def doTest(name, gen):
# Train the model...
df = DF()
df.setGoal(DensityGaussian(2)) # 2 = # of features
df.setGen(gen)
df.getPruner().setMinTrain(
48
) # Playing around shows that this is probably the most important number to get right when doing density estimation - the information gain heuristic just doesn't know when to stop.
global es
pb = ProgBar()
df.learn(
32, es, callback=pb.callback, mp=doMP
) # 32 = number of trees to learn - you need a lot to get a good answer.
del pb
# Drop some stats...
print '%i trees containing %i nodes.\nAverage error is %.3f.' % (
df.size(), df.nodes(), df.error())
# Visualise the density estimate...
global img
testSet = numpy.empty((pixel_width, 2), dtype=numpy.float32)
pb = ProgBar()
for y in xrange(pixel_width):
pb.callback(y, pixel_width)
i = 0
for x in xrange(pixel_width):
testSet[i, 0] = axis_half_width * float(
x - pixel_half_width) / pixel_half_width
testSet[i, 1] = axis_half_width * float(
y - pixel_half_width) / pixel_half_width
i += 1
test = MatrixES(testSet)
res = df.evaluate(test, mp=doMP)
i = 0
for x in xrange(pixel_width):
img[y, x, :] = res[i]
i += 1
del pb
print 'Maximum probability = %.2f' % img.max()
img /= img.max()
cv.SaveImage('test_de_circle_%s.png' % name, array2cv(img * 255))
def doTest(name, gen):
# Train the model...
df = DF()
df.setGoal(DensityGaussian(2)) # 2 = # of features
df.setGen(gen)
df.getPruner().setMinTrain(48) # Playing around shows that this is probably the most important number to get right when doing density estimation - the information gain heuristic just doesn't know when to stop.
global es
pb = ProgBar()
df.learn(32, es, callback = pb.callback, mp=doMP) # 32 = number of trees to learn - you need a lot to get a good answer.
del pb
# Drop some stats...
print '%i trees containing %i nodes.\nAverage error is %.3f.'%(df.size(), df.nodes(), df.error())
# Visualise the density estimate...
global img
testSet = numpy.empty((pixel_width,2), dtype=numpy.float32)
pb = ProgBar()
for y in xrange(pixel_width):
pb.callback(y,pixel_width)
i = 0
for x in xrange(pixel_width):
testSet[i,0] = axis_half_width * float(x - pixel_half_width) / pixel_half_width
testSet[i,1] = axis_half_width * float(y - pixel_half_width) / pixel_half_width
i += 1
test = MatrixES(testSet)
res = df.evaluate(test, mp=doMP)
i = 0
for x in xrange(pixel_width):
img[y,x,:] = res[i]
i += 1
del pb
print 'Maximum probability = %.2f'%img.max()
img /= img.max()
cv.SaveImage('test_de_circle_%s.png'%name,array2cv(img*255))
def doRun(tdc):
# Create a corpus...
c = lda.Corpus(4)
c.setWordCount(identCount()*4)
for i in xrange(tdc):
dic, abn = genDoc()
nDic = dict()
for key,item in dic.iteritems(): nDic[key[0]*4+key[1]] = item
doc = lda.Document(nDic)
doc.abn = abn
c.add(doc)
# Fit a model...
params = lda.Params()
params.setRuns(16)
print 'Fitting model...'
p = ProgBar()
c.fit(params,p.callback)
del p
tw = c.topicsWords()
# Test on a bunch of documents, creating a list of abnormality score/actually an abnormality pairs...
ab_gt = []
print 'Testing...'
p = ProgBar()
for i in xrange(testDocCount):
p.callback(i,testDocCount)
dic, abn = genDoc()
nDic = dict()
for key,item in dic.iteritems(): nDic[key[0]*4+key[1]] = item
doc = lda.Document(nDic)
doc.fit(tw)
ab_gt.append((doc.negLogLikelihood(tw),abn))
del p
ab_gt.sort(reverse=True)
# Use the pairs to construct a roc...
posCount = len(filter(lambda p:p[1]==True,ab_gt))
negCount = len(ab_gt) - posCount
print 'positive samples = ',posCount
print 'negative samples = ',negCount
truePos = 0
falsePos = 0
trueNeg = negCount
falseNeg = posCount
roc = []
for p in ab_gt:
if p[1]:
truePos += 1
falseNeg -= 1
else:
falsePos +=1
trueNeg -= 1
pnt = (float(falsePos)/float(falsePos+trueNeg), float(truePos)/float(truePos+falseNeg))
roc.append(pnt)
# Save the roc to disk...
if not sweep:
f = open('junction_roc.txt','w')
f.write('0.0 0.0\n')
for pnt in roc: f.write('%f %f\n'%pnt)
f.close()
# Calculate and print out the area under the roc...
area = 0.0
for i in xrange(1,len(roc)):
area += 0.5*(roc[i-1][1]+roc[i][1]) * (roc[i][0]-roc[i-1][0])
print 'area under roc =',area, '(above',(1.0-area),')'
return area
lambda: ms.set_scale(numpy.array([5.0, 5.0]))),
('Silverman', ms.scale_silverman), ('Scott', ms.scale_scott),
('loo_nll', scale_loo_nll)]:
# Calculate and print out the scales...
print '<', name, '>'
alg()
print 'Scale:', ms.get_scale()
print 'loo nll for this scale =', ms.loo_nll()
mean, sd = ms.stats()
print 'mean = (%f, %f); sd = (%f, %f)' % (mean[0], mean[1], sd[0], sd[1])
# Render out a normalised probability map...
image = numpy.zeros((dim, dim, 3), dtype=numpy.float32)
p = ProgBar()
for row in xrange(dim):
p.callback(row, dim)
sam = numpy.append(numpy.linspace(-size, size, dim).reshape(
(-1, 1)), ((row / (dim - 1.0) - 0.5) * 2.0 * size) *
numpy.ones(dim).reshape((-1, 1)),
axis=1)
image[row, :, :] = ms.probs(sam).reshape((-1, 1))
del p
print 'Largest sampled probability =', image.max()
image *= 255.0 / image.max()
image = array2cv(image)
cv.SaveImage('bandwidth_%s.png' % name, image)
print
# Now create some documents and test them...
# Below indexed by [actual, result], with 0 to mean normal and 1 for abnormal...
print 'Testing model on new documents...'
result = numpy.zeros((6,6),dtype=numpy.int32)
testCounts = [normal_doc_test, square_doc_test, hor_star_doc_test, vert_star_doc_test, hor_in_vert_doc_test, vert_in_hor_doc_test]
p = ProgBar()
stepsDone = 0
stepsTotal = sum(testCounts)
for ab in ([None] + [x for x in sam.getAbnormDict().iterkeys()]):
c = behClasses.index(ab)
for _ in xrange(testCounts[c]):
p.callback(stepsDone,stepsTotal)
stepsDone += 1
doc = ddhdp.Document(sampleDocument(ab))
abnormList = model.mlDocAbnorm(doc, lone = True, cap = 0) ######## Cap is super low for testing.
truth = c
guess = 0
if len(abnormList)!=0:
guess = behClasses.index(abnormList[0]) # Doesn't handle it thinking that multiple abnormalities are present.
result[truth,guess] += 1
del p
print 'Confusion matrix:'
print 'Converting... (%i pixels at a time)'%step
slices = map(lambda x: slice(x*step, (x+1)*step), xrange(data.shape[0]//step + 1))
if slices[-1].stop<data.shape[0]:
slices.append(slice(slices[-1].stop, data.shape[0]))
## Calculate each channel in turn...
out = data.copy()
for cc in xrange(3):
if not args.quiet:
print 'Converting - %s...'%(['red', 'green', 'blue'][cc])
p = ProgBar()
for i,s in enumerate(slices):
p.callback(i, len(slices))
out[s,cc] = model[cc](data[s,:].astype(numpy.float32))
del p
## Expand the data matrix back up to the order and length of the image, by expanding duplicates...
source = numpy.cumsum(keep) - 1
out = out[source,:]
out = out[numpy.argsort(index),:]
## Convert back from data matrix to image...
out = out.reshape(image.shape)
# Clamp unreasonable values, record where clamping occurs...
if not args.quiet:
# Choose a reasonable size...
print 'Selecting size using loo:'
p = ProgBar()
ms.scale_loo_nll(callback = p.callback)
del p
# Plot the pdf, for reference...
image = numpy.zeros((pixels, pixels, 3), dtype=numpy.float32)
print 'Rendering probability map:'
p = ProgBar()
for row in xrange(pixels):
p.callback(row, pixels)
sam = numpy.append(numpy.linspace(0.0, 1.0, pixels).reshape((-1,1)), (row / float(pixels-1)) * numpy.ones(pixels).reshape((-1,1)), axis=1)
image[row, :, :] = ms.probs(sam).reshape((-1,1))
del p
image *= 255.0 / image.max()
image = array2cv(image)
cv.SaveImage('draw_weighted_density.png', image)
# Draw a bunch of points from it and plot them...
samples = numpy.random.randint(16, 512)
draw = ms.draws(1024)
image = numpy.zeros((pixels, pixels, 3), dtype=numpy.float32)
def doRun(tdc):
# Create a corpus...
c = lda.Corpus(4)
c.setWordCount(identCount() * 4)
for i in xrange(tdc):
dic, abn = genDoc()
nDic = dict()
for key, item in dic.iteritems():
nDic[key[0] * 4 + key[1]] = item
doc = lda.Document(nDic)
doc.abn = abn
c.add(doc)
# Fit a model...
params = lda.Params()
params.setRuns(16)
print 'Fitting model...'
p = ProgBar()
c.fit(params, p.callback)
del p
tw = c.topicsWords()
# Test on a bunch of documents, creating a list of abnormality score/actually an abnormality pairs...
ab_gt = []
print 'Testing...'
p = ProgBar()
for i in xrange(testDocCount):
p.callback(i, testDocCount)
dic, abn = genDoc()
nDic = dict()
for key, item in dic.iteritems():
nDic[key[0] * 4 + key[1]] = item
doc = lda.Document(nDic)
doc.fit(tw)
ab_gt.append((doc.negLogLikelihood(tw), abn))
del p
ab_gt.sort(reverse=True)
# Use the pairs to construct a roc...
posCount = len(filter(lambda p: p[1] == True, ab_gt))
negCount = len(ab_gt) - posCount
print 'positive samples = ', posCount
print 'negative samples = ', negCount
truePos = 0
falsePos = 0
trueNeg = negCount
falseNeg = posCount
roc = []
for p in ab_gt:
if p[1]:
truePos += 1
falseNeg -= 1
else:
falsePos += 1
trueNeg -= 1
pnt = (float(falsePos) / float(falsePos + trueNeg),
float(truePos) / float(truePos + falseNeg))
roc.append(pnt)
# Save the roc to disk...
if not sweep:
f = open('junction_roc.txt', 'w')
f.write('0.0 0.0\n')
for pnt in roc:
f.write('%f %f\n' % pnt)
f.close()
# Calculate and print out the area under the roc...
area = 0.0
for i in xrange(1, len(roc)):
area += 0.5 * (roc[i - 1][1] + roc[i][1]) * (roc[i][0] - roc[i - 1][0])
print 'area under roc =', area, '(above', (1.0 - area), ')'
return area
ms = map(to_ms, samples)
# Infer a good loo value for the first one, then set them all to the same...
p = ProgBar()
ms[0].scale_loo_nll(callback=p.callback)
del p
for i in xrange(1, 4):
ms[i].copy_scale(ms[0])
# Visualise the distributions using KDE...
imgs = []
p = ProgBar()
for i in xrange(4):
p.callback(i, 4)
img = numpy.zeros((draw_scale * size[0], draw_scale * size[1]),
dtype=numpy.float32)
sweep0 = numpy.linspace(0, size[0], img.shape[0])
sweep1 = numpy.linspace(0, size[1], img.shape[1])
for ij, j in enumerate(sweep0):
points = numpy.append(j * numpy.ones(sweep1.shape[0]).reshape(
(-1, 1)),
sweep1.reshape((-1, 1)),
axis=1)
img[ij, :] = ms[i].probs(points)
img *= 255.0 / img.max()
imgs.append(img)
loo.addSample(numpy.reshape(data.getVectors()[i], (1,1))) loo.solve(p.callback) precision = loo.getBest() del p print 'Optimal standard deviation = %s'%str(math.sqrt(1.0/precision[0,0])) # Create and fill the pool... print 'Filling the pool...' pool = Pool() p = ProgBar() for i in xrange(data.getVectors().shape[0]): p.callback(i, data.getVectors().shape[0]) pool.store(numpy.reshape(data.getVectors()[i], (1,)), data.getClasses()[i]) del p # Create the classifier... classifier = ClassifyKDE(precision) # Calculate the dimensions for the visualisations... low = data.getVectors().min() high = data.getVectors().max() low -= 0.2*(high-low) high += 0.2*(high-low)
if img[t_y,t_x,0] < t:
img[t_y,t_x,:] = t
except:
pass
img = array2cv(255.0 * img)
cv.SaveImage('composite_draw.png', img)
# Visualise the probability - both spatial and rotational in a single image, with one colour channel each for 3 directions...
img = numpy.zeros((size, size, 3), dtype=numpy.float32)
p = ProgBar()
for y in xrange(size):
p.callback(y, size)
for index, orient_x, orient_y in [(0,1.0,0.0), (1,0.0,1.0), (2,-1.0,0.0)]:
block = numpy.concatenate(((scale * y / float(size-1)) * numpy.ones(size).reshape((-1,1)), numpy.linspace(0.0, scale, size).reshape((-1,1)), orient_x * numpy.ones(size).reshape((-1,1)), orient_y * numpy.ones(size).reshape((-1,1))), axis=1)
vals = ms.probs(block)
img[y,:,index] = vals
del p
img *= 255 / img.max()
img = array2cv(img)
cv.SaveImage('composite_prob.png', img)
slices = map(lambda x: slice(x * step, (x + 1) * step),
xrange(data.shape[0] // step + 1))
if slices[-1].stop < data.shape[0]:
slices.append(slice(slices[-1].stop, data.shape[0]))
## Calculate each channel in turn...
out = data.copy()
for cc in xrange(3):
if not args.quiet:
print 'Converting - %s...' % (['red', 'green', 'blue'][cc])
p = ProgBar()
for i, s in enumerate(slices):
p.callback(i, len(slices))
out[s, cc] = model[cc](data[s, :].astype(numpy.float32))
del p
## Expand the data matrix back up to the order and length of the image, by expanding duplicates...
source = numpy.cumsum(keep) - 1
out = out[source, :]
out = out[numpy.argsort(index), :]
## Convert back from data matrix to image...
out = out.reshape(image.shape)
# Clamp unreasonable values, record where clamping occurs...
if not args.quiet:
print 'Clamping...'
mask = numpy.zeros((out.shape[0], out.shape[1]), dtype=numpy.bool)
t_y = int(t * s_y + (1 - t) * e_y)
try:
if img[t_y, t_x, 0] < t:
img[t_y, t_x, :] = t
except:
pass
img = array2cv(255.0 * img)
cv.SaveImage('composite_draw.png', img)
# Visualise the probability - both spatial and rotational in a single image, with one colour channel each for 3 directions...
img = numpy.zeros((size, size, 3), dtype=numpy.float32)
p = ProgBar()
for y in xrange(size):
p.callback(y, size)
for index, orient_x, orient_y in [(0, 1.0, 0.0), (1, 0.0, 1.0),
(2, -1.0, 0.0)]:
block = numpy.concatenate(
((scale * y / float(size - 1)) * numpy.ones(size).reshape(
(-1, 1)), numpy.linspace(0.0, scale, size).reshape(
(-1, 1)), orient_x * numpy.ones(size).reshape(
(-1, 1)), orient_y * numpy.ones(size).reshape(
(-1, 1))),
axis=1)
vals = ms.probs(block)
img[y, :, index] = vals
del p
p = ProgBar()
loo = PrecisionLOO()
for i in xrange(data.getVectors().shape[0]):
loo.addSample(numpy.reshape(data.getVectors()[i], (1, 1)))
loo.solve(p.callback)
precision = loo.getBest()
del p
print 'Optimal standard deviation = %s' % str(math.sqrt(1.0 / precision[0, 0]))
# Create and fill the pool...
print 'Filling the pool...'
pool = Pool()
p = ProgBar()
for i in xrange(data.getVectors().shape[0]):
p.callback(i, data.getVectors().shape[0])
pool.store(numpy.reshape(data.getVectors()[i], (1, )),
data.getClasses()[i])
del p
# Create the classifier...
classifier = ClassifyKDE(precision)
# Calculate the dimensions for the visualisations...
low = data.getVectors().min()
high = data.getVectors().max()
low -= 0.2 * (high - low)
high += 0.2 * (high - low)
# Quickly visualise the dataset as lines in an image...
weight = 1.0
def doRun(tdc):
# Create a corpus...
vlda = lda.VLDA(4, identCount()*4)
abnDict = dict()
for i in xrange(tdc):
dic, abn = genDoc()
nDic = dict()
for key,item in dic.iteritems(): nDic[key[0]*4+key[1]] = item
doc = vlda.add(nDic)
abnDict[doc] = abn
# Fit a model...
print 'Fitting model...'
p = ProgBar()
vlda.solve()
del p
# Visualise the topics...
if not sweep:
for t in xrange(vlda.numTopics()):
prob = numpy.zeros((6,6,4),dtype=numpy.float_)
beta = vlda.getBeta(t)
for i in xrange(beta.shape[0]):
x,y = identToCoord(i//4)
w = i%4
prob[x,y,w] += beta[i]
multProb = 255.0/prob.max()
img = cv.CreateImage((6*25,6*25),cv.IPL_DEPTH_32F,3)
for y in xrange(6):
for x in xrange(6):
coords = [(x*25,y*25),((x+1)*25,y*25),((x+1)*25,(y+1)*25),(x*25,(y+1)*25)]
centre = (x*25+12,y*25+12)
for d in xrange(4):
if d%2==0:
col = cv.RGB(0.0,prob[x,y,d]*multProb,0.0)
else:
col = cv.RGB(prob[x,y,d]*multProb,0.0,0.0)
cv.FillPoly(img, [(coords[d],coords[(d+1)%4],centre)], col)
cv.SaveImage('junction_topic_%i.png'%t,img)
# Test on a bunch of documents, creating a list of abnormality score/actually an abnormality pairs...
ab_gt = []
print 'Testing...'
p = ProgBar()
for i in xrange(testDocCount):
p.callback(i,testDocCount)
dic, abn = genDoc()
nDic = dict()
for key,item in dic.iteritems(): nDic[key[0]*4+key[1]] = item
nll = vlda.getNewNLL(nDic)
ab_gt.append((nll,abn))
del p
ab_gt.sort(reverse=True)
# Use the pairs to construct a roc...
posCount = len(filter(lambda p:p[1]==True,ab_gt))
negCount = len(ab_gt) - posCount
print 'positive samples = ',posCount
print 'negative samples = ',negCount
truePos = 0
falsePos = 0
trueNeg = negCount
falseNeg = posCount
roc = []
for p in ab_gt:
if p[1]:
truePos += 1
falseNeg -= 1
else:
falsePos +=1
trueNeg -= 1
pnt = (float(falsePos)/float(falsePos+trueNeg), float(truePos)/float(truePos+falseNeg))
roc.append(pnt)
# Save the roc to disk...
if not sweep:
f = open('junction_roc.csv','w')
f.write('0.0, 0.0\n')
for pnt in roc: f.write('%f, %f\n'%pnt)
f.close()
# Calculate and print out the area under the roc...
area = 0.0
for i in xrange(1,len(roc)):
area += 0.5*(roc[i-1][1]+roc[i][1]) * (roc[i][0]-roc[i-1][0])
print 'area under roc =',area, '(above',(1.0-area),')'
return area
# Choose a reasonable size...
print 'Selecting size using loo:'
p = ProgBar()
ms.scale_loo_nll(callback = p.callback)
del p
# Render out a normalised probability map...
image = numpy.zeros((dim, dim, 3), dtype=numpy.float32)
print 'Rendering probability map:'
p = ProgBar()
for row in xrange(dim):
p.callback(row, dim)
sam = numpy.append(numpy.linspace(-size, size, dim).reshape((-1,1)), ((row / (dim-1.0) - 0.5) * 2.0 * size) * numpy.ones(dim).reshape((-1,1)), axis=1)
image[row, :, :] = ms.probs(sam).reshape((-1,1))
del p
image *= 255.0 / image.max()
image = array2cv(image)
cv.SaveImage('draw_density.png', image)
# Draw a new set of samples from the KDE approximation of the distribution, and visualise...
draw = ms.draws(data.shape[0])
image = numpy.zeros((dim, dim, 3), dtype=numpy.float32)
def doRun(tdc):
# Create directory to put images into...
if not sweep:
try:
os.makedirs('junction')
except:
pass
# Create a corpus...
c = rlda.Corpus(10,4)
c.setIdentWordCounts(identCount(),4)
for i in xrange(tdc):
dic, abn = genDoc(False)
doc = rlda.Document(dic)
doc.abn = abn
c.add(doc)
if not sweep:
prob = numpy.zeros((6,6,4),dtype=numpy.float_)
for key,item in dic.iteritems():
x,y = identToCoord(key[0])
prob[x,y,key[1]] = item
multProb = 255.0/prob.max()
img = cv.CreateImage((6*25,6*25),cv.IPL_DEPTH_32F,3)
for y in xrange(6):
for x in xrange(6):
coords = [(x*25,y*25),((x+1)*25,y*25),((x+1)*25,(y+1)*25),(x*25,(y+1)*25)]
centre = (x*25+12,y*25+12)
for d in xrange(4):
if d%2==0:
col = cv.RGB(0.0,prob[x,y,d]*multProb,0.0)
else:
col = cv.RGB(prob[x,y,d]*multProb,0.0,0.0)
cv.FillPoly(img, [(coords[d],coords[(d+1)%4],centre)], col)
cv.SaveImage('junction/xdoc_%i_%s.png'%(i,str(abn)),img)
# Fit a model...
params = rlda.Params()
params.setRuns(16)
print 'Fitting model...'
p = ProgBar()
c.fit(params,p.callback)
del p
ir = c.getIR()
wrt = c.getWRT()
# Visualise the regions...
if not sweep:
mult = 255.0/ir.max()
for r in xrange(ir.shape[1]):
rend = numpy.zeros((6,6),dtype=numpy.float_)
for i in xrange(ir.shape[0]): rend[identToCoord(i)] = ir[i,r] * mult
rend = numpy.repeat(numpy.repeat(rend,25,axis=0),25,axis=1)
cv.SaveImage('junction/region_%i.png'%r,array2cv(rend))
# Visualise the topics...
if not sweep:
for t in xrange(wrt.shape[2]):
prob = numpy.zeros((6,6,4),dtype=numpy.float_)
for i in xrange(ir.shape[0]):
x,y = identToCoord(i)
for r in xrange(wrt.shape[1]):
for w in xrange(wrt.shape[0]):
prob[x,y,w] += ir[i,r] * wrt[w,r,t]
multProb = 255.0/prob.max()
img = cv.CreateImage((6*25,6*25),cv.IPL_DEPTH_32F,3)
for y in xrange(6):
for x in xrange(6):
coords = [(x*25,y*25),((x+1)*25,y*25),((x+1)*25,(y+1)*25),(x*25,(y+1)*25)]
centre = (x*25+12,y*25+12)
for d in xrange(4):
if d%2==0:
col = cv.RGB(0.0,prob[x,y,d]*multProb,0.0)
else:
col = cv.RGB(prob[x,y,d]*multProb,0.0,0.0)
cv.FillPoly(img, [(coords[d],coords[(d+1)%4],centre)], col)
cv.SaveImage('junction/topic_%i.png'%t,img)
# Test on a bunch of documents, creating a list of abnormality score/actually an abnormality pairs...
ab_gt = []
print 'Testing...'
p = ProgBar()
for i in xrange(testDocCount):
p.callback(i,testDocCount)
dic, abn = genDoc()
doc = rlda.Document(dic)
doc.fit(ir,wrt)
ab_gt.append((doc.negLogLikeRegionVec().max(),abn))
del p
ab_gt.sort(reverse=True)
# Use the pairs to construct a roc...
posCount = len(filter(lambda p:p[1]==True,ab_gt))
negCount = len(ab_gt) - posCount
print 'positive samples = ',posCount
print 'negative samples = ',negCount
truePos = 0
falsePos = 0
trueNeg = negCount
falseNeg = posCount
roc = []
for p in ab_gt:
if p[1]:
truePos += 1
falseNeg -= 1
else:
falsePos +=1
trueNeg -= 1
pnt = (float(falsePos)/float(falsePos+trueNeg), float(truePos)/float(truePos+falseNeg))
roc.append(pnt)
# Save the roc to disk...
if not sweep:
f = open('junction_roc.txt','w')
f.write('0.0 0.0\n')
for pnt in roc: f.write('%f %f\n'%pnt)
f.close()
# Calculate and print out the area under the roc...
area = 0.0
for i in xrange(1,len(roc)):
area += 0.5*(roc[i-1][1]+roc[i][1]) * (roc[i][0]-roc[i-1][0])
print 'area under roc =',area, '(above',(1.0-area),')'
return area
ms.set_kernel('triangular')
ms.set_spatial('kd_tree')
# Choose a reasonable size...
print 'Selecting size using loo:'
p = ProgBar()
ms.scale_loo_nll(callback=p.callback)
del p
# Plot the pdf, for reference...
image = numpy.zeros((pixels, pixels, 3), dtype=numpy.float32)
print 'Rendering probability map:'
p = ProgBar()
for row in xrange(pixels):
p.callback(row, pixels)
sam = numpy.append(numpy.linspace(0.0, 1.0, pixels).reshape((-1, 1)),
(row / float(pixels - 1)) * numpy.ones(pixels).reshape(
(-1, 1)),
axis=1)
image[row, :, :] = ms.probs(sam).reshape((-1, 1))
del p
image *= 255.0 / image.max()
image = array2cv(image)
cv.SaveImage('draw_weighted_density.png', image)
# Draw a bunch of points from it and plot them...
samples = numpy.random.randint(16, 512)
draw = ms.draws(1024)
def doRun(tdc):
# Create directory to put images into...
if not sweep:
try:
os.makedirs('junction')
except:
pass
# Create a corpus...
c = rlda.Corpus(10,4)
c.setIdentWordCounts(identCount(),4)
for i in xrange(tdc):
dic, abn = genDoc(False)
doc = rlda.Document(dic)
doc.abn = abn
c.add(doc)
if not sweep:
prob = numpy.zeros((6,6,4),dtype=numpy.float_)
for key,item in dic.iteritems():
x,y = identToCoord(key[0])
prob[x,y,key[1]] = item
multProb = 255.0/prob.max()
img = cv.CreateImage((6*25,6*25),cv.IPL_DEPTH_32F,3)
for y in xrange(6):
for x in xrange(6):
coords = [(x*25,y*25),((x+1)*25,y*25),((x+1)*25,(y+1)*25),(x*25,(y+1)*25)]
centre = (x*25+12,y*25+12)
for d in xrange(4):
if d%2==0:
col = cv.RGB(0.0,prob[x,y,d]*multProb,0.0)
else:
col = cv.RGB(prob[x,y,d]*multProb,0.0,0.0)
cv.FillPoly(img, [(coords[d],coords[(d+1)%4],centre)], col)
cv.SaveImage('junction/xdoc_%i_%s.png'%(i,str(abn)),img)
# Fit a model...
params = rlda.Params()
params.setRuns(16)
print 'Fitting model...'
p = ProgBar()
c.fit(params,p.callback)
del p
ir = c.getIR()
wrt = c.getWRT()
# Visualise the regions...
if not sweep:
mult = 255.0/ir.max()
for r in xrange(ir.shape[1]):
rend = numpy.zeros((6,6),dtype=numpy.float_)
for i in xrange(ir.shape[0]): rend[identToCoord(i)] = ir[i,r] * mult
rend = numpy.repeat(numpy.repeat(rend,25,axis=0),25,axis=1)
cv.SaveImage('junction/region_%i.png'%r,array2cv(rend))
# Visualise the topics...
if not sweep:
for t in xrange(wrt.shape[2]):
prob = numpy.zeros((6,6,4),dtype=numpy.float_)
for i in xrange(ir.shape[0]):
x,y = identToCoord(i)
for r in xrange(wrt.shape[1]):
for w in xrange(wrt.shape[0]):
prob[x,y,w] += ir[i,r] * wrt[w,r,t]
multProb = 255.0/prob.max()
img = cv.CreateImage((6*25,6*25),cv.IPL_DEPTH_32F,3)
for y in xrange(6):
for x in xrange(6):
coords = [(x*25,y*25),((x+1)*25,y*25),((x+1)*25,(y+1)*25),(x*25,(y+1)*25)]
centre = (x*25+12,y*25+12)
for d in xrange(4):
if d%2==0:
col = cv.RGB(0.0,prob[x,y,d]*multProb,0.0)
else:
col = cv.RGB(prob[x,y,d]*multProb,0.0,0.0)
cv.FillPoly(img, [(coords[d],coords[(d+1)%4],centre)], col)
cv.SaveImage('junction/topic_%i.png'%t,img)
# Test on a bunch of documents, creating a list of abnormality score/actually an abnormality pairs...
ab_gt = []
print 'Testing...'
p = ProgBar()
for i in xrange(testDocCount):
p.callback(i,testDocCount)
dic, abn = genDoc()
doc = rlda.Document(dic)
doc.fit(ir,wrt)
ab_gt.append((doc.negLogLikeRegionVec().max(),abn))
del p
ab_gt.sort(reverse=True)
# Use the pairs to construct a roc...
posCount = len(filter(lambda p:p[1]==True,ab_gt))
negCount = len(ab_gt) - posCount
print 'positive samples = ',posCount
print 'negative samples = ',negCount
truePos = 0
falsePos = 0
trueNeg = negCount
falseNeg = posCount
roc = []
for p in ab_gt:
if p[1]:
truePos += 1
falseNeg -= 1
else:
falsePos +=1
trueNeg -= 1
pnt = (float(falsePos)/float(falsePos+trueNeg), float(truePos)/float(truePos+falseNeg))
roc.append(pnt)
# Save the roc to disk...
if not sweep:
f = open('junction_roc.txt','w')
f.write('0.0 0.0\n')
for pnt in roc: f.write('%f %f\n'%pnt)
f.close()
# Calculate and print out the area under the roc...
area = 0.0
for i in xrange(1,len(roc)):
area += 0.5*(roc[i-1][1]+roc[i][1]) * (roc[i][0]-roc[i-1][0])
print 'area under roc =',area, '(above',(1.0-area),')'
return area
# Now create some documents and test them...
# Below indexed by [actual, result], with 0 to mean normal and 1 for abnormal...
print 'Testing model on new documents...'
result = numpy.zeros((6,6),dtype=numpy.int32)
testCounts = [normal_doc_test, square_doc_test, hor_star_doc_test, vert_star_doc_test, hor_in_vert_doc_test, vert_in_hor_doc_test]
p = ProgBar()
stepsDone = 0
stepsTotal = sum(testCounts)
for ab in ([None] + [x for x in sam.getAbnormDict().iterkeys()]):
c = behClasses.index(ab)
for _ in xrange(testCounts[c]):
p.callback(stepsDone,stepsTotal)
stepsDone += 1
doc = ddhdp.Document(sampleDocument(ab))
abnormList = model.mlDocAbnorm(doc, lone = True, cap = 0) ######## Cap is super low for testing.
truth = c
guess = 0
if len(abnormList)!=0:
guess = behClasses.index(abnormList[0]) # Doesn't handle it thinking that multiple abnormalities are present.
result[truth,guess] += 1
del p
print 'Confusion matrix:'