def computeGibbsProbabilities(crf,
blocksFunction,
choiceFunction,
xs,
samples=2000):
"""
Empirically estimate the probabilities of various tag sequences. You
should count the number of labelings over many samples from the
Gibbs sampler.
@param xs list string - Observation sequence
@param samples int - Number of epochs to produce samples
@return Counter - A counter of tag sequences with an empirical
estimate of their probabilities.
Example output:
Counter({
('-FEAT-', '-SIZE-', '-SIZE-'): 0.379,
('-SIZE-', '-SIZE-', '-SIZE-'): 0.189,
('-FEAT-', '-SIZE-', '-FEAT-'): 0.166,
('-SIZE-', '-SIZE-', '-FEAT-'): 0.135,
('-FEAT-', '-FEAT-', '-SIZE-'): 0.053,
('-SIZE-', '-FEAT-', '-SIZE-'): 0.052,
('-FEAT-', '-FEAT-', '-FEAT-'): 0.018,
('-SIZE-', '-FEAT-', '-FEAT-'): 0.008})
Possibly useful:
* Counter
* gibbsRun
"""
# BEGIN_YOUR_CODE (around 2 lines of code expected)
particles = Counter(
gibbsRun(crf, blocksFunction, choiceFunction, xs, samples))
Counters.scale(particles, 1.0 / sum(particles.values()))
return particles
def computeGibbsProbabilities(crf, blocksFunction, choiceFunction, xs, samples = 2000):
"""
Empirically estimate the probabilities of various tag sequences. You
should count the number of labelings over many samples from the
Gibbs sampler.
@param xs list string - Observation sequence
@param samples int - Number of epochs to produce samples
@return Counter - A counter of tag sequences with an empirical
estimate of their probabilities.
Example output:
Counter({
('-FEAT-', '-SIZE-', '-SIZE-'): 0.379,
('-SIZE-', '-SIZE-', '-SIZE-'): 0.189,
('-FEAT-', '-SIZE-', '-FEAT-'): 0.166,
('-SIZE-', '-SIZE-', '-FEAT-'): 0.135,
('-FEAT-', '-FEAT-', '-SIZE-'): 0.053,
('-SIZE-', '-FEAT-', '-SIZE-'): 0.052,
('-FEAT-', '-FEAT-', '-FEAT-'): 0.018,
('-SIZE-', '-FEAT-', '-FEAT-'): 0.008})
Possibly useful:
* Counter
* gibbsRun
"""
# BEGIN_YOUR_CODE (around 2 lines of code expected)
particles = Counter(gibbsRun(crf,blocksFunction,choiceFunction,xs,samples))
Counters.scale(particles, 1.0/sum(particles.values()))
return particles
def part2a_0():
"""Check that you have all the features we expect"""
xs = exampleInput
phi = Counter({('-BEGIN-', '-FEAT-'): 1.0, ('-FEAT-', 'Beautiful'): 1.0, ('-FEAT-', 'PREV:-BEGIN-'): 1.0, ('-FEAT-', 'NEXT:2'): 1.0, ('-FEAT-', '-CAPITALIZED-'): 1.0, ('-FEAT-', '-POST-CAPITALIZED-'): 0.0})
phi_ = submission.nerFeatureFunction(0, '-BEGIN-', '-FEAT-', xs)
grader.requireIsTrue( Counters.approximateEquals(phi, phi_) )
phi = Counter({('-FEAT-', '-SIZE-'): 1.0, ('-SIZE-', 'PREV:Beautiful'): 1.0, ('-SIZE-', 'NEXT:bedroom'): 1.0, ('-SIZE-', '-PRE-CAPITALIZED-'): 1.0, ('-SIZE-', '2'): 1.0, ('-SIZE-', '-POST-CAPITALIZED-'): 0.0, ('-SIZE-', '-CAPITALIZED-'): 0.0})
phi_ = submission.nerFeatureFunction(1, '-FEAT-', '-SIZE-', xs)
grader.requireIsTrue( Counters.approximateEquals(phi, phi_) )
phi = Counter({('-SIZE-', '-SIZE-'): 1.0, ('-SIZE-', 'PREV:2'): 1.0, ('-SIZE-', 'bedroom'): 1.0, ('-SIZE-', 'NEXT:-END-'): 1.0, ('-SIZE-', '-CAPITALIZED-'): 0.0, ('-SIZE-', '-PRE-CAPITALIZED-'): 0.0})
phi_ = submission.nerFeatureFunction(2, '-SIZE-', '-SIZE-', xs)
grader.requireIsTrue( Counters.approximateEquals(phi, phi_) )
def part1b_3():
"""Check that values match our reference"""
xs = exampleInput
backward = [
Counter({'-SIZE-': 0.564, '-FEAT-': 0.435}),
Counter({'-SIZE-': 0.567, '-FEAT-': 0.432}),
Counter({'-FEAT-': 0.5, '-SIZE-': 0.5})]
backward_ = submission.computeBackward(simpleCRF, xs)
for vec, vec_ in zip( backward, backward_):
grader.requireIsTrue( Counters.approximateEquals( vec, vec_ ) )
def G(self, t, y_, y, xs):
r"""
Computes one of the potentials in the CRF.
@param t int - index in the observation sequence, 0-based.
@param y_ string - value of of tag at time t-1 (y_{t-1}),
@param y string - value of of tag at time t (y_{t}),
@param xs list string - The full observation seqeunce.
@return double - G_t(y_{t-1}, y_t ; x, \theta)
"""
return math.exp( Counters.dot( self.parameters, self.featureFunction(t, y_, y, xs) ) )
def G(self, t, y_, y, xs):
r"""
Computes one of the potentials in the CRF.
@param t int - index in the observation sequence, 0-based.
@param y_ string - value of of tag at time t-1 (y_{t-1}),
@param y string - value of of tag at time t (y_{t}),
@param xs list string - The full observation seqeunce.
@return double - G_t(y_{t-1}, y_t ; x, \theta)
"""
return math.exp( Counters.dot( self.parameters, self.featureFunction(t, y_, y, xs) ) )
def part1c_1():
"""Check that values match our reference"""
xs = exampleInput
T = [ Counter({('-BEGIN-', '-FEAT-'): 0.561, ('-BEGIN-', '-SIZE-'): 0.439}),
Counter({('-FEAT-', '-SIZE-'): 0.463, ('-SIZE-', '-SIZE-'): 0.343,
('-SIZE-', '-FEAT-'): 0.096, ('-FEAT-', '-FEAT-'): 0.096}),
Counter({('-SIZE-', '-SIZE-'): 0.590, ('-SIZE-', '-FEAT-'): 0.217,
('-FEAT-', '-SIZE-'): 0.151, ('-FEAT-', '-FEAT-'): 0.041})
]
T_ = submission.computeEdgeMarginals(simpleCRF, xs)
for t, t_ in zip(T, T_):
grader.requireIsTrue( Counters.approximateEquals(t, t_) )
def part2a_0():
"""Check that you have all the features we expect"""
xs = exampleInput
phi = Counter({
('-BEGIN-', '-FEAT-'): 1.0,
('-FEAT-', 'Beautiful'): 1.0,
('-FEAT-', 'PREV:-BEGIN-'): 1.0,
('-FEAT-', 'NEXT:2'): 1.0,
('-FEAT-', '-CAPITALIZED-'): 1.0,
('-FEAT-', '-POST-CAPITALIZED-'): 0.0
})
phi_ = submission.nerFeatureFunction(0, '-BEGIN-', '-FEAT-', xs)
grader.requireIsTrue(Counters.approximateEquals(phi, phi_))
phi = Counter({
('-FEAT-', '-SIZE-'): 1.0,
('-SIZE-', 'PREV:Beautiful'): 1.0,
('-SIZE-', 'NEXT:bedroom'): 1.0,
('-SIZE-', '-PRE-CAPITALIZED-'): 1.0,
('-SIZE-', '2'): 1.0,
('-SIZE-', '-POST-CAPITALIZED-'): 0.0,
('-SIZE-', '-CAPITALIZED-'): 0.0
})
phi_ = submission.nerFeatureFunction(1, '-FEAT-', '-SIZE-', xs)
grader.requireIsTrue(Counters.approximateEquals(phi, phi_))
phi = Counter({
('-SIZE-', '-SIZE-'): 1.0,
('-SIZE-', 'PREV:2'): 1.0,
('-SIZE-', 'bedroom'): 1.0,
('-SIZE-', 'NEXT:-END-'): 1.0,
('-SIZE-', '-CAPITALIZED-'): 0.0,
('-SIZE-', '-PRE-CAPITALIZED-'): 0.0
})
phi_ = submission.nerFeatureFunction(2, '-SIZE-', '-SIZE-', xs)
grader.requireIsTrue(Counters.approximateEquals(phi, phi_))
def part1b_3():
"""Check that values match our reference"""
xs = exampleInput
backward = [
Counter({
'-SIZE-': 0.564,
'-FEAT-': 0.435
}),
Counter({
'-SIZE-': 0.567,
'-FEAT-': 0.432
}),
Counter({
'-FEAT-': 0.5,
'-SIZE-': 0.5
})
]
backward_ = submission.computeBackward(simpleCRF, xs)
for vec, vec_ in zip(backward, backward_):
grader.requireIsTrue(Counters.approximateEquals(vec, vec_))
def part1c_1():
"""Check that values match our reference"""
xs = exampleInput
T = [
Counter({
('-BEGIN-', '-FEAT-'): 0.561,
('-BEGIN-', '-SIZE-'): 0.439
}),
Counter({
('-FEAT-', '-SIZE-'): 0.463,
('-SIZE-', '-SIZE-'): 0.343,
('-SIZE-', '-FEAT-'): 0.096,
('-FEAT-', '-FEAT-'): 0.096
}),
Counter({
('-SIZE-', '-SIZE-'): 0.590,
('-SIZE-', '-FEAT-'): 0.217,
('-FEAT-', '-SIZE-'): 0.151,
('-FEAT-', '-FEAT-'): 0.041
})
]
T_ = submission.computeEdgeMarginals(simpleCRF, xs)
for t, t_ in zip(T, T_):
grader.requireIsTrue(Counters.approximateEquals(t, t_))