def train(self):
'''
Estimate model parameters from data.
'''
self.model.init_training()
# Initialize frequency accumulators
transmission_acc = numpy.zeros((self.num_states, self.num_states))
emission_acc = numpy.zeros(self.num_states * self.num_features *
self.num_plif_nodes)
# Go through training data, counting state transitions and feature emissions
assert (self.model.get_labels().get_num_labels() ==
self.model.get_features().get_num_vectors())
for i in xrange(self.model.get_labels().get_num_labels()):
# Update transmission and emission counts internally in the model
self.model.get_joint_feature_vector(i, i)
# Update accumulators
transmission_acc += self.model.get_transmission_weights()
emission_acc += self.model.get_emission_weights()
# Compute model parameters from the accumulators
state_model = TwoStateModel()
w = state_model.weights_to_vector(transmission_acc, emission_acc,
self.num_features,
self.num_plif_nodes)
## print w
## print '# transmission & emission weights = ',
## numpy.prod(transmission_acc.shape)+emission_acc.shape[0]
## print '# params = ', num_free_states*(num_free_states + num_features*num_plif_nodes)
# Smooth counts equal to zero to avoid zero-division warnings
w[w == 0.0] = 1e-100
# Compute transition scores as transition frequencies, i.e. estimate
# probabilities using the ML rule
self.transition_scores[0] = numpy.log(w[0] / w[0:2].sum())
self.transition_scores[1] = numpy.log(w[1] / w[0:2].sum())
self.transition_scores[2] = numpy.log(w[2] / w[2:4].sum())
self.transition_scores[3] = numpy.log(w[3] / w[2:4].sum())
# Compute emission or feature scores as frequencies of occurrence
# Offset given by the number of transition_scores, used to index w
offset = self.transition_scores.shape[0]
self.feature_scores = numpy.zeros(
(self.num_features, self.num_free_states, self.num_plif_nodes))
for s in xrange(self.num_free_states):
state_idx = s * self.num_features * self.num_plif_nodes
for f in xrange(self.num_features):
feat_idx = f * self.num_plif_nodes
idx = offset + state_idx + feat_idx
self.feature_scores[f, s, :] = numpy.log(
w[idx:idx + self.num_plif_nodes] /
w[idx:idx + self.num_plif_nodes].sum())
def structure_plif_hmsvm_bmrm (num_examples, example_length, num_features, num_noise_features): from modshogun import RealMatrixFeatures, TwoStateModel, DualLibQPBMSOSVM, StructuredAccuracy model = TwoStateModel.simulate_data(num_examples, example_length, num_features, num_noise_features) sosvm = DualLibQPBMSOSVM(model, model.get_labels(), 5000.0) sosvm.train() #print sosvm.get_w() predicted = sosvm.apply(model.get_features()) evaluator = StructuredAccuracy() acc = evaluator.evaluate(predicted, model.get_labels())
def structure_plif_hmsvm_mosek (num_examples, example_length, num_features, num_noise_features):
from modshogun import RealMatrixFeatures, TwoStateModel, StructuredAccuracy
try:
from modshogun import PrimalMosekSOSVM
except ImportError:
print("Mosek not available")
return
model = TwoStateModel.simulate_data(num_examples, example_length, num_features, num_noise_features)
sosvm = PrimalMosekSOSVM(model, model.get_labels())
sosvm.train()
#print(sosvm.get_w())
predicted = sosvm.apply(model.get_features())
evaluator = StructuredAccuracy()
acc = evaluator.evaluate(predicted, model.get_labels())
def structure_plif_hmsvm_bmrm (num_examples, example_length, num_features, num_noise_features):
from modshogun import RealMatrixFeatures, TwoStateModel, StructuredAccuracy
try:
from modshogun import DualLibQPBMSOSVM
except ImportError:
print("DualLibQPBMSOSVM not available")
exit(0)
model = TwoStateModel.simulate_data(num_examples, example_length, num_features, num_noise_features)
sosvm = DualLibQPBMSOSVM(model, model.get_labels(), 5000.0)
sosvm.set_store_train_info(False)
sosvm.train()
#print sosvm.get_w()
predicted = sosvm.apply(model.get_features())
evaluator = StructuredAccuracy()
acc = evaluator.evaluate(predicted, model.get_labels())
def structure_plif_hmsvm_mosek(num_examples, example_length, num_features,
num_noise_features):
from modshogun import RealMatrixFeatures, TwoStateModel, StructuredAccuracy
try:
from modshogun import PrimalMosekSOSVM
except ImportError:
print("Mosek not available")
return
model = TwoStateModel.simulate_data(num_examples, example_length,
num_features, num_noise_features)
sosvm = PrimalMosekSOSVM(model, model.get_labels())
sosvm.train()
#print(sosvm.get_w())
predicted = sosvm.apply(model.get_features())
evaluator = StructuredAccuracy()
acc = evaluator.evaluate(predicted, model.get_labels())
