def train(self, weighted_ink_groups, verbose=True):
self.trained_prototypes = []
for weighted_ink_list in weighted_ink_groups:
# we skip small clusters
if len(weighted_ink_list) > self.min_cluster_size:
ink_data, weights = zip(*weighted_ink_list)
proto = PrototypeDTW(self.label, alpha=self.alpha)
avg_score = proto.train(
ink_data,
obs_weights=weights,
center_type=self.center_type,
state_reduction=self.state_reduction)
self.trained_prototypes.append(proto)
def closest_prototype(con, user_id, label, user_data):
cur = con.cursor()
cur.execute("""
SELECT t1.protoset_json
FROM protosets as t1
JOIN
(SELECT MAX(protoset_id) as pid
FROM protosets
WHERE label=%s AND user_id!=%s
GROUP BY user_id) as t2
ON t1.protoset_id = t2.pid
""", (label,user_id,))
rows = cur.fetchall()
# no prototypes found, return empty
if len(rows) == 0:
return []
# create a list of prototypes
proto_list = []
for row in rows:
protoset_json = json.loads(row[0])
for prototype_json in protoset_json['prototypes']:
p = PrototypeDTW(label)
p.fromJSON(prototype_json)
proto_list.append(p)
# select the best prototype
count = np.zeros(len(proto_list))
for ink in user_data:
scores = np.zeros(len(proto_list))
for i,p in enumerate(proto_list):
scores[i] = p.score(ink)
count[scores.argmax()] += 1
best_proto_ink = proto_list[count.argmax()].model
# make sure the penup is binary
best_proto_ink[:,_PU_IDX] = best_proto_ink[:,_PU_IDX].round()
return [best_proto_ink]
