def plotMotif(seq, motif, showExtracted=True, color='gray',
title=None, saveas=None):
start1 = motif[3]
start2 = motif[4]
end1 = start1 + len(motif[0]) - 1
end2 = start2 + len(motif[1]) - 1
# just show where the motif is in the original signal
if not showExtracted:
_, ax = plt.subplots()
ax.autoscale(False)
ax.plot(seq)
plotRect(ax, start1, end1, color=color)
plotRect(ax, start2, end2, color=color)
if saveas:
plt.savefig(saveas)
else:
plt.show()
return ax
# set up axes
ax1 = plt.subplot2grid((2,2), (0,0), colspan=2)
ax2 = plt.subplot2grid((2,2), (1,0))
ax3 = plt.subplot2grid((2,2), (1,1))
ax1.autoscale(tight=True)
ax2.autoscale(tight=True)
ax3.autoscale(tight=True)
# plot raw ts on top and motif instances on the bottom
ax1.plot(seq, lw=2)
ax2.plot(motif[0], lw=2)
ax3.plot(motif[1], lw=2)
ax1.set_title('Original Signal')
ax2.set_title('Motif Instance at %d' % start1)
ax3.set_title('Motif Instance at %d' % start2)
# draw rects in the ts where the motif is
plotRect(ax1, start1, end1, color=color)
plotRect(ax1, start2, end2, color=color)
plt.tight_layout()
if saveas:
plt.savefig(saveas)
else:
plt.show()
return ax1, ax2, ax3
def plotFFOutput(ts,
startIdxs,
endIdxs,
featureMat,
model,
cleanOverlap=False,
highlightInFeatureMat=False,
showGroundTruth=False):
axSeq, axSim, axPattern = plotSeqAndFeatures(ts.data,
featureMat,
model,
createModelAx=True,
padBothSides=True)
# our algorithm is allowed to return overlapping regions, but
# they look visually cluttered, so crudely un-overlap them
if cleanOverlap:
for i in range(1, len(startIdxs)):
start, end = startIdxs[i], endIdxs[i - 1]
if start < end:
middle = int((start + end) // 2)
startIdxs[i] = middle + 1
endIdxs[i - 1] = middle
# plot estimated regions
for startIdx, endIdx in zip(startIdxs, endIdxs):
plotRect(axSeq, startIdx, endIdx)
if highlightInFeatureMat:
# this is a hack to plot where instances are in the feature
# matrix; ideally, this info should come directly from
# the learning function
matOffset = 0 # 0, not below line, since we told it to pad both sides
matStartIdx = startIdx - matOffset
matEndIdx = endIdx - matOffset
plotRect(axSim, matStartIdx, matEndIdx, alpha=.2)
# plot ground truth regions
if showGroundTruth:
for startIdx, endIdx in zip(ts.startIdxs, ts.endIdxs):
plotRect(axSeq,
startIdx,
endIdx,
color='none',
hatch='---',
alpha=.3)
axSim.set_xlabel("Time")
axSim.set_ylabel("Feature")
plt.tight_layout()
def plotFFOutput(ts, startIdxs, endIdxs, featureMat, model,
cleanOverlap=False, highlightInFeatureMat=False,
showGroundTruth=False):
axSeq, axSim, axPattern = plotSeqAndFeatures(ts.data, featureMat, model,
createModelAx=True, padBothSides=True)
# our algorithm is allowed to return overlapping regions, but
# they look visually cluttered, so crudely un-overlap them
if cleanOverlap:
for i in range(1, len(startIdxs)):
start, end = startIdxs[i], endIdxs[i-1]
if start < end:
middle = int((start + end) // 2)
startIdxs[i] = middle + 1
endIdxs[i-1] = middle
# plot estimated regions
for startIdx, endIdx in zip(startIdxs, endIdxs):
plotRect(axSeq, startIdx, endIdx)
if highlightInFeatureMat:
# this is a hack to plot where instances are in the feature
# matrix; ideally, this info should come directly from
# the learning function
matOffset = 0 # 0, not below line, since we told it to pad both sides
matStartIdx = startIdx - matOffset
matEndIdx = endIdx - matOffset
plotRect(axSim, matStartIdx, matEndIdx, alpha=.2)
# plot ground truth regions
if showGroundTruth:
for startIdx, endIdx in zip(ts.startIdxs, ts.endIdxs):
plotRect(axSeq, startIdx, endIdx, color='none', hatch='---', alpha=.3)
axSim.set_xlabel("Time")
axSim.set_ylabel("Feature")
plt.tight_layout()
def showPairwiseSims(origSignal, m, simMat, clamp=True, pruneCorrAbove=-1,
plotMotifs=True, showEigenVect=False, hasPadding=True, saveas=None):
print "origSignal shape", origSignal.shape
# padLen = len(origSignal) - simMat.shape[1]
padLen = m - 1 if hasPadding else 0
subseqLen = m
plt.figure(figsize=(8,10))
if showEigenVect:
ax1 = plt.subplot2grid((20,10), (0,0), colspan=8, rowspan=5)
ax2 = plt.subplot2grid((20,10), (5,0), colspan=8, rowspan=15)
ax3 = plt.subplot2grid((20,10), (5,8), colspan=2, rowspan=15)
ax3.autoscale(tight=True)
ax3.set_title('Extracted')
else:
ax1 = plt.subplot2grid((4,2), (0,0), colspan=2)
ax2 = plt.subplot2grid((4,2), (1,0), colspan=2, rowspan=3)
ax1.autoscale(tight=True)
ax2.autoscale(tight=True)
ax1.set_title('Original Signal')
ax1.set_ylabel('Value')
if pruneCorrAbove > 0:
ax2.set_title('Subsequence Cosine Similarities to Dictionary Sequences')
else:
ax2.set_title('Subsequence Pairwise Cosine Similarities')
ax2.set_xlabel('Subsequence Start Index')
ax2.set_ylabel('"Dictionary" Sequence Number')
seq = origSignal
imgMat = simMat
print "imgMat shape: ", imgMat.shape
# # show magnitude of similarities in each row in descending order; there are
# # only about 60 entries > .01 in *any* row for msrc, and way fewer in most
# # plt.figure()
# # thresh = .5
# # sortedSimsByRow = np.sort(imgMat, axis=1)
# # sortedSimsByRow = sortedSimsByRow[:, ::-1]
# # nonzeroCols = np.sum(sortedSimsByRow, axis=0) > thresh # ignore tiny similarities
# # sortedSimsByRow = sortedSimsByRow[:, nonzeroCols]
# # # plt.imshow(sortedSimsByRow)
# # # plt.plot(np.mean(sortedSimsByRow, axis=1))
# # plt.plot(np.sum(sortedSimsByRow > thresh, axis=1)) # entries > thresh per row
# if pruneCorrAbove > 0.:
# print "ImgMat Shape:"
# print imgMat.shape
# imgMat = removeCorrelatedRows(imgMat, pruneCorrAbove)
# print imgMat.shape
# print "NaNs at:", np.where(np.isnan(imgMat))[0]
# print "Infs at:", np.where(np.isinf(imgMat))[0]
# power iteration to see what we get
if showEigenVect:
width = int(subseqLen * 1.5)
nRows, nCols = imgMat.shape
nPositions = nCols - width + 1
if nPositions > 1:
elementsPerPosition = nRows * width # size of 2d slice
dataMat = np.empty((nPositions, elementsPerPosition))
# for i in range(nPositions): # step by 1
for i in range(0, nPositions, width): # step by width, so non-overlapping
startCol = i
endCol = startCol + width
data = imgMat[:, startCol:endCol]
dataMat[i] = data.flatten()
# ah; power iteration is for cov matrix, cuz needs a square mat
# v = np.ones(elementsPerPosition) / elementsPerPosition # uniform start vect
# for i in range(3):
# v = np.dot(dataMat.T, v)
svd = TruncatedSVD(n_components=1, random_state=42)
svd.fit(dataMat)
v = svd.components_[0]
learnedFilt = v.reshape((nRows, width))
ax3.imshow(learnedFilt) # seems to be pretty good
# plt.show()
ax1.plot(seq)
ax2.imshow(imgMat, interpolation='nearest', aspect='auto')
plt.tight_layout()
if plotMotifs:
searchSeq = seq
print "searchSeq shape:", searchSeq.shape
motif = findMotif([searchSeq], subseqLen) # motif of min length
start1 = motif[3]
start2 = motif[4]
end1 = start1 + len(motif[0]) - 1
end2 = start2 + len(motif[1]) - 1
ax2.autoscale(False)
color = 'grey'
plotRect(ax1, start1, end1, color=color)
plotRect(ax2, start1, end1, color=color)
plotRect(ax1, start2, end2, color=color)
plotRect(ax2, start2, end2, color=color)
print "imgMat shape: ", imgMat.shape
print "padLen: ", padLen
if padLen:
searchSeq = imgMat[:,:-padLen].T
else:
searchSeq = imgMat.T
print "searchSeq shape:", searchSeq.shape
print "subseqLen:", subseqLen
motif = findMotif([searchSeq], subseqLen) # motif of min length
start1 = motif[3]
start2 = motif[4]
end1 = start1 + len(motif[0]) - 1
end2 = start2 + len(motif[1]) - 1
print [start1, end1, start2, end2]
color = 'm' # magenta
plotRect(ax1, start1, end1, color=color)
plotRect(ax2, start1, end1, color=color)
plotRect(ax1, start2, end2, color=color)
plotRect(ax2, start2, end2, color=color)
if saveas:
plt.savefig(saveas)
else:
plt.show()
if showEigenVect:
return ax1, ax2, ax3
return ax1, ax2
