def test_2():
n_features = 3
length = 32
for n_states in [4]:
t1 = np.random.randn(length, n_features)
means = np.random.randn(n_states, n_features)
vars = np.random.rand(n_states, n_features)
transmat = np.random.rand(n_states, n_states)
transmat = transmat / np.sum(transmat, axis=1)[:, None]
startprob = np.random.rand(n_states)
startprob = startprob / np.sum(startprob)
chmm = GaussianHMMCPUImpl(n_states, n_features)
chmm._sequences = [t1]
pyhmm = GaussianHMM(n_components=n_states,
init_params='',
params='',
covariance_type='diag')
chmm.means_ = means.astype(np.float32)
chmm.vars_ = vars.astype(np.float32)
chmm.transmat_ = transmat.astype(np.float32)
chmm.startprob_ = startprob.astype(np.float32)
clogprob, cstats = chmm.do_estep()
pyhmm.means_ = means
pyhmm.covars_ = vars
pyhmm.transmat_ = transmat
pyhmm.startprob_ = startprob
framelogprob = pyhmm._compute_log_likelihood(t1)
fwdlattice = pyhmm._do_forward_pass(framelogprob)[1]
bwdlattice = pyhmm._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
stats = pyhmm._initialize_sufficient_statistics()
pyhmm._accumulate_sufficient_statistics(stats, t1, framelogprob,
posteriors, fwdlattice,
bwdlattice, 'stmc')
yield lambda: np.testing.assert_array_almost_equal(
stats['trans'], cstats['trans'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(
stats['post'], cstats['post'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(
stats['obs'], cstats['obs'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(
stats['obs**2'], cstats['obs**2'], decimal=3)
class _SklearnGaussianHMMCPUImpl(object):
def __init__(self, n_states, n_features):
from sklearn.hmm import GaussianHMM
self.impl = GaussianHMM(n_states, params='stmc')
self._sequences = None
self.means_ = None
self.vars_ = None
self.transmat_ = None
self.startprob_ = None
def do_estep(self):
from sklearn.utils.extmath import logsumexp
self.impl.means_ = self.means_.astype(np.double)
self.impl.covars_ = self.vars_.astype(np.double)
self.impl.transmat_ = self.transmat_.astype(np.double)
self.impl.startprob_ = self.startprob_.astype(np.double)
stats = self.impl._initialize_sufficient_statistics()
curr_logprob = 0
for seq in self._sequences:
seq = seq.astype(np.double)
framelogprob = self.impl._compute_log_likelihood(seq)
lpr, fwdlattice = self.impl._do_forward_pass(framelogprob)
bwdlattice = self.impl._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
curr_logprob += lpr
self.impl._accumulate_sufficient_statistics(
stats, seq, framelogprob, posteriors, fwdlattice,
bwdlattice, self.impl.params)
return curr_logprob, stats
def do_viterbi(self):
logprob = 0
state_sequences = []
for obs in self._sequences:
lpr, ss = self.impl._decode_viterbi(obs)
logprob += lpr
state_sequences.append(ss)
return logprob, state_sequences
class _SklearnGaussianHMMCPUImpl(object):
def __init__(self, n_states, n_features):
from sklearn.hmm import GaussianHMM
self.impl = GaussianHMM(n_states, params='stmc')
self._sequences = None
self.means_ = None
self.vars_ = None
self.transmat_ = None
self.startprob_ = None
def do_estep(self):
from sklearn.utils.extmath import logsumexp
self.impl.means_ = self.means_.astype(np.double)
self.impl.covars_ = self.vars_.astype(np.double)
self.impl.transmat_ = self.transmat_.astype(np.double)
self.impl.startprob_ = self.startprob_.astype(np.double)
stats = self.impl._initialize_sufficient_statistics()
curr_logprob = 0
for seq in self._sequences:
seq = seq.astype(np.double)
framelogprob = self.impl._compute_log_likelihood(seq)
lpr, fwdlattice = self.impl._do_forward_pass(framelogprob)
bwdlattice = self.impl._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
curr_logprob += lpr
self.impl._accumulate_sufficient_statistics(
stats, seq, framelogprob, posteriors, fwdlattice, bwdlattice,
self.impl.params)
return curr_logprob, stats
def do_viterbi(self):
logprob = 0
state_sequences = []
for obs in self._sequences:
lpr, ss = self.impl._decode_viterbi(obs)
logprob += lpr
state_sequences.append(ss)
return logprob, state_sequences
def test_2():
n_features = 3
length = 32
for n_states in [4]:
t1 = np.random.randn(length, n_features)
means = np.random.randn(n_states, n_features)
vars = np.random.rand(n_states, n_features)
transmat = np.random.rand(n_states, n_states)
transmat = transmat / np.sum(transmat, axis=1)[:, None]
startprob = np.random.rand(n_states)
startprob = startprob / np.sum(startprob)
chmm = GaussianHMMCPUImpl(n_states, n_features)
chmm._sequences = [t1]
pyhmm = GaussianHMM(n_components=n_states, init_params='', params='', covariance_type='diag')
chmm.means_ = means.astype(np.float32)
chmm.vars_ = vars.astype(np.float32)
chmm.transmat_ = transmat.astype(np.float32)
chmm.startprob_ = startprob.astype(np.float32)
clogprob, cstats = chmm.do_estep()
pyhmm.means_ = means
pyhmm.covars_ = vars
pyhmm.transmat_ = transmat
pyhmm.startprob_ = startprob
framelogprob = pyhmm._compute_log_likelihood(t1)
fwdlattice = pyhmm._do_forward_pass(framelogprob)[1]
bwdlattice = pyhmm._do_backward_pass(framelogprob)
gamma = fwdlattice + bwdlattice
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
stats = pyhmm._initialize_sufficient_statistics()
pyhmm._accumulate_sufficient_statistics(
stats, t1, framelogprob, posteriors, fwdlattice,
bwdlattice, 'stmc')
yield lambda: np.testing.assert_array_almost_equal(stats['trans'], cstats['trans'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(stats['post'], cstats['post'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(stats['obs'], cstats['obs'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(stats['obs**2'], cstats['obs**2'], decimal=3)
def hmm(samples):
model = GaussianHMM(n_components=3)
samples = samples.dropna()
idx = samples.index
if samples.values.ndim < 2:
#import pdb; pdb.set_trace()
m = samples.values.shape
samples = samples.values.reshape(m[0],1)
model.fit([samples])
#_, states = model.decode(samples, algorithm='map')
framelogprob = model._compute_log_likelihood(samples)
logprob, fwdlattice = model._do_forward_pass(framelogprob)
n, _ = model.means_.shape
frame = pd.DataFrame(
framelogprob, index=idx, columns=map(lambda x: "frame_"+str(x), range(n)) )
forward = pd.DataFrame(
fwdlattice, index=idx, columns=map(lambda x: "forward_"+str(x), range(n)) )
#import pdb; pdb.set_trace()
predict = pd.DataFrame(
(fwdlattice-framelogprob)[1:, :], index=idx[:-1], columns=map(lambda x: "predict_"+str(x), range(n)))
import pdb; pdb.set_trace()
return model, frame.join(forward)
def test_2():
np.random.seed(42)
n_features = 32
length = 20
#for n_states in [3, 4, 5, 7, 8, 9, 15, 16, 17, 31, 32]:
for n_states in [8]:
t1 = np.random.randn(length, n_features)
means = np.random.randn(n_states, n_features)
vars = np.random.rand(n_states, n_features)
transmat = np.random.rand(n_states, n_states)
transmat = transmat / np.sum(transmat, axis=1)[:, None]
startprob = np.random.rand(n_states)
startprob = startprob / np.sum(startprob)
cuhmm = GaussianHMMCUDAImpl(n_states, n_features)
cuhmm._sequences = [t1]
pyhmm = GaussianHMM(n_components=n_states, init_params='', params='', covariance_type='diag')
cuhmm.means_ = means
cuhmm.vars_ = vars
cuhmm.transmat_ = transmat
cuhmm.startprob_ = startprob
logprob, custats = cuhmm.do_estep()
pyhmm.means_ = means
pyhmm.covars_ = vars
pyhmm.transmat_ = transmat
pyhmm.startprob_ = startprob
pyhmm._initialize_sufficient_statistics()
framelogprob = pyhmm._compute_log_likelihood(t1)
cuframelogprob = cuhmm._get_framelogprob()
yield lambda: np.testing.assert_array_almost_equal(framelogprob, cuframelogprob, decimal=3)
fwdlattice = pyhmm._do_forward_pass(framelogprob)[1]
cufwdlattice = cuhmm._get_fwdlattice()
yield lambda: np.testing.assert_array_almost_equal(fwdlattice, cufwdlattice, decimal=3)
bwdlattice = pyhmm._do_backward_pass(framelogprob)
cubwdlattice = cuhmm._get_bwdlattice()
yield lambda: np.testing.assert_array_almost_equal(bwdlattice, cubwdlattice, decimal=3)
gamma = fwdlattice + bwdlattice
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
cuposteriors = cuhmm._get_posteriors()
yield lambda: np.testing.assert_array_almost_equal(posteriors, cuposteriors, decimal=3)
stats = pyhmm._initialize_sufficient_statistics()
pyhmm._accumulate_sufficient_statistics(
stats, t1, framelogprob, posteriors, fwdlattice,
bwdlattice, 'stmc')
print 'ref transcounts'
print transitioncounts(cufwdlattice, cubwdlattice, cuframelogprob, np.log(transmat))
print 'cutranscounts'
print custats['trans']
yield lambda: np.testing.assert_array_almost_equal(stats['trans'], custats['trans'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(stats['post'], custats['post'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(stats['obs'], custats['obs'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(stats['obs**2'], custats['obs**2'], decimal=3)
def test_2():
np.random.seed(42)
n_features = 32
length = 20
#for n_states in [3, 4, 5, 7, 8, 9, 15, 16, 17, 31, 32]:
for n_states in [8]:
t1 = np.random.randn(length, n_features)
means = np.random.randn(n_states, n_features)
vars = np.random.rand(n_states, n_features)
transmat = np.random.rand(n_states, n_states)
transmat = transmat / np.sum(transmat, axis=1)[:, None]
startprob = np.random.rand(n_states)
startprob = startprob / np.sum(startprob)
cuhmm = GaussianHMMCUDAImpl(n_states, n_features)
cuhmm._sequences = [t1]
pyhmm = GaussianHMM(n_components=n_states,
init_params='',
params='',
covariance_type='diag')
cuhmm.means_ = means
cuhmm.vars_ = vars
cuhmm.transmat_ = transmat
cuhmm.startprob_ = startprob
logprob, custats = cuhmm.do_estep()
pyhmm.means_ = means
pyhmm.covars_ = vars
pyhmm.transmat_ = transmat
pyhmm.startprob_ = startprob
pyhmm._initialize_sufficient_statistics()
framelogprob = pyhmm._compute_log_likelihood(t1)
cuframelogprob = cuhmm._get_framelogprob()
yield lambda: np.testing.assert_array_almost_equal(
framelogprob, cuframelogprob, decimal=3)
fwdlattice = pyhmm._do_forward_pass(framelogprob)[1]
cufwdlattice = cuhmm._get_fwdlattice()
yield lambda: np.testing.assert_array_almost_equal(
fwdlattice, cufwdlattice, decimal=3)
bwdlattice = pyhmm._do_backward_pass(framelogprob)
cubwdlattice = cuhmm._get_bwdlattice()
yield lambda: np.testing.assert_array_almost_equal(
bwdlattice, cubwdlattice, decimal=3)
gamma = fwdlattice + bwdlattice
posteriors = np.exp(gamma.T - logsumexp(gamma, axis=1)).T
cuposteriors = cuhmm._get_posteriors()
yield lambda: np.testing.assert_array_almost_equal(
posteriors, cuposteriors, decimal=3)
stats = pyhmm._initialize_sufficient_statistics()
pyhmm._accumulate_sufficient_statistics(stats, t1, framelogprob,
posteriors, fwdlattice,
bwdlattice, 'stmc')
print 'ref transcounts'
print transitioncounts(cufwdlattice, cubwdlattice, cuframelogprob,
np.log(transmat))
print 'cutranscounts'
print custats['trans']
yield lambda: np.testing.assert_array_almost_equal(
stats['trans'], custats['trans'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(
stats['post'], custats['post'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(
stats['obs'], custats['obs'], decimal=3)
yield lambda: np.testing.assert_array_almost_equal(
stats['obs**2'], custats['obs**2'], decimal=3)
