def main(): return from flock import learnFFfromSeq from dig import datasets np.random.seed(12345) # read in the first few time series from the TIDIGITS dataset; the return # value is a collection of LabeledTimeSeries (see datasets.utils). You # will of course need to have the relevant dataset on your machine, as # well as update datasets/paths.py to point to it. For TIDIGITS # specifically, you will also need to have librosa installed. For the # UCR datasets, the whichExamples argument takes this many examples from # all 20 datasets whichExamples = np.arange(2) tsList = datasets.loadDataset(datasets.TIDIGITS, whichExamples=whichExamples) # uncomment any of these to use a different dataset # tsList = datasets.loadDataset(datasets.DISHWASHER, whichExamples=whichExamples) # tsList = datasets.loadDataset(datasets.MSRC, whichExamples=whichExamples) # tsList = datasets.loadDataset(datasets.UCR, whichExamples=[0]) Lmin, Lmax = 1./20, 1./10 # fractions of time series length for ts in tsList: startIdxs, endIdxs, model, featureMat, featureMatBlur = learnFFfromSeq( ts.data, Lmin, Lmax) plotFFOutput(ts, startIdxs, endIdxs, featureMat, model) # you can also call this if you just want to see what the data looks like # ts.plot() # plt.savefig(ts.name + '.pdf') # use this to save it plt.show()
def main():
# tsList = ds.loadDataset(ds.SINES)
# tsList = ds.loadDataset(ds.MSRC, whichExamples=range(1))
# tsList = ds.loadDataset(ds.MSRC, whichExamples=range(2))
# tsList = ds.loadDataset(ds.DISHWASHER, whichExamples=range(1))
tsList = ds.loadDataset(ds.DISHWASHER, whichExamples=range(5))
# tsList = ds.loadDataset(ds.TIDIGITS, whichExamples=range(2))
# tsList = ds.loadDataset(ds.TIDIGITS, whichExamples=range(10))
for ts in tsList:
# ts.data = ts.data[:, :5]
# ts.data = ts.data[::2, :]
# make it have 3 signals with increasing amounts of noise
# n = ts.data.shape[0]
# noiseDim1 = ts.data + np.random.randn(n, 1) * .25
# noiseDim2 = np.random.randn(n).reshape((-1, 1))
# ts.data = np.hstack((ts.data, noiseDim1, noiseDim2))
# seq = ts.data.ravel().reshape((-1, 1))
# print seq.shape
# print ts.data.shape
# return
# Lmin = .1
# Lmax = .2
Lmin = .05
Lmax = .1
Lmin_int = int(Lmin * len(ts.data))
Lmax_int = int(Lmax * len(ts.data))
# ts.data = ts.data[:, ::5] # downsample by factor of 5
# plt.plot(ts.data[:, :5])
# plt.show()
# return
ff = dig.FlockLearner()
ff.learn(ts.data.T, Lmin, Lmax)
T = ff.getTimeSeries()
Phi = ff.getFeatureMat()
PhiBlur = ff.getBlurredFeatureMat()
W = ff.getPattern()
Wsum = ff.getPatternSum()
starts = ff.getInstanceStartIdxs()
ends = ff.getInstanceEndIdxs()
windowLen = ff.getWindowLen()
print "windowLen: ", windowLen
# scores = ff.getSeedScores()
seeds = ff.getSeeds()
# startIdxs, endIdxs, model, X, Xblur = flock.learnFFfromSeq(ts.data, Lmin,
# Lmax)
# ts.plot()
printCandidates = False
if printCandidates:
FPhi = Phi
FPhiBlur = PhiBlur
for FPhi, FPhiBlur in [(Phi, PhiBlur), (X, Xblur)]:
for seed in seeds:
# Lmax = int(len(ts.data))
# dotProds = dig.dotProdsForSeed(FPhi, FPhiBlur, seed, windowLen)
dotProds = dig.dotProdsForSeed(FPhi, FPhiBlur, seed, windowLen)
dotProds2 = flock.dotProdsForSeed(FPhi, FPhiBlur, seed, windowLen)
# plt.figure()
# plt.plot(dotProds)
# plt.plot(dotProds2)
candidates = dig.candidatesFromDotProds(dotProds, Lmin_int)
candidates2 = flock.candidatesFromDotProds(dotProds, Lmin_int)
# candidates = ff.getCandidatesForSeed(FPhi, FPhiBlur, long(seed))
# candidates2 = flock.get
print "{}) {} vs {}".format(seed, candidates, candidates2)
assert np.allclose(dotProds, dotProds2)
# cpp impl seems to be much more accurate than python
# version, which has a known bug in the func that
# finds spaced relative maxima
# if len(candidates) != len(candidates2):
# plt.plot(dotProds)
# plt.plot(dotProds2)
# plt.show()
# assert(False)
# okay, so it's receiving, storing, and returning mats correctly
# print "T shape", T.shape
# ts.plot()
# plt.plot(T.T)
# plt.title("Orig TS and TS from FF object")
# alright, seed scores look awesome...
# scores = ff.getSeedScores()
# plt.figure()
# plt.plot(scores)
# are structure score computations the same?
# -> little bit of noise, but basically yes
# subseqLen = 40
# walks = dig.createRandomWalks(seq.ravel(), subseqLen, 100)
# structureScores = dig.structureScores1D(seq.ravel(), subseqLen, walks)
# structureScores2 = feat.windowScoresRandWalk(ts.data, subseqLen)
# plt.figure()
# plt.plot(structureScores / np.max(structureScores))
# plt.plot(structureScores2)
# are seed scores about right?
# scores = ff.getSeedScores()
# scores2 = flock._seedScores(seq, Lmin_int)
# plt.figure()
# plt.plot(scores)
# plt.plot(scores2)
# okay, so python impl nails it; clearly we have a bug somewhere...
# startIdxs, endIdxs, model, X, Xblur = flock.learnFFfromSeq(ts.data, Lmin,
# Lmax)
# plt.figure()
# viz.imshowBetter(model)
# plt.colorbar()
# plt.figure()
# viz.imshowBetter(X)
# plt.colorbar()
# plt.figure()
# viz.imshowBetter(Xblur)
# plt.colorbar()
print "cpp startIdxs:", starts
print "cpp endIdxs:", ends
# print "python startIdxs:", startIdxs
# print "python endIdxs:", endIdxs
# random walks seem to be good
# walks = dig.createRandomWalks(ts.data.ravel(), 80, 100)
# plt.plot(walks.T)
# for walk in walks:
# if np.abs(np.mean(walk)) > .0001:
# print np.mean(walk)
# print "walk is not zero-meaned!"
# phi has reasonable shape (a few rows, same num cols = ts len)
# print "Phi shape", Phi.shape
# print "PhiBlur shape", PhiBlur.shape
# print "Phi:"
# for row in Phi[:20]:
# print row
ploFeatureMats = 0
if ploFeatureMats:
# yep, phi and phi_blur look right
# well, sort of...phi consistently has fewer 1s...
_, axes = plt.subplots(2,2)
viz.imshowBetter(Phi, ax=axes[0,0])
viz.imshowBetter(PhiBlur, ax=axes[1,0])
viz.imshowBetter(X, ax=axes[0,1])
viz.imshowBetter(Xblur, ax=axes[1,1])
# ya, phi_blur and blurring of phi using python code are identical
# Lfilt = int(Lmin * len(ts.data))
# print "Lfilt", Lfilt
# # PhiBlur2 = _filterRows(Phi, Lfilt)
# _, PhiBlur2 = preprocessFeatureMat(Phi, Lfilt)
# plt.figure()
# viz.imshowBetter(PhiBlur2)
# plt.colorbar()
# for row in W:
# print row
# print "True pattern sum", Wsum
# print "sum of rx'd pattern", np.sum(W)
# print "W nonzeros: ", np.nonzero(W)
# plt.figure()
# viz.imshowBetter(W)
# plt.colorbar()
# print "python Phi shape, Phi sum, PhiBlur sum", X.shape, np.sum(X), np.sum(Xblur)
print "cpp Phi shape, Phi sum, PhiBlur sum", Phi.shape, np.sum(Phi), np.sum(PhiBlur)
plotOutput = False
# plotOutput = True
if plotOutput:
axSeq, axSim, axPattern = flockviz.plotFFOutput(ts, starts, ends, Phi, W)
# flockviz.plotFFOutput(ts, starts, ends, PhiBlur, W)
axSeq.set_title("cpp output")
# axes = flockviz.plotFFOutput(ts, startIdxs, endIdxs, X, W)
# axes[0].set_title("python output")
plt.show()
