def run(options):
restaurant = lp.ReinstantiatingCompactRestaurant()
nodeManager = lp.SimpleNodeManager(restaurant.getFactory())
discounts = lp.VectorDouble(DISCOUNTS)
parameters = lp.SimpleParameters(discounts, options.alpha)
seq = lp.VectorInt()
lp.pushIntFileToVec(options.train_file, seq)
print >> sys.stderr, "Train seq length: %i" % (seq.size(), )
# initialize model
model = lp.HPYPModel(seq, nodeManager, restaurant, parameters, NUM_TYPES)
#insert training observations into model using particle filter
model.computeLosses(0, seq.size())
# add test observations to underlying sequence
testOffset = seq.size()
lp.pushIntFileToVec(options.test_file, seq)
print >> sys.stderr, "Test seq length: %i" % (seq.size() - testOffset, )
if options.prediction == 2:
predictMode = lp.HPYPModel.BELOW
elif options.prediction == 1:
predictMode = lp.HPYPModel.FRAGMENT
else:
predictMode = lp.HPYPModel.ABOVE
if options.inference == 1:
for i in xrange(BURN_IN_SAMPLES):
print >> sys.stderr, "Burn in iteration %i" % (i, )
model.runGibbsSampler()
if options.prediction != 3:
loss = float(
lp.prob2loss(
model.predictSequence(testOffset, seq.size(), predictMode)))
else:
loss = float(np.mean(model.computeLosses(testOffset, seq.size())))
if options.inference == 2 and options.prediction != 3:
losses = np.zeros((PREDICT_SAMPLES, seq.size() - testOffset))
for i in xrange(BURN_IN_SAMPLES):
print >> sys.stderr, "Burn in iteration %i" % (i, )
model.runGibbsSampler()
for i in xrange(PREDICT_SAMPLES):
print >> sys.stderr, "Prediction iteration %i" % (i, )
model.runGibbsSampler()
losses[i, :] = model.predictSequence(testOffset, seq.size(),
predictMode)
loss = float(np.mean(-np.log2(np.mean(losses, 0))))
print loss
# make sure destructors are called in correct order
del model
del nodeManager
def buildModel(fn):
"""Build a byte-level SM model from the given file."""
global seq, model, nodeManager, parameters, restaurant
#restaurant = libplump.SimpleFullRestaurant()
#restaurant = libplump.HistogramRestaurant()
#restaurant = libplump.KneserNeyRestaurant()
restaurant = libplump.ReinstantiatingCompactRestaurant()
#restaurant = libplump.StirlingCompactRestaurant()
nodeManager = libplump.SimpleNodeManager(restaurant.getFactory())
parameters = libplump.SimpleParameters(DISCOUNTS, CONCENTRATION)
seq = libplump.VectorInt()
libplump.pushCharFileToVec(fn, seq)
numTypes = 256
model = libplump.HPYPModel(seq, nodeManager, restaurant, parameters,
numTypes)
model.computeLosses(0, seq.size())
import libplump
restaurant = libplump.SimpleFullRestaurant()
#restaurant = libplump.FractionalRestaurant()
#restaurant = libplump.HistogramRestaurant()
#restaurant = libplump.KneserNeyRestaurant()
#restaurant = libplump.ReinstantiatingCompactRestaurant()
#restaurant = libplump.StirlingCompactRestaurant()
nodeManager = libplump.SimpleNodeManager(restaurant.getFactory())
parameters = libplump.SimpleParameters()
#seq = libplump.vectori(range(10))
seq = libplump.VectorInt([0, 1, 2, 1, 2])
#seq = libplump.VectorInt(map(ord,'oacac'))
#numTypes = max(seq)
numTypes = 3
model = libplump.HPYPModel(seq, nodeManager, restaurant, parameters, numTypes)
print model.computeLosses(0, len(seq))
for i in range(seq.size()):
print model.toString()
model.runGibbsSampler()
print "Predictions after training:"
for i in range(len(seq)):
#print model.predict(0,i,i)
dist = model.predictiveDistribution(0, i)
print dist, sum(dist)
# save model to file