def construct_measured_vector(self, train_subset):
num_seqs = num_rows(train_subset)
measured_vals = np.matlib.zeros((1, num_seqs), dtype=np.float)
for ctr in range(num_seqs):
seq = str(train_subset[ctr, :])
measured_vals[0, ctr] = self.sequence_matrix.seq_to_measurement[seq]
return measured_vals
def update_dictionaries(self):
"""
When the training data is changed (features or examples added/removed) the dictionary keys need to be updated
# using a dictionary to look up measured values an inequalities means that sequences must be unique in training data
# however, this doesn't seem to be an issue and it is the simplest way to associate sequences with the measured
# values and inequalities across shuffling and CV splits
:return:
"""
self.seq_to_measurement.clear()
self.seq_to_inequality.clear()
for ctr in range(num_rows(self.training_data)):
seq = str(self.training_data[ctr, :])
self.seq_to_inequality[seq] = self.inequalities[ctr]
self.seq_to_measurement[seq] = self.measured_values[ctr, 0]
def distance(self, log_lambdas):
print('calculating distance')
dist = 0.0
count = 0
for fold_num in range(self.num_cv_folds):
train = self.cv_train_data[fold_num]
blind = self.cv_blind_data[fold_num]
if fold_num == 0:
print('starting solve_x for fold ', fold_num)
self.solve_x(train, log_lambdas, fold_num)
dist += self.x_distance(blind, fold_num)[0, 0]
# num rows is num elements in blind set
count += num_rows(blind)
dist /= float(count)
print('dist: ' + str(dist) + '\n')
print('\n')
sys.stdout.flush()
return dist
def add_pair_coeffs(self, pairs):
num_pairs = len(pairs)
# extend columns for pair coeff binary features
coeff_feature_matrix = np.matlib.zeros((num_rows(self.training_data), num_pairs), dtype=np.float)
# TODO: 20 should be dynamic alphabet size
base_index = 1 + self.seq_length * 20
feature_index_to_pair = [None] * (base_index + len(pairs))
for offset, pair in enumerate(pairs):
feature_index_to_pair[base_index + offset] = pair
seq_index = 0
for seq in self.training_data:
for offset, pair in enumerate(pairs):
if seq[0, pair[0]] == 1 and seq[0, pair[1]] == 1:
coeff_feature_matrix[seq_index, offset] = 1.0
seq_index += 1
self.training_data = np.concatenate((self.training_data, coeff_feature_matrix), axis=1)
self.update_dictionaries()
def create_pair_training_matrix(self, pairs):
"""
The 0th pair will be the 1st column of features due to bias
:param pairs:
:return:
"""
num_pairs = len(pairs)
# feature_index_to_pair = [None] * (1 + len(pairs))
# for offset, pair in enumerate(pairs):
# feature_index_to_pair[1 + offset] = pair
coeff_feature_matrix = np.matlib.zeros((num_rows(self.training_data), 1 + num_pairs), dtype=np.float)
seq_index = 0
# can't use np.ndenumerate() because that is by element, not row
for seq in self.training_data:
# pair is tuple of index locations of binary pair feature
coeff_feature_matrix[seq_index, 0] = 1.0
for offset, pair in enumerate(pairs):
if seq[0, pair[0]] == 1 and seq[0, pair[1]] == 1: # does seq contain the pair
coeff_feature_matrix[seq_index, offset + 1] = 1.0 # plus 1 to account for bias being first column
seq_index += 1
return coeff_feature_matrix