def avaliate(classes, ctrs, dataset, n):
sum = 0
qtd = 0
total_wrong = 0
for user, fts in dataset.items():
true_centroid = [idc for idc, c in enumerate(classes) if user in c]
report = {}
for i in range(len(fts[0])):
winners = compute_distance(ctrs, fts[0][i], n)
st = "centers_"
st2 = ''
for i in range(len(winners)):
st2 += str(winners[i]) + "_"
st += st2
if st not in report.keys():
report.update({st: []})
distances = []
for i in range(len(winners)):
distances.append(l2(ctrs[true_centroid] - ctrs[winners[i]]))
# distance = l2(ctrs[true_centroid] - ctrs[winner])
minimum = min(distances)
if minimum > 0:
total_wrong += 1
report[st].append(minimum)
sum += minimum
qtd += 1
logger.info("Report for user {}: {}".format(user, report))
logger.info("Quantization error: {}".format(sum / qtd))
logger.info("Miss prediction tax: {}".format(total_wrong / qtd))
def normalize(v):
"""
Normalize vectors which don't have length 0
"""
l_ = l2(v)
if l_ == 0:
l_ = 1.0
return v / l_
def save_train_features(data):
start_timestamp = time.time()
dataset = {}
full_dataset = {}
logger.info("Entered on reduce train features process")
for user, paths in data.items():
features = [extract_features_of_roi(path)[0] for path in paths]
avg = np.array(Average(features))
arg = np.array(l2(vector - avg) for vector in features).argmin()
dataset.update({user: avg})
full_dataset.update({user: [features, paths]})
logger.info("Extraction concluded in {} seconds.".format(time.time() -
start_timestamp))
with open(DATAKEYS_PATH, "wb") as file:
pickle.dump(dataset, file)
with open(TRAIN_PATH, "wb") as file:
pickle.dump(full_dataset, file)
return 0
def second_relaxation(X, loss):
tau = loss / (l2(X) + (1 / (2 * self.C)))
return tau
def first_relaxation(X, loss):
tau = min(self.C, loss / l2(X))
return tau
def classic(X, loss):
tau = loss / l2(X)
return tau
def compute_distance(centroids, vector, n):
return n_greatest([l2(vector - c) for c in centroids], n)
def compute_lowest_distance(centroids, vector):
return np.array([l2(vector - c) for c in centroids]).argmin()
def vector_cos(v, w):
l = l2(v) * l2(w)
if l == 0:
l = 1.0
return dot(v, w) / l