def extract_false_positives(num_mix_params,first_pass_fns):
for num_mix in num_mix_params:
affinities = np.load('data/%d_affinities.npy' % (num_mix))
for fnr in first_pass_fnrs:
out = open('data/false_pos_times_%d_%d.pkl' % (num_mix,fnr),'rb')
false_pos_times=pickle.load(out)
out.close()
false_positives = rf.get_false_positives(false_pos_times,
S_config=sp,
E_config=ep,
offset=0,
waveform_offset=7)
example_mat = gtrd.recover_example_map(false_positives)
false_pos_assigned_phns,false_pos_phn_contexts,utt_paths,file_idx,start_ends = gtrd.recover_assigned_phns(false_positives,example_mat)
np.savez('data/false_pos_phns_assigned_contexts_%d_%d.npy' % (num_mix,fnr),
example_mat=example_mat,
assigned_phns=false_pos_assigned_phns,
phn_contexts=false_pos_phn_contexts,
utt_paths=utt_paths,
file_idx=file_idx,
start_ends=start_ends)
np.save('data/false_positives_example_mat_%d_%d.npy' % (num_mix,fnr),example_mat)
lengths,waveforms = gtrd.recover_waveforms(false_positives,example_mat)
np.savez('data/false_positives_waveforms_lengths_%d_%d.npz' % (num_mix,fnr),
lengths=lengths,
waveforms=waveforms)
Slengths,Ss = gtrd.recover_specs(false_positives,example_mat)
np.savez('data/false_positives_Ss_lengths_%d_%d.npz' % (num_mix,fnr),
lengths=Slengths,
Ss=Ss)
Elengths,Es = gtrd.recover_edgemaps(false_positives,example_mat)
np.savez('data/false_positives_Es_lengths_%d_%d.npz' % (num_mix,fnr),
lengths=Elengths,
Es=Es)
templates = et.recover_different_length_templates(affinities,
Es,
Elengths)
spec_templates = et.recover_different_length_templates(affinities,
Ss,
Slengths)
cluster_counts = np.zeros(num_mix)
for mix_id in xrange(num_mix):
os.system("mkdir -p data/%d_%d_%d" % (num_mix,fnr,mix_id))
for example_id in xrange(waveforms.shape[0]):
if affinities[example_id].max() > .999:
cluster_id = np.argmax(affinities[example_id])
wavfile.write('data/%d_%d_%d/%d.wav' % (num_mix,fnr,cluster_id,cluster_counts[cluster_id]),16000,((2**15-1)*waveforms[example_id]).astype(np.int16))
cluster_counts[cluster_id] += 1
ep = gtrd.EdgemapParameters(block_length=40,
spread_length=1,
threshold=.7)
utterances_path = '/home/mark/Template-Speech-Recognition/Data/Train/'
file_indices = gtrd.get_data_files_indices(utterances_path)
syllable=('aa','r')
syllable_features,avg_bgd=gtrd.get_syllable_features_directory(utterances_path,file_indices,syllable,
S_config=sp,E_config=ep,offset=0,
E_verbose=False,return_avg_bgd=True)
np.save('data/aar_bgd_mel.npy',avg_bgd.E)
example_mat = gtrd.recover_example_map(syllable_features)
lengths,waveforms = gtrd.recover_waveforms(syllable_features,example_mat)
np.savez('data/aar_waveforms_lengths.npz',waveforms,lengths,example_mat)
Slengths,Ss = gtrd.recover_specs(syllable_features,example_mat)
np.savez('data/aar_Ss_mel_lengths.npz',Ss,Slengths,example_mat)
lengths,Es = gtrd.recover_edgemaps(syllable_features,example_mat)
np.savez('data/aar_Es_mel_lengths.npz',Es,lengths,example_mat)
#
# Now that we have the Es we estimate the model
#
def perform_phn_template_estimation(phn,utterances_path,
file_indices,sp,ep,
num_mix_params,
phn_mapping=None,
waveform_offset=15,
chunk_length=1000):
phn_tuple = (phn,)
print phn
phn_features,avg_bgd=gtrd.get_syllable_features_directory(utterances_path,file_indices,phn_tuple,
S_config=sp,E_config=ep,offset=0,
E_verbose=False,return_avg_bgd=True,
waveform_offset=15,
phn_mapping=phn_mapping)
bgd = np.clip(avg_bgd.E,.01,.99)
np.save('data/bgd.npy',bgd)
example_mat = gtrd.recover_example_map(phn_features)
lengths,waveforms = gtrd.recover_waveforms(phn_features,example_mat)
np.savez('data/waveforms_lengths.npz',waveforms=waveforms,
lengths=lengths,
example_mat=example_mat)
Slengths,Ss = gtrd.recover_specs(phn_features,example_mat)
Ss = Ss.astype(np.float32)
np.savez('data/Ss_lengths.npz' ,Ss=Ss,Slengths=Slengths,example_mat=example_mat)
Elengths,Es = gtrd.recover_edgemaps(phn_features,example_mat,bgd=bgd)
Es = Es.astype(np.uint8)
np.savez('data/Es_lengths.npz' ,Es=Es,Elengths=Elengths,example_mat=example_mat)
# the Es are padded from recover_edgemaps
f = open('data/mixture_estimation_stats_%s.data' % phn,'w')
for num_mix in num_mix_params:
print num_mix
if num_mix == 1:
affinities = np.ones((Es.shape[0],1),dtype=np.float64)
mean_length = int(np.mean(Elengths) + .5)
templates = (np.mean(Es,0)[:mean_length],)
spec_templates = (np.mean(Ss,0)[:mean_length],)
np.save('data/%d_affinities.npy' % (num_mix),
affinities)
np.save('data/%d_templates.npy' % (num_mix),
templates)
np.save('data/%d_spec_templates.npy' % (num_mix),
spec_templates)
np.save('data/%d_templates_%s.npy' % (num_mix,phn),
templates)
np.save('data/%d_spec_templates_%s.npy' % (num_mix,phn),
spec_templates)
#
# write the data to the mixture file for checking purposes
# format is:
# num_components total c0 c1 c2 ... ck
f.write('%d %d %g\n' % (num_mix,
len(affinities),np.sum(affinities[:,0])))
else:
if len(Es) > chunk_length:
bem = bm.BernoulliMixture(num_mix,Es[:chunk_length])
bem.run_EM(.000001)
for i in xrange(1,len(Es)/chunk_length):
start_idx = i*chunk_length
block_length = min(chunk_length,len(Es)-start_idx)
if block_length < chunk_length:
end_idx = len(Es)
start_idx = len(Es)-chunk_length
block_length = chunk_length
else:
end_idx = start_idx + block_length
bem.data_mat = Es[start_idx:end_idx].reshape(
block_length,bem.data_length)
bem.run_EM(.000001)
else:
bem = bm.BernoulliMixture(num_mix,Es)
bem.run_EM(.000001)
templates = et.recover_different_length_templates(bem.affinities,
Es[start_idx:end_idx],
Elengths[start_idx:end_idx])
spec_templates = et.recover_different_length_templates(bem.affinities,
Ss[start_idx:end_idx],
Slengths[start_idx:end_idx])
np.save('data/%d_affinities.npy' % (num_mix),
bem.affinities)
np.savez('data/%d_templates.npz' % (num_mix),
*templates)
np.savez('data/%d_spec_templates.npz' % (num_mix),
*spec_templates)
np.savez('data/%d_templates_%s.npz' % (num_mix,phn),
*templates)
np.savez('data/%d_spec_templates_%s.npz' % (num_mix,phn),
*spec_templates)
f.write('%d %d ' % (num_mix,
len(affinities))
+ ' '.join(str(np.sum(affinities[:,i]))
for i in xrange(affinities.shape[1]))
+'\n')
f.close()
for phn_idx, phn in enumerate(use_phns[12:]):
print phn_idx, phn
phn_tuple = (phn,)
print phn
waveform_offset = 15
phn_mapping=leehon_mapping
chunk_length=1000
phn_features,avg_bgd=gtrd.get_syllable_features_directory(utterances_path,file_indices,phn_tuple,
S_config=sp,E_config=ep,offset=0,
E_verbose=False,return_avg_bgd=True,
waveform_offset=15,
phn_mapping=phn_mapping)
bgd = np.clip(avg_bgd.E,.01,.99)
np.save('data/bgd.npy',bgd)
example_mat = gtrd.recover_example_map(phn_features)
lengths,waveforms = gtrd.recover_waveforms(phn_features,example_mat)
np.savez('data/waveforms_lengths.npz',waveforms=waveforms,
lengths=lengths,
example_mat=example_mat)
Slengths,Ss = gtrd.recover_specs(phn_features,example_mat)
Ss = Ss.astype(np.float32)
np.savez('data/Ss_lengths.npz' ,Ss=Ss,Slengths=Slengths,example_mat=example_mat)
Elengths,Es = gtrd.recover_edgemaps(phn_features,example_mat,bgd=bgd)
Es = Es.astype(np.uint8)
np.savez('data/Es_lengths.npz' ,Es=Es,Elengths=Elengths,example_mat=example_mat)
# the Es are padded from recover_edgemaps
f = open('data/mixture_estimation_stats_%s.data' % phn,'w')
for num_mix in num_mix_params:
print num_mix
if num_mix == 1:
affinities = np.ones((Es.shape[0],1),dtype=np.float64)
def save_syllable_features_to_data_dir(
args,
phn_tuple,
utterances_path,
file_indices,
sp,
ep,
phn_mapping,
tag_data_with_syllable_string=False,
save_tag="train",
waveform_offset=10,
block_features=False,
):
"""
Wrapper function to get all the examples processed
"""
print "Collecting the data for phn_tuple " + " ".join("%s" % k for k in phn_tuple)
syllable_string = "_".join(p for p in phn_tuple)
phn_features, avg_bgd = gtrd.get_syllable_features_directory(
utterances_path,
file_indices,
phn_tuple,
S_config=sp,
E_config=ep,
offset=0,
E_verbose=False,
return_avg_bgd=True,
waveform_offset=15,
phn_mapping=phn_mapping,
)
bgd = np.clip(avg_bgd.E, 0.01, 0.4)
np.save("data/bgd.npy", bgd)
example_mat = gtrd.recover_example_map(phn_features)
lengths, waveforms = gtrd.recover_waveforms(phn_features, example_mat)
if tag_data_with_syllable_string:
np.savez(
"data/%s_waveforms_lengths_%s.npz" % (syllable_string, save_tag),
waveforms=waveforms,
lengths=lengths,
example_mat=example_mat,
)
else:
np.savez(
"data/waveforms_lengths_%s.npz" % save_tag, waveforms=waveforms, lengths=lengths, example_mat=example_mat
)
Slengths, Ss = gtrd.recover_specs(phn_features, example_mat)
Ss = Ss.astype(np.float32)
if tag_data_with_syllable_string:
np.savez(
"data/%s_Ss_lengths_%s.npz" % (syllable_string, save_tag), Ss=Ss, Slengths=Slengths, example_mat=example_mat
)
else:
np.savez("data/Ss_lengths_%s.npz" % (save_tag), Ss=Ss, Slengths=Slengths, example_mat=example_mat)
Elengths, Es = gtrd.recover_edgemaps(phn_features, example_mat, bgd=bgd)
Es = Es.astype(np.uint8)
if tag_data_with_syllable_string:
np.savez(
"data/%s_Es_lengths_%s.npz" % (syllable_string, save_tag), Es=Es, Elengths=Elengths, example_mat=example_mat
)
else:
np.savez("data/Es_lengths_%s.npz" % (save_tag), Es=Es, Elengths=Elengths, example_mat=example_mat)
if args.doBlockFeatures:
out = code_parts.code_parts(E.astype(np.uint8), logParts, logInvParts, args.bernsteinEdgeThreshold)
max_responses = np.argmax(out, -1)
Bs = code_parts.spread_patches(max_responses, 2, 2, out.shape[-1] - 1)
