = [8,16,32,64,128,256,512,1024,2048]
for dpc in out:
print '%i datapoints:'%(dpc)
# Fill in the model...
model = DPGMM(dims)
for point in samples[:dpc]: model.add(point)
model.setPrior()
# Solve...
p = ProgBar()
model = model.multiGrowSolve(8)
del p
# Now plot the estimated distribution against the actual distribution...
img = numpy.ones((height,width,3))
draw = model.sampleMixture()
for px in xrange(width):
x = float(px)/float(width) * (high-low) + low
# Output parameters...
low = -2.0
high = 14.0
width = 800
height = 400
scale = 1.5 * max(
map(lambda i: gt_weight[i] * gt[i].prob(gt[i].getMean()), xrange(len(gt))))
# Iterate a number of sample counts...
out = [8, 16, 32, 64, 128, 256, 512, 1024, 2048]
for dpc in out:
print '%i datapoints:' % (dpc)
# Fill in the model...
model = DPGMM(dims)
for point in samples[:dpc]:
model.add(point)
model.setPrior()
# Solve...
p = ProgBar()
model = model.multiGrowSolve(8)
del p
# Now plot the estimated distribution against the actual distribution...
img = numpy.ones((height, width, 3))
draw = model.sampleMixture()
for px in xrange(width):
x = float(px) / float(width) * (high - low) + low
try:
shutil.rmtree(out_dir)
except:
pass
os.mkdir(out_dir)
# Output parameters...
low = -2.0
high = 14.0
width = 800
height = 400
scale = 1.5 * max(
map(lambda i: gt_weight[i] * gt[i].prob(gt[i].getMean()), xrange(len(gt))))
# Fill in the model...
model = DPGMM(dims)
for point in samples:
model.add(point)
model.setPrior()
# Iterate over the number of sticks, increasing till it stops getting better...
prev = None
while True:
print 'Stick count = %i' % model.getStickCap()
p = ProgBar()
it = model.solve()
del p
print 'Updated fitting in %i iterations' % it
# Now plot the estimated distribution against the actual distribution...
img = numpy.ones((height, width, 3))
