class TestPrototypeDTW(_BaseTest):
def setUp(self):
super(TestPrototypeDTW,self).setUp()
self.p = PrototypeDTW(self.label, alpha=0.5)
self.p.train(self.ink_data)
def test_serialization(self):
p_data = self.p.toJSON()
q = PrototypeDTW(None)
q.fromJSON(p_data)
self.assertEqual(p_data, q.toJSON())
def test_score(self):
score0 = self.p.score(self.ink_data[0])
if VERBOSE: print score0
self.assertAlmostEqual(score0, -0.088, delta=1e-3)
score1 = self.p.score(self.ink_data[1])
if VERBOSE: print score1
self.assertAlmostEqual(score1, -0.049, delta=1e-3)
def test_centroid(self):
q = PrototypeDTW('u', alpha=0.5)
q.train(self.ink_data, center_type="centroid")
score0 = q.score(self.ink_data[0])
if VERBOSE: print score0
self.assertAlmostEqual(score0, -0.088, delta=1e-3)
def test_state_reduction(self):
q = PrototypeDTW('u', alpha=0.5)
q.train(self.ink_data, center_type="centroid", state_reduction=True)
score0 = q.score(self.ink_data[0])
if VERBOSE: print score0
self.assertAlmostEqual(score0, -0.089, delta=1e-3)
def train(self, clustered_ink_data, center_type="centroid", verbose=False, state_reduction=False):
"""Trains the classifier.
Parameters
----------
clustered_ink_data : dictionary
Training data with the following structure.
clustered_ink_data[label][cluster_id] = [ink1, ink2, ... inkN]
center_type : {'medoid', 'centroid'}
Type of the DTW center.
state_reduction : bool
If state_reduction=True, state-reduction procedure is performed
after each prototype is trained.
"""
self._trained_prototypes = []
for label in clustered_ink_data:
for ink_list in clustered_ink_data[label]:
if len(ink_list) > self.min_cluster_size:
ink_data, weights = zip(*ink_list)
proto = PrototypeDTW(label, alpha=self.alpha)
avg_score = proto.train(
ink_data, obs_weights=weights, center_type=center_type, state_reduction=state_reduction
)
self._trained_prototypes.append(proto)
if verbose:
print(
"Prototype for " "%s (%d instances, avg.score = %0.2f)" % (label, len(ink_list), avg_score)
)
self._compute_log_priors()
def fromJSON(self, json_obj):
prototypes = json_obj["prototypes"]
self._trained_prototypes = []
self.log_priors = np.zeros(len(prototypes))
for i, prototype in enumerate(prototypes):
p = PrototypeDTW(prototype["label"], prototype["alpha"])
p.fromJSON(prototype)
self._trained_prototypes.append(p)
self.log_priors[i] = np.log(prototype["prior"])
def _train_prototypes(weighted_ink, partition, label,
min_cluster_size=5, center_type='centroid'):
obs, obs_weights = zip(*weighted_ink)
obs_weights = np.asarray(obs_weights)
trained_prototypes = []
for k in range(np.max(partition)):
clustered_obs = [obs[j]
for j in range(len(obs))
if partition[j] == (k+1)]
clustered_weights = obs_weights[partition == (k+1)]
if len(clustered_obs) > min_cluster_size:
p_dtw = PrototypeDTW(label)
p_dtw.train(clustered_obs,
obs_weights=clustered_weights,
center_type=center_type)
trained_prototypes.append(p_dtw)
return trained_prototypes
def test_state_reduction(self):
q = PrototypeDTW('u', alpha=0.5)
q.train(self.ink_data, center_type="centroid", state_reduction=True)
score0 = q.score(self.ink_data[0])
if VERBOSE: print score0
self.assertAlmostEqual(score0, -0.089, delta=1e-3)
def test_centroid(self):
q = PrototypeDTW('u', alpha=0.5)
q.train(self.ink_data, center_type="centroid")
score0 = q.score(self.ink_data[0])
if VERBOSE: print score0
self.assertAlmostEqual(score0, -0.088, delta=1e-3)
def test_serialization(self):
p_data = self.p.toJSON()
q = PrototypeDTW(None)
q.fromJSON(p_data)
self.assertEqual(p_data, q.toJSON())
def setUp(self):
super(TestPrototypeDTW,self).setUp()
self.p = PrototypeDTW(self.label, alpha=0.5)
self.p.train(self.ink_data)
