def get_detection_scores(self,E,template,bgd_length, mean_background,
edge_feature_row_breaks,
edge_orientations,
abst_threshold=-np.inf *np.ones(8),
spread_length=3):
""" gets the detection scores for later processing
"""
template_height,template_length = template.shape
num_detections = E.shape[1]-template_length+1
E_background, estimated_background_idx = self._get_E_background(E,num_detections,bgd_length, mean_background,
edge_feature_row_breaks,
edge_orientations,
abst_threshold=abst_threshold,
spread_length=spread_length)
E_stack= self._get_E_stack(E,template.shape,num_detections)
print "E_stack has shape",E_stack.shape
# perform thresholding for the stack of E features
for frame_idx in xrange(E_stack.shape[1]):
E_segment = E_stack[:,frame_idx].reshape(template_height,template_length)
esp.threshold_edgemap(E_segment,.30,edge_feature_row_breaks,abst_threshold=abst_threshold)
esp.spread_edgemap(E_segment,edge_feature_row_breaks,edge_orientations,spread_length=spread_length)
bg_stack = np.tile(E_background,(template.shape[1],1))
T_stack = np.tile(template.transpose().reshape(np.prod(template.shape),1),
(1,num_detections))
T_inv = 1 - T_stack
bg_inv = 1- bg_stack
C_exp_inv_long = T_inv/bg_inv
C = np.log(C_exp_inv_long).sum(0)
return (E_stack*np.log(T_stack/bg_stack /\
C_exp_inv_long)).sum(0),\
C
def next(self):
if self.cur_window_id < self.num_detections-1:
self.cur_window_id+=1
else:
print "Error: End of utterance reached, reset window"
if self.pattern_times:
self.compute_roc()
return
self.cur_window = self.E[:,self.cur_window_id\
:self.cur_window_id+\
self.window_length].copy()
esp.threshold_edgemap(self.cur_window,self.edge_quantile,
self.edge_feature_row_breaks,
abst_threshold=self.abst_threshold)
esp.spread_edgemap(self.cur_window,
self.edge_feature_row_breaks,
self.edge_orientations,
spread_length=self.spread_length)
self.j0_scores[self.cur_window_id] =\
np.sum(cur_window[self.mask])
if self.cur_window_id >=2:
if self.j0_scores[self.cur_window_id-1]>\
max(self.j0_scores[self.cur_window_id],
self.j0_scores[self.cur_window_id-2]):
self.j0_maxima.append(cur_window_id-1)
def get_hyp_segment_bgd(self,E,E_loc,
edge_feature_row_breaks,
abst_threshold,
edge_orientations,
spread_length=3,bg_len=26):
E_segment = E[:,E_loc-self.front_def_radius-bg_len:
E_loc-self.front_def_radius]
esp.threshold_edgemap(E_segment,.30,edge_feature_row_breaks,abst_threshold=abst_threshold)
esp.spread_edgemap(E_segment,edge_feature_row_breaks,edge_orientations,spread_length=spread_length)
def get_hyp_segment(self,E,E_loc,
edge_feature_row_breaks,
abst_threshold,
edge_orientations,
spread_length=3):
E_segment = E[:,E_loc:
E_loc+self.deformed_max_length].copy()
esp.threshold_edgemap(E_segment,.30,edge_feature_row_breaks,abst_threshold=abst_threshold)
esp.spread_edgemap(E_segment,edge_feature_row_breaks,edge_orientations,spread_length=spread_length)
return E_segment
def _get_bgd(self,E_bgd_window,
edge_feature_row_breaks,
edge_orientations,
abst_threshold=-np.inf*np.ones(8),
spread_length=3):
esp.threshold_edgemap(E_bgd_window,.30,edge_feature_row_breaks,
abst_threshold=abst_threshold)
esp.spread_edgemap(E_bgd_window,edge_feature_row_breaks,edge_orientations,
spread_length=spread_length)
E_bgd_window = np.mean(E_bgd_window,axis=1)
E_bgd_window = np.maximum(np.minimum(E_bgd_window,.4),.1)
return E_bgd_window
def get_detection_scores_slow(self,E,template,bgd_length,mean_background,
edge_feature_row_breaks,
edge_orientations,
abst_threshold=-np.inf *np.ones(8),
spread_length=3):
template_height,template_length = template.shape
num_detections = E.shape[1]-template_length+1
E_background, estimated_background_idx = self._get_E_background(E,num_detections,bgd_length, mean_background,
edge_feature_row_breaks,
edge_orientations,
abst_threshold=abst_threshold,
spread_length=spread_length)
Ps = np.zeros(num_detections)
Cs = np.zeros(num_detections)
for frame_idx in xrange(num_detections):
E_segment = E[:,frame_idx:frame_idx+template_length].copy()
esp.threshold_edgemap(E_segment,.30,edge_feature_row_breaks,abst_threshold=abst_threshold)
esp.spread_edgemap(E_segment,edge_feature_row_breaks,edge_orientations,spread_length=spread_length)
Ps[frame_idx],Cs[frame_idx] = tt.score_template_background_section(template,E_background[:,frame_idx],E_segment)
return Ps,Cs
def get_roc_full(data_iter, classifier, coarse_thresh,
allowed_overlap = .2,
edge_feature_row_breaks= np.array([ 0.,
45.,
90.,
138.,
186.,
231.,
276.,
321.,
366.]),
edge_orientations=np.array([[ 1., 0.],
[-1., 0.],
[ 0., 1.],
[ 0., -1.],
[ 1., 1.],
[-1., -1.],
[ 1., -1.],
[-1., 1.]]),
abst_threshold=np.array([.025,.025,.015,.015,
.02,.02,.02,.02]),
spread_radius=3):
"""
Find the appropriate threshold for the coarse classifier, this
should be run on tuning data, and then we can get a level for the
tradeoff between false positives and false negatives the first
pair is the roc curve for the count test and the second is for the
coarse likelihood test
The way this works is that we find the scores for each window and
then we rank them and remove overlapping windows that are lower
rank if the overlap is below a certain percentage
We are assuming that the classifier has been implemented and initialized properly
"""
num_frames = 0
all_positive_counts = []
all_positive_likes = []
all_negative_counts = []
all_negative_likes = []
data_iter.reset_exp()
for datum_id in xrange(data_iter.num_data):
if datum_id % 10 == 0:
print "working on example", datum_id
if data_iter.next(compute_pattern_times=True,
max_template_length=classifier.window[1]):
pattern_times = data_iter.pattern_times
num_detections = data_iter.E.shape[1] - classifier.window[1]
num_frames += data_iter.E.shape[1]
scores = -np.inf * np.ones(num_detections)
for d in xrange(num_detections):
E_segment = data_iter.E[:,d:d+classifier.window[1]].copy()
esp.threshold_edgemap(E_segment,.30,edge_feature_row_breaks,report_level=False,abst_threshold=abst_threshold)
esp.spread_edgemap(E_segment,edge_feature_row_breaks,edge_orientations,spread_length=3)
coarse_count_score = classifier.coarse_score_count(E_segment)
if coarse_count_score > coarse_thresh:
scores[d] = classifier.score_no_bg(E_segment)
# now we get the indices sorted
score_indices = remove_overlapping_examples(np.argsort(scores),
classifier.window[1],
int(allowed_overlap*classifier.coarse_length))
positive_scores, negative_scores = get_pos_neg_scores(score_indices,pattern_times,
scores)
positive_likes, negative_likes = get_pos_neg_scores(like_indices,pattern_times,
coarse_like_scores)
all_positive_counts.extend(positive_counts)
all_negative_counts.extend(negative_counts)
all_positive_likes.extend(positive_likes)
all_negative_likes.extend(negative_likes)
else:
break
count_roc = get_roc(np.sort(all_positive_counts)[::-1],
np.sort(all_negative_counts)[::-1],
num_frames)
like_roc = get_roc(np.sort(all_positive_likes)[::-1],
np.sort(all_negative_likes)[::-1],
num_frames)
return count_roc, like_roc