def get_roc_curves(phn, num_mix_params,
train_example_lengths,bgd,
train_path,file_indices,
sp,ep):
FOMS = collections.defaultdict(list)
for num_mix in num_mix_params:
if num_mix > 1:
outfile = np.load('data/%d_templates.npz' % num_mix)
templates = tuple( outfile['arr_%d'%i] for i in xrange(len(outfile.files)))
else:
templates = (np.load('data/1_templates.npy')[0],)
detection_array = np.zeros((train_example_lengths.shape[0],
train_example_lengths.max() + 2),dtype=np.float32)
linear_filters_cs = et.construct_linear_filters(templates,
bgd)
np.savez('data/linear_filter_%d.npz'% num_mix,*(tuple(lfc[0] for lfc in linear_filters_cs)))
np.savez('data/c_%d.npz'%num_mix,*(tuple(lfc[1] for lfc in linear_filters_cs)))
syllable = np.array((phn,))
(detection_array,
example_start_end_times,
detection_lengths,
detection_template_ids)=gtrd.get_detection_scores_mixture_named_params(
train_path,
file_indices,
detection_array,
syllable,
linear_filters_cs,S_config=sp,
E_config=ep,
verbose = True,
num_examples =-1,
return_detection_template_ids=True)
np.save('data/detection_array_%d.npy' % num_mix,detection_array)
np.save('data/detection_template_ids_%d.npy' % num_mix,detection_template_ids)
np.save('data/detection_lengths_%d.npy' % num_mix,detection_lengths)
if num_mix == num_mix_params[0]:
out = open('data/example_start_end_times.pkl','wb')
cPickle.dump(example_start_end_times,out)
out.close()
window_start = -int(np.mean(tuple( t.shape[0] for t in templates))/3.+.5)
window_end = -window_start
max_detect_vals = rf.get_max_detection_in_syllable_windows(detection_array,
example_start_end_times,
detection_lengths,
window_start,
window_end)
max_detect_vals = max_detect_vals[:1000]
np.save('data/max_detect_vals_%d_%s.npy' % (num_mix,phn),max_detect_vals)
C0 = int(np.mean(tuple( t.shape[0] for t in templates))/3.+.5)
C1 = int( 33 * 1.5 + .5)
frame_rate = 1/.005
fpr, tpr = rf.get_roc_curve(max_detect_vals,
detection_array,
np.array(detection_lengths),
example_start_end_times,
C0,C1,frame_rate)
np.save('data/fpr_%d_%s.npy' % (num_mix,phn),
fpr)
np.save('data/tpr_%d_%s.npy' % (num_mix,phn),
tpr)
# Now we are going to test the mixtures
# First thing to do is to test the registered mixtures
# then we are going to try the padded mixtures
import template_speech_rec.bernoulli_mixture as bm
bem = bm.BernoulliMixture(2,aar_registered)
bem.run_EM(.000001)
detection_array = np.zeros((test_example_lengths.shape[0],
test_example_lengths.max() + 2),dtype=np.float32)
linear_filters_cs = et.construct_linear_filters(bem.templates,
clipped_bgd)
detection_array,example_start_end_times, detection_lengths = gtrd.get_detection_scores_mixture(data_path+'Test/',
detection_array,
syllable,
linear_filters_cs,verbose=True)
np.save(tmp_data_path+'detection_array_aar_2.npy',detection_array)
out = open(tmp_data_path+'example_start_end_times_aar_2.pkl','wb')
cPickle.dump(example_start_end_times,out)
out.close()
out = open(tmp_data_path+'detection_lengths_aar_2.pkl','wb')
cPickle.dump(detection_lengths,out)
np.save("data/test_example_lengths.npy",test_example_lengths)
np.save("data/test_file_indices.npy",test_file_indices)
import collections
FOMS = collections.defaultdict(list)
bgd = np.clip(np.load("data/aar_bgd_mel.npy"),.1,.4)
test_example_lengths =np.load("data/test_example_lengths.npy")
test_path = "/home/mark/Template-Speech-Recognition/Data/Test/"
test_file_indices =np.load("data/test_file_indices.npy")
for num_mix in num_mix_params:
templates = tuple(np.clip(T,.01,.99) for T in (np.load('aar1_templates_%d.npz' % num_mix))['arr_0'])
detection_array = np.zeros((test_example_lengths.shape[0],
test_example_lengths.max() + 2),dtype=np.float32)
linear_filters_cs = et.construct_linear_filters(templates,
bgd)
syllable = np.array(['aa','r'])
detection_array,example_start_end_times, detection_lengths = gtrd.get_detection_scores_mixture_named_params(test_path,test_file_indices,
detection_array,
syllable,
linear_filters_cs,S_config=sp,
E_config=ep,
verbose = False,
num_examples =-1)
np.save('data/detection_array_aar_%d.npy' % num_mix,detection_array)
if num_mix == 2:
out = open('data/example_start_end_times_aar.pkl','wb')
cPickle.dump(example_start_end_times,out)
out.close()
out = open('data/detection_lengths_aar.pkl','wb')
cPickle.dump(detection_lengths,out)
for i in xrange(2):
np.save(tmp_data_path + 'aar2_%d.npy' % i,aar2[i])
aar_mixture = tuple(
np.load(tmp_data_path+'aar2_%d.npy' % i)
for i in xrange(2))
# now we will get the detection array associated with these two
test_example_lengths = np.load(tmp_data_path+'test_example_lengths.npy')
detection_array = np.zeros((test_example_lengths.shape[0],
int(test_example_lengths.max()/float(log_part_blocks.shape[1]) + .5) + 2),dtype=np.float32)
linear_filters_cs = et.construct_linear_filters(aar_mixture,
clipped_bgd)
# need to state the syllable we are working with
syllable = np.array(['aa','r'])
detection_array,example_start_end_times, detection_lengths = gtrd.get_detection_scores_mixture(data_path+'Test/',
detection_array,
syllable,
linear_filters_cs,
log_part_blocks,
log_invpart_blocks,verbose=True)
np.save(tmp_data_path+'detection_array_aar_2.npy',detection_array)
out = open(tmp_data_path+'example_start_end_times_aar_2.pkl','wb')
def perform_phn_train_detection_SVM(phn, num_mix_params,
train_example_lengths,bgd,
train_path,file_indices,
first_pass_fnrs):
FOMS = collections.defaultdict(list)
for num_mix in num_mix_params:
if num_mix > 1:
outfile = np.load('data/%d_templates.npz' % num_mix)
templates = tuple( outfile['arr_%d'%i] for i in xrange(len(outfile.files)))
else:
templates = (np.load('data/1_templates.npy')[0],)
detection_array = np.zeros((train_example_lengths.shape[0],
train_example_lengths.max() + 2),dtype=np.float32)
linear_filters_cs = et.construct_linear_filters(templates,
bgd)
np.savez('data/linear_filter_%d.npy'% num_mix,*(tuple(lfc[0] for lfc in linear_filters_cs)))
np.savez('data/c_%d.npy'%num_mix,*(tuple(lfc[1] for lfc in linear_filters_cs)))
syllable = np.array((phn,))
(detection_array,
example_start_end_times,
detection_lengths,
detection_template_ids)=get_detection_scores_mixture_named_params(
utterances_path,
file_indices,
detection_array,
syllable,
linear_filters_cs,S_config=sp,
E_config=ep,
verbose = True,
num_examples =-1,
return_detection_template_ids=True)
np.save('data/detection_array_%d.npy' % num_mix,detection_array)
np.save('data/detection_template_ids_%d.npy' % num_mix,detection_template_ids)
if num_mix == num_mix[0]:
out = open('data/example_start_end_times.pkl','wb')
cPickle.dump(example_start_end_times,out)
out.close()
out = open('data/detection_lengths.pkl','wb')
cPickle.dump(detection_lengths,out)
out.close()
window_start = -int(np.mean(tuple( t.shape[0] for t in templates))/3.+.5)
window_end = -window_start
max_detect_vals = rf.get_max_detection_in_syllable_windows(detection_array,
example_start_end_times,
detection_lengths,
window_start,
window_end)
np.save('data/max_detect_vals_%d.npy' % num_mix,max_detect_vals)
C0 = int(np.mean(tuple( t.shape[0] for t in templates))/3.+.5)
C1 = int( 33 * 1.5 + .5)
frame_rate = 1/.005
fpr, tpr = rf.get_roc_curve(max_detect_vals,
detection_array,
np.array(detection_lengths),
example_start_end_times,
C0,C1,frame_rate)
np.save('data/fpr_%d.npy' % num_mix,
fpr)
np.save('data/tpr_%d.npy' % num_mix,
tpr)
detection_clusters = rf.get_detect_clusters_threshold_array(max_detect_vals,
detection_array,
np.array(detection_lengths),
C0,C1)
out = open('data/detection_clusters_%d.npy' % num_mix,
'wb')
cPickle.dump(detection_clusters,out)
out.close()
for i in xrange(1,11):
thresh_idx = np.arange(fpr.shape[0])[fpr*60 <= i].min()
FOMS[num_mix].append(tpr[thresh_idx])
template_lengths = tuple(t.shape[0] for t in templates)
for fnr in first_pass_fnrs:
print "num_mix=%d,fnr=%d" % (num_mix,fnr)
thresh_id = int(len(detection_clusters)* fnr/100. + 5)
(pos_times,
false_pos_times,
false_neg_times) = rf.get_pos_false_pos_false_neg_detect_points(
detection_clusters[thresh_id],
detection_array,
detection_template_ids,
template_lengths,
window_start,
window_end,example_start_end_times,
utterances_path,
train_file_indices,
verbose=True)
out = open('data/false_pos_times_%d_%d.pkl' % (num_mix,fnr),'wb')
pickle.dump(false_pos_times,out)
out.close()
out = open('data/pos_times_%d_%d.pkl' % (num_mix,fnr),'wb')
pickle.dump(pos_times,out)
out.close()
out = open('data/false_neg_times_%d_%d.pkl' % (num_mix,fnr),'wb')
pickle.dump(false_pos_times,out)
out.close()
return FOMS
