def eval_global(self, measures, test_sets):
histograms, testset_pdfs = {}, {}
dist_vals, conf_mats, results = {}, {}, {}
for m in measures:
dist_vals[m], conf_mats[m], results[m] = {}, {}, {}
for c in self.cl:
dist_vals[m][c], conf_mats[m][c], results[m][c] = {}, {}, {}
for o in self.obs:
results[m][c][o] = {'right': 0, 'wrong': 0}
for c in self.cl:
histograms[c], testset_pdfs[c] = {}, {}
for m in measures:
for c_pred in self.cl:
conf_mats[m][c][c_pred] = 0
for file_path in test_sets[c]:
histograms[c][file_path], testset_pdfs[c][
file_path] = {}, {}
dist_vals[m][c][file_path] = {}
data = np.load(file_path)
data_A, data_B = tools.extract_individual_data(data)
obs_data = tools.compute_observables(data_A, data_B)
for o in self.obs:
dist_vals[m][c][file_path][o] = {}
histograms[c][file_path][
o] = tools.initialize_histogram(o)
histograms[c][file_path][o] = tools.compute_histogram(
o, obs_data[o])
testset_pdfs[c][file_path][o] = tools.compute_pdf(
o, histograms[c][file_path][o])
for c_query in self.cl:
dist_vals[m][c][file_path][o][
c_query] = self.get_dists(
m, testset_pdfs[c][file_path][o],
histograms[c][file_path][o],
self.pdfs[o][c_query])
c_pred = min(dist_vals[m][c][file_path][o].items(),
key=operator.itemgetter(1))[0]
conf_mats[m][c][c_pred] += 1
if c_pred == c:
results[m][c][o]['right'] += 1
else:
results[m][c][o]['wrong'] += 1
return results
def train(self, train_sets):
histograms, self.joint_pdfs, jaccard_dist = {}, {}, {}
# initialize empty histograms
for o in self.obs:
histograms[o], self.pdfs[o] = {}, {}
for c in self.cl:
histograms[o][c] = tools.initialize_histogram(o)
# compute histograms for each classes
obs_data_cum = {}
for c in self.cl:
obs_data_cum[c] = {}
for o in self.obs:
obs_data_cum[c][o] = []
for file_path in train_sets[c]:
data = np.load(file_path)
data_A, data_B = tools.extract_individual_data(data)
obs_data = tools.compute_observables(data_A, data_B)
for o in self.obs:
histograms[o][c] += tools.compute_histogram(o, obs_data[o])
obs_data_cum[c][o].extend(obs_data[o])
for o in self.obs:
for c in self.cl:
self.pdfs[o][c] = tools.compute_pdf(o, histograms[o][c])
for c in self.cl:
self.joint_pdfs[c], jaccard_dist[c] = {}, {}
for o1 in self.obs:
self.joint_pdfs[c][o1], jaccard_dist[c][o1] = {}, {}
for o2 in self.obs:
self.joint_pdfs[c][o1][o2] = tools.compute_joint_pdf(
tools.compute_joint_histogram(o1, obs_data_cum[c][o1],
o2, obs_data_cum[c][o2]))
joint_ent = tools.get_joint_ent(self.joint_pdfs[c][o1][o2],
self.pdfs[o1][c],
self.pdfs[o2][c])
mutual_inf = tools.get_mutual_inf(
self.joint_pdfs[c][o1][o2], self.pdfs[o1][c],
self.pdfs[o2][c])
# i should not need th follwoign wheck but all is nan
if mutual_inf is not 0:
jaccard_dist[c][o1][o2] = (joint_ent -
mutual_inf) / joint_ent
return jaccard_dist
def get_jaccard_dist(train_fnames):
histograms, pdfs, joint_pdfs, jaccard_dist = {}, {}, {}, {}
# initialize empty histograms
for o in preferences.OBSERVABLES:
histograms[o], pdfs[o] = {}, {}
for c in preferences.CLASSES:
histograms[o][c] = tools.initialize_histogram(o)
# compute histograms for each classes
obs_data_cum = {}
for c in preferences.CLASSES:
obs_data_cum[c] = {}
for o in preferences.OBSERVABLES:
obs_data_cum[c][o] = []
for file_path in train_fnames[c]:
data = np.load(file_path)
data_A, data_B = tools.extract_individual_data(data)
obs_data = tools.compute_observables(data_A, data_B)
for o in preferences.OBSERVABLES:
histograms[o][c] += tools.compute_histogram(o, obs_data[o])
obs_data_cum[c][o].extend(obs_data[o])
for o in preferences.OBSERVABLES:
for c in preferences.CLASSES:
pdfs[o][c] = tools.compute_pdf(o, histograms[o][c])
for c in preferences.CLASSES:
joint_pdfs[c], jaccard_dist[c] = {}, {}
for o1 in preferences.OBSERVABLES:
joint_pdfs[c][o1], jaccard_dist[c][o1] = {}, {}
for o2 in preferences.OBSERVABLES:
joint_pdfs[c][o1][o2] = tools.compute_joint_pdf(
tools.compute_joint_histogram(o1, obs_data_cum[c][o1], o2,
obs_data_cum[c][o2]))
joint_ent = tools.get_joint_ent(joint_pdfs[c][o1][o2],
pdfs[o1][c], pdfs[o2][c])
mutual_inf = tools.get_mutual_inf(joint_pdfs[c][o1][o2],
pdfs[o1][c], pdfs[o2][c])
# i should not need th follwoign wheck but all is nan
if mutual_inf is not 0:
jaccard_dist[c][o1][o2] = (joint_ent -
mutual_inf) / joint_ent
return jaccard_dist
def evaluate_distance(self, alpha, test_sets):
results = {}
confusion_matrix = {}
# print('-------------------------------')
# print('\t Right \t Wrong \t Rate\n')
t = 0
for c in self.cl:
results[c] = {'right': 0, 'wrong': 0}
# init condusion matrix
confusion_matrix[c] = {}
for c_pred in self.cl:
confusion_matrix[c][c_pred] = 0
for file_path in test_sets[c]:
data = np.load(file_path)
pdfs, distances = {}, {}
# initialize distances
for c_pred in self.cl:
distances[c_pred] = 0
data_A, data_B = tools.extract_individual_data(data)
obs_data = tools.compute_observables(data_A, data_B)
for o in self.obs:
pdfs[o] = tools.compute_pdf(
o, tools.compute_histogram(o, obs_data[o]))
for c_pred in self.cl:
distances[c_pred] += stats.energy_distance(
pdfs[o], self.pdfs[o][c_pred])
# t += i
class_max = min(distances.items(),
key=operator.itemgetter(1))[0]
confusion_matrix[c][class_max] += 1
if class_max == c:
results[c]['right'] += 1
else:
results[c]['wrong'] += 1
rate = results[c]['right'] / (results[c]['right'] +
results[c]['wrong'])
# print('{}\t {}\t {}\t {}'.format(c, results[c]['right'], results[c]['wrong'], rate))
# tools.print_confusion_matrix(self.cl, confusion_matrix)
# print(t)
return results