def kernel_fisher_modular(fm_train_dna=traindat,
fm_test_dna=testdat,
label_train_dna=label_traindat,
N=1,
M=4,
pseudo=1e-1,
order=1,
gap=0,
reverse=False,
kargs=[1, False, True]):
from shogun.Features import StringCharFeatures, StringWordFeatures, FKFeatures, DNA
from shogun.Kernel import PolyKernel
from shogun.Distribution import HMM, BW_NORMAL #, MSG_DEBUG
# train HMM for positive class
charfeat = StringCharFeatures(fm_hmm_pos, DNA)
#charfeat.io.set_loglevel(MSG_DEBUG)
hmm_pos_train = StringWordFeatures(charfeat.get_alphabet())
hmm_pos_train.obtain_from_char(charfeat, order - 1, order, gap, reverse)
pos = HMM(hmm_pos_train, N, M, pseudo)
pos.baum_welch_viterbi_train(BW_NORMAL)
# train HMM for negative class
charfeat = StringCharFeatures(fm_hmm_neg, DNA)
hmm_neg_train = StringWordFeatures(charfeat.get_alphabet())
hmm_neg_train.obtain_from_char(charfeat, order - 1, order, gap, reverse)
neg = HMM(hmm_neg_train, N, M, pseudo)
neg.baum_welch_viterbi_train(BW_NORMAL)
# Kernel training data
charfeat = StringCharFeatures(fm_train_dna, DNA)
wordfeats_train = StringWordFeatures(charfeat.get_alphabet())
wordfeats_train.obtain_from_char(charfeat, order - 1, order, gap, reverse)
# Kernel testing data
charfeat = StringCharFeatures(fm_test_dna, DNA)
wordfeats_test = StringWordFeatures(charfeat.get_alphabet())
wordfeats_test.obtain_from_char(charfeat, order - 1, order, gap, reverse)
# get kernel on training data
pos.set_observations(wordfeats_train)
neg.set_observations(wordfeats_train)
feats_train = FKFeatures(10, pos, neg)
feats_train.set_opt_a(-1) #estimate prior
kernel = PolyKernel(feats_train, feats_train, *kargs)
km_train = kernel.get_kernel_matrix()
# get kernel on testing data
pos_clone = HMM(pos)
neg_clone = HMM(neg)
pos_clone.set_observations(wordfeats_test)
neg_clone.set_observations(wordfeats_test)
feats_test = FKFeatures(10, pos_clone, neg_clone)
feats_test.set_a(feats_train.get_a()) #use prior from training data
kernel.init(feats_train, feats_test)
km_test = kernel.get_kernel_matrix()
return km_train, km_test, kernel
def kernel_top_modular(fm_train_dna=traindat,
fm_test_dna=testdat,
label_train_dna=label_traindat,
pseudo=1e-1,
order=1,
gap=0,
reverse=False,
kargs=[1, False, True]):
from shogun.Features import StringCharFeatures, StringWordFeatures, TOPFeatures, DNA
from shogun.Kernel import PolyKernel
from shogun.Distribution import HMM, BW_NORMAL
N = 1 # toy HMM with 1 state
M = 4 # 4 observations -> DNA
# train HMM for positive class
charfeat = StringCharFeatures(fm_hmm_pos, DNA)
hmm_pos_train = StringWordFeatures(charfeat.get_alphabet())
hmm_pos_train.obtain_from_char(charfeat, order - 1, order, gap, reverse)
pos = HMM(hmm_pos_train, N, M, pseudo)
pos.baum_welch_viterbi_train(BW_NORMAL)
# train HMM for negative class
charfeat = StringCharFeatures(fm_hmm_neg, DNA)
hmm_neg_train = StringWordFeatures(charfeat.get_alphabet())
hmm_neg_train.obtain_from_char(charfeat, order - 1, order, gap, reverse)
neg = HMM(hmm_neg_train, N, M, pseudo)
neg.baum_welch_viterbi_train(BW_NORMAL)
# Kernel training data
charfeat = StringCharFeatures(fm_train_dna, DNA)
wordfeats_train = StringWordFeatures(charfeat.get_alphabet())
wordfeats_train.obtain_from_char(charfeat, order - 1, order, gap, reverse)
# Kernel testing data
charfeat = StringCharFeatures(fm_test_dna, DNA)
wordfeats_test = StringWordFeatures(charfeat.get_alphabet())
wordfeats_test.obtain_from_char(charfeat, order - 1, order, gap, reverse)
# get kernel on training data
pos.set_observations(wordfeats_train)
neg.set_observations(wordfeats_train)
feats_train = TOPFeatures(10, pos, neg, False, False)
kernel = PolyKernel(feats_train, feats_train, *kargs)
km_train = kernel.get_kernel_matrix()
# get kernel on testing data
pos_clone = HMM(pos)
neg_clone = HMM(neg)
pos_clone.set_observations(wordfeats_test)
neg_clone.set_observations(wordfeats_test)
feats_test = TOPFeatures(10, pos_clone, neg_clone, False, False)
kernel.init(feats_train, feats_test)
km_test = kernel.get_kernel_matrix()
return km_train, km_test, kernel
def _run_top_fisher():
"""Run Linear Kernel with {Top,Fisher}Features."""
# put some constantness into randomness
Math_init_random(dataop.INIT_RANDOM)
data = dataop.get_cubes(4, 8)
prefix = 'topfk_'
params = {
prefix + 'N': 3,
prefix + 'M': 6,
prefix + 'pseudo': 1e-1,
prefix + 'order': 1,
prefix + 'gap': 0,
prefix + 'reverse': False,
prefix + 'alphabet': 'CUBE',
prefix + 'feature_class': 'string_complex',
prefix + 'feature_type': 'Word',
prefix + 'data_train': numpy.matrix(data['train']),
prefix + 'data_test': numpy.matrix(data['test'])
}
wordfeats = featop.get_features(params[prefix + 'feature_class'],
params[prefix + 'feature_type'], data,
eval(params[prefix + 'alphabet']),
params[prefix + 'order'],
params[prefix + 'gap'],
params[prefix + 'reverse'])
pos_train = HMM(wordfeats['train'], params[prefix + 'N'],
params[prefix + 'M'], params[prefix + 'pseudo'])
pos_train.train()
pos_train.baum_welch_viterbi_train(BW_NORMAL)
neg_train = HMM(wordfeats['train'], params[prefix + 'N'],
params[prefix + 'M'], params[prefix + 'pseudo'])
neg_train.train()
neg_train.baum_welch_viterbi_train(BW_NORMAL)
pos_test = HMM(pos_train)
pos_test.set_observations(wordfeats['test'])
neg_test = HMM(neg_train)
neg_test.set_observations(wordfeats['test'])
feats = {}
feats['train'] = TOPFeatures(10, pos_train, neg_train, False, False)
feats['test'] = TOPFeatures(10, pos_test, neg_test, False, False)
params[prefix + 'name'] = 'TOP'
_compute_top_fisher(feats, params)
feats['train'] = FKFeatures(10, pos_train, neg_train)
feats['train'].set_opt_a(-1) #estimate prior
feats['test'] = FKFeatures(10, pos_test, neg_test)
feats['test'].set_a(feats['train'].get_a()) #use prior from training data
params[prefix + 'name'] = 'FK'
_compute_top_fisher(feats, params)
def distribution_hmm_modular(fm_cube, N, M, pseudo, order, gap, reverse, num_examples): from shogun.Features import StringWordFeatures, StringCharFeatures, CUBE from shogun.Distribution import HMM, BW_NORMAL charfeat=StringCharFeatures(CUBE) charfeat.set_features(fm_cube) feats=StringWordFeatures(charfeat.get_alphabet()) feats.obtain_from_char(charfeat, order-1, order, gap, reverse) hmm=HMM(feats, N, M, pseudo) hmm.train() hmm.baum_welch_viterbi_train(BW_NORMAL) num_examples=feats.get_num_vectors() num_param=hmm.get_num_model_parameters() for i in range(num_examples): for j in range(num_param): hmm.get_log_derivative(j, i) best_path=0 best_path_state=0 for i in range(num_examples): best_path+=hmm.best_path(i) for j in range(N): best_path_state+=hmm.get_best_path_state(i, j) lik_example = hmm.get_log_likelihood() lik_sample = hmm.get_log_likelihood_sample() return lik_example, lik_sample, hmm
def _evaluate_top_fisher(indata, prefix):
feats = {}
wordfeats = util.get_features(indata, prefix)
pos_train = HMM(wordfeats['train'], indata[prefix + 'N'],
indata[prefix + 'M'], indata[prefix + 'pseudo'])
pos_train.train()
pos_train.baum_welch_viterbi_train(BW_NORMAL)
neg_train = HMM(wordfeats['train'], indata[prefix + 'N'],
indata[prefix + 'M'], indata[prefix + 'pseudo'])
neg_train.train()
neg_train.baum_welch_viterbi_train(BW_NORMAL)
pos_test = HMM(pos_train)
pos_test.set_observations(wordfeats['test'])
neg_test = HMM(neg_train)
neg_test.set_observations(wordfeats['test'])
if indata[prefix + 'name'] == 'TOP':
feats['train'] = TOPFeatures(10, pos_train, neg_train, False, False)
feats['test'] = TOPFeatures(10, pos_test, neg_test, False, False)
else:
feats['train'] = FKFeatures(10, pos_train, neg_train)
feats['train'].set_opt_a(-1) #estimate prior
feats['test'] = FKFeatures(10, pos_test, neg_test)
feats['test'].set_a(
feats['train'].get_a()) #use prior from training data
prefix = 'kernel_'
args = util.get_args(indata, prefix)
kernel = PolyKernel(feats['train'], feats['train'], *args)
# kernel=PolyKernel(*args)
# kernel.init(feats['train'], feats['train'])
km_train = max(
abs(indata[prefix + 'matrix_train'] - kernel.get_kernel_matrix()).flat)
kernel.init(feats['train'], feats['test'])
km_test = max(
abs(indata[prefix + 'matrix_test'] - kernel.get_kernel_matrix()).flat)
return util.check_accuracy(indata[prefix + 'accuracy'],
km_train=km_train,
km_test=km_test)