def compute_percentile_param(percentile, features_file, comparison='within',
samples=100000, with_vad=False, lsh=False):
"""Compute the threshold parameter from a percentile value
percentile: int or list of int
"""
assert comparison in set(['general', 'within', 'across'])
if comparison != 'within':
raise NotImplementedError
if not lsh:
if not with_vad:
features_dict = h5features.read(features_file)[1]
else:
features_dict = {}
vad = read_vad(with_vad)
for f, intervals in vad.iteritems():
feats = [h5features.read(features_file, from_item=f, from_time=interval[0], to_time=interval[1])[1][f] for interval in intervals]
features_dict[f] = np.concatenate(feats, axis=0)
else:
features_dict = read_sigs_remove_sils(features_file)
# else:
# features_dict = {}
# vad = read_vad(with_vad)
# for f, intervals in vad.iteritems():
# feats = []
# for interval in intervals:
# interval = (np.array(interval) * 100).astype(int)
# feats.append(np.fromfile(features_file[f]['sig'], dtype=dtype)[interval[0]:interval[1]])
# features_dict[f] = np.concatenate(feats, axis=0)
samples_per_file = samples // len(features_dict)
def estimate_similarities_distribution(feats1, feats2, n_samples, S=False):
samples1 = np.random.choice(feats1.shape[0], n_samples)
samples2 = np.random.choice(feats2.shape[0], n_samples)
sampled_feats1 = feats1[samples1]
sampled_feats2 = feats2[samples2]
if not S:
fun = do_cosine_similarity
else:
fun = lambda x, y: do_norm_hamming_sim(x, y, S)
return fun(sampled_feats1, sampled_feats2)
similarities = []
for feats in features_dict.itervalues():
S = False
if lsh:
S = features_file.stats['S']
similarities.append(estimate_similarities_distribution(
feats, feats, samples_per_file, S=S))
similarities = np.concatenate(similarities)
return np.percentile(similarities, percentile)
def test_read_works(self):
fname = self.file_v1
t1, f1 = h5f_1_0.read(fname, 'features')
t2, f2 = h5f_1_1.read(fname, 'features')
for tt1, tt2 in zip(t1, t2):
assert tt1 == tt2
for ff1, ff2 in zip(f1, f2):
assert ff1 == ff2
def test_read_works(self):
fname = self.file_v1
t1, f1 = h5f_1_0.read(fname, 'features')
t2, f2 = h5f_1_1.read(fname, 'features')
for tt1, tt2 in zip(t1, t2):
assert tt1 == tt2
for ff1, ff2 in zip(f1, f2):
assert ff1 == ff2
def do_transform(fname):
times, feats = h5features.read(stacked_fb_file,
from_internal_file=fname,
index=index)
times = times[fname]
feats = feats[fname]
X = np.asarray((feats - mean) / std, dtype='float32')
# times = np.arange(0.01, 0.01*npz.shape[0], 0.01)
emb_wrd, emb_spkr = transform(X)
return emb_wrd
def h5features_stack_fbanks(fbanks_file, stacked_fbanks_file):
import h5features
index = h5features.read_index(fbanks_file)
files = index['files']
for f in files:
times, fbanks = h5features.read(
fbanks_file, 'features', from_internal_file=f, index=index)
stacked_fbanks = stack_fbanks(fbanks[f])
h5features.write(stacked_fbanks_file, 'features', [f],
[times[f]], [stacked_fbanks])
def _test_wr(self, labeldim):
"""Test retrieving labels and files after a write/read operation."""
items, t_gold, feat = generate.full(self.nbitems, tformat=labeldim)
write(self.filename, self.group, items, t_gold, feat)
t, _ = read(self.filename, self.group)
assert len(t) == self.nbitems
if not labeldim == 1:
assert all([tt.shape[1] == labeldim for tt in t.values()])
# build a dict from gold to compare with t
d = dict(zip(items, t_gold))
for dd, tt in zip(d, t):
assert tt == dd
def _test_wr(self, labeldim):
"""Test retrieving labels and files after a write/read operation."""
items, t_gold, feat = generate.full(self.nbitems, tformat=labeldim)
write(self.filename, self.group, items, t_gold, feat)
t, _ = read(self.filename, self.group)
assert len(t) == self.nbitems
if not labeldim == 1:
assert all([tt.shape[1] == labeldim for tt in t.values()])
# build a dict from gold to compare with t
d = dict(zip(items, t_gold))
for dd, tt in zip(d, t):
assert tt == dd
def write_kl_to_column(distance_list, PG_file, root):
""" Write distances into original table """
hf5_file = root + PG_file
times_r, features_r = h5f.read(hf5_file, 'features')
items = h5f.Reader(hf5_file, 'features').items.data[0:]
oth_x_array = np.array([])
tgt_x_array = np.array([])
for TRIP_NUM in range(1, 113):
# select only item names which correspond to same triplet
trip_id = 'triplet' + str('{0:03}'.format(TRIP_NUM))
trip_items = [itm for itm in items if trip_id in itm]
# trace the 01 = OTH, 02 = TGT, 03 = X
item_oth = [oth for oth in trip_items if '_01' in oth][0]
item_tgt = [tgt for tgt in trip_items if '_02' in tgt][0]
item_x = [x for x in trip_items if '_03' in x][0]
# find vectors
feat_vector_oth = features_r[item_oth]
feat_vector_tgt = features_r[item_tgt]
feat_vector_x = features_r[item_x]
# time_vector = times_r[item]
# get KL divergence for TGT-X and OTH-X
kl_oth_x = dtw_kl_divergence(feat_vector_oth, feat_vector_x)
kl_tgt_x = dtw_kl_divergence(feat_vector_tgt, feat_vector_x)
# put them into an array
oth_x_array = np.append(oth_x_array, kl_oth_x)
tgt_x_array = np.append(tgt_x_array, kl_tgt_x)
name_othX = PG_file.split('.')[0] + '_oth_x'
name_tgtX = PG_file.split('.')[0] + '_tgt_x'
distance_list[name_othX] = pd.Series(oth_x_array, \
index=distance_list.index)
distance_list[name_tgtX] = pd.Series(tgt_x_array, \
index=distance_list.index)
return distance_list
def segment_features(features_file, segments_file, out_file):
"""
Segment h5features file containing features for whole wavefiles
of an abkhazia corpus (or split of a corpus) into features for
segments as described in the provided segments.txt file.
"""
utt_ids, wavefiles, starts, stops = io.read_segments(segments_file)
if all([e is None for e in starts]) and all([e is None for e in stops]):
# TODO use a log instead of a print statement
print(
"segment_features: segments already match wavefiles, "
"doing nothing...")
else:
# Group utterances by wavefiles
data = zip(utt_ids, wavefiles, starts, stops)
for wav, utts in groupby(data, lambda e: e[1]):
# TODO use a log instead of a print statement
print "Segmenting features for file {} by utterance".format(wav)
# load features for whole wavefile
wav_id = os.path.splitext(wav)[0]
# TODO fix that
times, features = h5features.read(features_file,
from_internal_file=wav_id)
# no need for dict here
times, features = times[wav_id], features[wav_id]
utt_ids, utt_times, utt_features = [], [], []
for utt_id, _, start, stop in utts:
# select features for appropriate segment
utt_ids.append(utt_id)
indices = np.where(
np.logical_and(times >= start, times <= stop))[0]
# get times relative to beginning of utterance
utt_times.append(times[indices] - start)
utt_features.append(features[indices, :])
# write to out_file once for each wavefile
h5features.write(out_file, 'features', utt_ids, utt_times,
utt_features)
def get_features_flat(fn, gn, feature_index, key, stat):
features_dict = h5features.read(fn, gn, key, index=feature_index)[1]
features = features_dict[features_dict.keys()[0]]
return stat(features)
def time_f(f):
return h5features.read(fb_h5f, from_item=f)[0][f]
def aux(f):
return stack_fbanks(h5features.read(fb_h5f, from_item=f)[1][f],
nframes=nframes)
def get_features_from_file(self, fileid):
"""return the features associated to a file"""
return h5features.read(self.feature_file, from_item=fileid)[1][fileid]
def run_distance_job(job_description, distance_file, distance, feature_files,
feature_groups, splitted_features, job_id, normalize):
if lock is None:
synchronize = False
else:
synchronize = True
if not (splitted_features):
times = {}
features = {}
for feature_file, feature_group in zip(feature_files, feature_groups):
t, f = h5features.read(feature_file, feature_group)
assert not (set(times.keys()).intersection(
t.keys())), ("The same file is indexed by (at least) two "
"different feature files")
times.update(t)
features.update(f)
get_features = Features_Accessor(times, features).get_features_from_raw
pair_file = job_description['pair_file']
n_blocks = len(job_description['by'])
for b in range(n_blocks):
print('Job %d: computing distances for block %d on %d' %
(job_id, b, n_blocks))
# get block spec
by = job_description['by'][b]
start = job_description['start'][b]
stop = job_description['stop'][b]
if splitted_features:
# FIXME modify feature_file/feature_group to adapt to 'by'
# FIXME any change needed when several feature files before
# splitting ?
times = {}
features = {}
for feature_file, feature_group in zip(feature_files,
feature_groups):
t, f = h5features.read(feature_file, feature_group)
assert not (set(times.keys()).intersection(
t.keys())), ("The same file is indexed by (at least) two "
"different feature files")
times.update(t)
features.update(f)
accessor = Features_Accessor(times, features)
get_features = accessor.get_features_from_splitted
# load pandas dataframe containing info for loading the features
if synchronize:
lock.acquire()
store = pandas.HDFStore(pair_file)
by_db = store['feat_dbs/' + by]
store.close()
# load pairs to be computed
# indexed relatively to the above dataframe
with h5py.File(pair_file) as fh:
attrs = fh['unique_pairs'].attrs[by]
pair_list = fh['unique_pairs/data'][attrs[1] + start:attrs[1] +
stop, 0]
base = attrs[0]
if synchronize:
lock.release()
A = np.mod(pair_list, base)
B = pair_list // base
pairs = np.column_stack([A, B])
n_pairs = pairs.shape[0]
# get dataframe with one entry by item involved in this block
# indexed by its 'by'-specific index
by_inds = np.unique(np.concatenate([A, B]))
items = by_db.iloc[by_inds]
# get a dictionary whose keys are the 'by' indices
features = get_features(items)
dis = np.empty(shape=(n_pairs, 1))
# FIXME: second dim is 1 because of the way it is stored to disk,
# but ultimately it shouldn't be necessary anymore
# (if using axis arg in np2h5, h52np and h5io...)
for i in range(n_pairs):
dataA = features[pairs[i, 0]]
dataB = features[pairs[i, 1]]
if dataA.shape[0] == 0:
warnings.warn(
'No features found for file {}, {} - {}'.format(
items['file'][pairs[i, 0]], items['onset'][pairs[i,
0]],
items['offset'][pairs[i, 0]]), UserWarning)
if dataB.shape[0] == 0:
warnings.warn(
'No features found for file {}, {} - {}'.format(
items['file'][pairs[i, 1]], items['onset'][pairs[i,
1]],
items['offset'][pairs[i, 1]]), UserWarning)
try:
if normalize is not None:
if normalize == 1:
normalize = True
elif normalize == 0:
normalize = False
else:
print('normalized parameter neither 1 nor 0,'
'using normalization')
normalize = True
dis[i, 0] = distance(dataA, dataB, normalized=normalize)
else:
dis[i, 0] = distance(dataA, dataB)
except:
sys.stderr.write(
'Error when calculating the distance between item {}, {} - {} '
'and item {}, {} - {}\n'.format(
items['file'][pairs[i, 0]], items['onset'][pairs[i,
0]],
items['offset'][pairs[i, 0]], items['file'][pairs[i,
1]],
items['onset'][pairs[i,
1]], items['offset'][pairs[i,
1]]), )
raise
if synchronize:
lock.acquire()
with h5py.File(distance_file) as fh:
fh['distances/data'][attrs[1] + start:attrs[1] + stop, :] = dis
if synchronize:
lock.release()
def get_features(self, segment):
"""return the features associated to a segment = (file, start, end)"""
fileid, start, end = segment
return h5features.read(self.feature_file, from_internal_file=fileid,
from_time=start, to_time=end,
index=self.index)[1][fileid]
def time_f(f):
return h5features.read(fb_h5f, from_item=f)[0][f]
def aux(f):
return stack_fbanks(h5features.read(fb_h5f, from_item=f)[1][f],
nframes=nframes)
# Author: Nika Jurov
import h5features as h5f
import numpy as np
import pandas as pd
import sys
import os
hf5_file = sys.argv[1] # the posterior gram file in .h5 format
TRIPLET_NAME = sys.argv[2] # which index / utterance do we want
NAMED_PG = sys.argv[3] # name the extracted PG
times_r, features_r = h5f.read(hf5_file, 'features')
#items = h5f.Reader(hf5_file, 'features').items.data[0:]
#utterance = items[int(NUMBER)]
f = pd.DataFrame(features_r[TRIPLET_NAME])
f['times'] = ["time_" + str('{0:03}'.format(i)) for i in range(0, f.shape[0])]
f.to_csv("model/supervised/posterior_grams/extracted_pgs/" \
+ NAMED_PG + ".csv", index=False)
#print "The utterance ID is: " + str(utterance)
def run_distance_job(job_description, distance_file, distance,
feature_files, feature_groups, splitted_features,
job_id, distance_file_lock=None):
if distance_file_lock is None:
synchronize = False
else:
synchronize = True
if not(splitted_features):
times = {}
features = {}
for feature_file, feature_group in zip(feature_files, feature_groups):
t, f = h5features.read(feature_file, feature_group)
assert not(set(times.keys()).intersection(
t.keys())), ("The same file is indexed by (at least) two "
"different feature files")
times.update(t)
features.update(f)
get_features = Features_Accessor(times, features).get_features_from_raw
pair_file = job_description['pair_file']
n_blocks = len(job_description['by'])
for b in range(n_blocks):
print('Job %d: computing distances for block %d on %d' % (job_id, b,
n_blocks))
# get block spec
by = job_description['by'][b]
start = job_description['start'][b]
stop = job_description['stop'][b]
if splitted_features:
# FIXME modify feature_file/feature_group to adapt to 'by'
# FIXME any change needed when several feature files before
# splitting ?
times = {}
features = {}
for feature_file, feature_group in zip(feature_files,
feature_groups):
t, f = h5features.read(feature_file, feature_group)
assert not(set(times.keys()).intersection(
t.keys())), ("The same file is indexed by (at least) two "
"different feature files")
times.update(t)
features.update(f)
accessor = Features_Accessor(times, features)
get_features = accessor.get_features_from_splitted
# load pandas dataframe containing info for loading the features
store = pandas.HDFStore(pair_file)
by_db = store['feat_dbs/' + by]
store.close()
# load pairs to be computed
# indexed relatively to the above dataframe
with h5py.File(pair_file) as fh:
pair_list = fh['unique_pairs/' + by][start:stop, 0]
base = fh['unique_pairs'].attrs[by]
A = np.mod(pair_list, base)
B = pair_list // base
pairs = np.column_stack([A, B])
n_pairs = pairs.shape[0]
# get dataframe with one entry by item involved in this block
# indexed by its 'by'-specific index
by_inds = np.unique(np.concatenate([A, B]))
items = by_db.iloc[by_inds]
# get a dictionary whose keys are the 'by' indices
features = get_features(items)
dis = np.empty(shape=(n_pairs, 1))
# FIXME: second dim is 1 because of the way it is stored to disk,
# but ultimately it shouldn't be necessary anymore
# (if using axis arg in np2h5, h52np and h5io...)
for i in range(n_pairs):
dataA = features[pairs[i, 0]]
dataB = features[pairs[i, 1]]
if dataA.shape[0] == 0:
warnings.warn('No features found for file {}, {} - {}'
.format(items['file'][pairs[i, 0]],
items['onset'][pairs[i, 0]],
items['offset'][pairs[i, 0]]),
UserWarning)
if dataB.shape[0] == 0:
warnings.warn('No features found for file {}, {} - {}'
.format(items['file'][pairs[i, 1]],
items['onset'][pairs[i, 1]],
items['offset'][pairs[i, 1]]),
UserWarning)
try:
dis[i, 0] = distance(dataA, dataB)
except:
sys.stderr.write(
'Error when calculating the distance between item {}, {} - {} '
'and item {}, {} - {}\n'
.format(items['file'][pairs[i, 0]],
items['onset'][pairs[i, 0]],
items['offset'][pairs[i, 0]],
items['file'][pairs[i, 0]],
items['onset'][pairs[i, 0]],
items['offset'][pairs[i, 0]]),
)
raise
if synchronize:
distance_file_lock.acquire()
with h5py.File(distance_file) as fh:
fh['distances/' + by][start:stop, :] = dis
if synchronize:
distance_file_lock.release()
def launch_lsh(features_file, featsdir, S=64, files=None, with_vad=None,
split=False):
"""Launch lsh for specified features, return a dictionnary containing for
all files the path to their signature, features and vad.
Parameters:
----------
features_file: h5features file name
featsdir: output folder where the features will be written
S: int, number of bits of the lsh signature
files: list, only launch lsh on the specified files.
(must be the basename of the file, like in the h5features file)
with_vad: optional VAD file
"""
if not os.path.isfile(features_file):
raise ValueError('file {} doesn\'t exist'.format(features_file))
if S not in VALID_BITS:
raise ValueError('S={} must be 32 or 64'.format(S))
def aux(f, feats, S, D, featfile, sigfile, vadfile=None, vad=None):
with open(featfile, 'wb') as fout:
fout.write(feats.tobytes())
# lsh creates .sig file
command_ = '{}/lsh -S {} -D {} -projfile proj_b{}xd{}_seed1 -featfile {} -sigfile {}'
command_ = command_.format(binpath, S, D, S, D, featfile, sigfile)
if vadfile:
with open(vadfile, 'w') as fout:
for interval in vad[f]:
fout.write(' '.join(map(str, interval)) + '\n')
command_ += ' -vadfile {}'.format(vadfile)
p_ = Popen(command_, shell=True, stdout=PIPE, stderr=PIPE)
output_, err_ = p_.communicate()
if 'usage' in err_:
print(err_)
raise NameError('Cannot run command: {}'.format(command_))
if os.stat(sigfile).st_size == 0: # warn lsh without results
warnings.warn('no results from lsh ({} empty)'.format(sigfile))
vad = {}
if with_vad:
vad_ = defaultdict(list)
with open(with_vad) as fin:
for line in fin:
fname, start, end = line.strip().split()
start, end = map(lambda t: int(float(t) * 100), (start, end))
vad_[fname].append((start, end))
vad = dict(vad_)
# generate a file with DxS normal random values [=numpy.random.norm(0.0,1.0,D*S)]
# TODO: can be different the output file from genproj?
D = h5py.File(features_file)['features']['features'].shape[1]
proj_f_ = 'proj_b{}xd{}_seed1'.format(S, D)
command_ = '{}/genproj -S {} -D {} -seed 1'.format(binpath, S, D)
p_ = Popen(command_, shell=True, stdout=PIPE, stderr=PIPE)
output_, err_ = p_.communicate()
if 'unknown' in err_:
print(err_)
raise NameError('Cannot run command: {}'.format(command_))
if os.stat(proj_f_).st_size == 0:
warnings.warn('genproj didn\'t create file {} or empty'.format(proj_f_))
res = fdict()
res.stats = {'S':S, 'D':D}
if not os.path.exists(featsdir):
try:
os.makedirs(featsdir)
except:
pass
if files == None:
files = h5features.read(features_file)[0].keys()
for f in files:
spk = get_speaker(f)
if not split:
sigfile = os.path.join(featsdir, f + ".sig")
vadfile = os.path.join(featsdir, f + ".vad")
featfile = os.path.join(featsdir, f + ".fea")
feats = h5features.read(features_file, from_item=f)[1][f]
if with_vad:
aux(f, feats, S, D, featfile, sigfile, vadfile, vad)
else:
aux(f, feats, S, D, featfile, sigfile)
try :
res[spk][f] = {'sig': sigfile, 'fea': featfile}
except KeyError:
res[spk] = {f: {'sig': sigfile, 'fea': featfile}}
if with_vad:
res[spk][f]['vad'] = vadfile
else:
intervals = vad[f]
for i, (start, end) in enumerate(intervals):
fi = '{}_{}'.format(f, i)
sigfile = os.path.join(featsdir, fi + ".sig")
vadfile = os.path.join(featsdir, fi + ".vad")
featfile = os.path.join(featsdir, fi + ".fea")
feats = h5features.read(
features_file, from_item=f,
from_time=float(start)/100, to_time=float(end)/100)[1][f]
aux(fi, feats, S, D, featfile, sigfile)
try:
res[spk][fi] = {'sig': sigfile, 'fea': featfile}
except KeyError:
res[spk] = {fi: {'sig': sigfile, 'fea': featfile}}
return res
def get_features_from_file(self, fileid):
"""return the features accosiated to a file"""
return h5features.read(self.feature_file, from_internal_file=fileid,
index=self.index)[1][fileid]
def get_features_from_file(self, fileid):
"""return the features associated to a file"""
return h5features.read(
self.feature_file, from_item=fileid)[1][fileid]
