def index_train(feat_filename, index_filename, index_type="IVF1024,Flat"):
print('Loading feature data...')
max_feats = 10000
feats, utt_ids = kaldi_io.readArk(feat_filename, limit=max_feats)
complete_feat_len = 0
for feat in feats:
complete_feat_len += feat.shape[0]
utt_map = np.zeros(complete_feat_len, dtype=np.int32)
pos_map = np.zeros(complete_feat_len, dtype=np.int32)
# create the utt_map and pos_map that map a pos id to utt id and position inside an utterance
pos = 0
for i, (feat, utt_id) in enumerate(zip(feats, utt_ids)):
for j in range(feat.shape[0]):
utt_map[pos] = i
pos_map[pos] = j
pos += 1
complete_feats = np.concatenate(feats, axis=0)
print('complete_feats:', complete_feats)
print('complete_feats.min():', complete_feats.min())
print('complete_feats.max():', complete_feats.max())
print('complete_feats.sum()', complete_feats.sum())
print('Going to index features of shape:', complete_feats.shape)
index = faiss.index_factory(complete_feats.shape[1],
index_type.encode("ascii"))
index.train(complete_feats)
index.add(complete_feats)
print('is_trained:', index.is_trained)
print('ntotal:', index.ntotal)
print('Indexing finished.')
search_vec_index = 100
search_vec = complete_feats[search_vec_index]
search_vec = search_vec.reshape((1, len(search_vec)))
print("Shape of search_vec:", search_vec.shape)
D, I = index.search(search_vec, 50)
print("Neighboors of ", search_vec_index)
print("D:")
print(D)
print("I:")
print(I)
def visualize_kaldi_bin_feats(feat_filename,
max_frames,
num_feat=0,
phn_file='',
phn_offset=5,
wav_file='',
do_tsne=False):
feats, utt_ids = kaldi_io.readArk(feat_filename, limit=10000)
print([feat.shape for feat in feats], utt_ids)
print('showing features for utt_id:', utt_ids[num_feat])
print('min vector:')
print(np.min(feats[num_feat], axis=1))
print('max vector:')
print(np.max(feats[num_feat], axis=1))
print('sum vector:')
print(np.sum(feats[num_feat], axis=1))
print(feats[num_feat].shape)
if phn_file == '':
plt.matshow(feats[num_feat][:max_frames].T)
plt.show()
else:
plt.matshow(feats[num_feat][:max_frames].T)
positions, names = utils.loadPhnFile(phn_file)
xpositions = [
float(pos[1]) / samples_per_frame - phn_offset for pos in positions
if float(pos[1]) / samples_per_frame < max_frames
]
for xc in xpositions:
plt.axvline(x=xc, color='k', linestyle='--')
plt.show()
if do_tsne:
plt.figure(1)
print('Calculating TSNE:')
model = TSNE(n_components=2, random_state=0)
tsne_data = model.fit_transform(feats[num_feat])
plt.plot(tsne_data[:, 0], tsne_data[:, 1], '--')
print('Now showing tsne plot:')
plt.show()
def cluster_rnn_phn(n_clusters,
wav_files,
ark_file,
hopping_size,
window_size,
subsample,
n_jobs=4):
feats, uttids = kaldi_io.readArk(ark_file)
tf.segment_mean()
#tf.
#from https://github.com/tensorflow/tensorflow/issues/7389
ones = tf.ones_like(x)
count = tf.unsorted_segment_sum(ones, ids, 2)
sums = tf.unsorted_segment_sum(x, ids, 2)
mean = tf.divide(sums, count)
def load_feats_flat(ark_file):
feats, uttids = kaldi_io.readArk(ark_file)
# preallocate the array and establish array sizes
inner_dim = feats[0][0].shape[1]
sum_len = feats2sumlen(feats)
feats_flat = np.zeros(sum_len, inner_dim)
uttids_flat = []
pos_flat = []
pos = 0
for uttid, feat in zip(feats):
feats_flat[pos:pos + feat.shape[0]] = feat
pos += feat.shape[0]
#repeating uttid feat.shape[0] times
uttids_flat += [uttid] * feat.shape[0]
pos_flat = np.array(float(x)
for x in range(feat.shape[0])) * hopping_size
return feats, feats_flat, uttids_flat, pos_flat
def visualize_stats(feat_filename,
max_feats,
abs_feats=True,
reverse_sort=True):
feats, utt_ids = kaldi_io.readArk(feat_filename, limit=max_feats)
feats_len = len(feats)
print("Loaded:" + str(feats_len) + "feats")
sums = []
for feat in feats:
if abs_feats:
feat = np.abs(feat)
local_sum = np.sum(feat, axis=0) / float(len(feat))
print(local_sum.shape)
sums.append(local_sum)
sums = np.stack(sums, axis=0)
print(sums.shape)
finalsum = np.sum(sums, axis=0) / float(feats_len)
finalsum_sorted = np.sort(np.array(finalsum))
if reverse_sort:
finalsum_sorted = finalsum_sorted[::-1]
print(finalsum)
print(finalsum_sorted)
plt.plot(finalsum_sorted)
plt.figure(1)
plt.matshow([finalsum])
plt.figure(2)
plt.matshow([finalsum_sorted])
plt.show()
def get_vectors():
# possible parameters
# feat_file -> path to feat_file, must be one returned by /list_avail_reps
# half_index -> cut vectors at this position (optional, default: -1)
# limit -> max vectors to return (optional, numeric)
# average_utts -> average vector for each utterance (optional, default: True)
# normalize -> normalize vectors to unit length (optional, default: False)
# Reading parameters from POST request:
if 'feat_file' in flask.request.form:
feat_filename = flask.request.form['feat_file']
else:
response_str = json.dumps({
'status':
'fail',
'reason':
'You must supply a feat_file for /get_vectors'
})
response = Response(response_str, mimetype='application/json')
return response
if 'half_index' in flask.request.form:
half_index = int(flask.request.form['half_index'])
else:
print(
'POST /get_vectors called without half_index parameter, setting to default -1 (disable)'
)
half_index = -1
if 'limit' in flask.request.form:
limit = int(flask.request.form['limit'])
else:
print(
'POST /get_vectors called without limit parameter, setting to default -1 (disable)'
)
limit = -1
if 'average_utts' in flask.request.form:
average_utts = flask.request.form['average_utts']
else:
print(
'POST /get_vectors called without average_utts parameter, setting to default true (enable)'
)
average_utts = True
if not average_utts or average_utts == 'False' or average_utts == 'false':
if 'stride' in flask.request.form:
stride = int(flask.request.form['stride'])
else:
print(
'POST /get_vectors called with average_utts = False, but stride parameter is not set, setting it to the default value (1)'
)
stride = 1
normalize = ('normalize' in flask.request.form)
feats, utt_ids = kaldi_io.readArk(feat_filename, limit=limit)
feats_len = len(feats)
assert (len(utt_ids) == len(feats))
print("Loaded:" + str(feats_len) + " feats.")
if average_utts or average_utts == 'True' or average_utts == 'true':
feats = [feat.mean(0) for feat in feats]
if half_index != -1:
print('Cutting vectors at ', half_index,
'and normalize to unit length' if normalize else '')
feats = [
feat[:half_index] /
(np.linalg.norm(feat[:half_index]) if normalize else 1.0)
for feat in feats
]
else:
if normalize:
print('Normalize to unit length.')
feats = [feat / np.linalg.norm(feat) for feat in feats]
response_vec_dict = {}
for utt_id, feat in zip(utt_ids, feats):
response_vec_dict[utt_id] = feat.tolist()
response_str = json.dumps({
'status': 'success',
'vectors': response_vec_dict
})
else:
if stride != 1:
feats = [feat[::stride] for feat in feats]
if half_index != -1:
print('Cutting vectors at ', half_index)
print('Not yet supported')
if normalize:
feats = [(feat.T / np.linalg.norm(feat, axis=1)).T
for feat in feats]
response_vec_dict = {}
for utt_id, feat in zip(utt_ids, feats):
response_vec_dict[utt_id] = feat.tolist()
response_str = json.dumps({
'status': 'success',
'vectors': response_vec_dict
})
response = Response(response_str, mimetype='application/json')
return response
def visualize_classes_tsne(feat_filename,
utt_2_class_filename,
half_index=-1,
normalize=True,
class_mean_vector=False):
feats, utt_ids = kaldi_io.readArk(feat_filename, limit=25000)
feats_len = len(feats)
assert (len(utt_ids) == len(feats))
print("Loaded:" + str(feats_len) + " feats.")
feats = [feat.mean(0) for feat in feats]
if half_index != -1:
print('Cutting vectors at ', half_index,
'and normalize to unit length' if normalize else '')
feats = [
feat[:half_index] /
(np.linalg.norm(feat[:half_index]) if normalize else 1.0)
for feat in feats
]
else:
if normalize:
print('Normalize to unit length.')
feats = [feat / np.linalg.norm(feat) for feat in feats]
utt_2_class = utils.loadUtt2Spk(utt_2_class_filename)
ground_truth_utt_2_class = [
utt_2_class[utt_id] for utt_id in utt_ids if utt_id in utt_2_class
]
utt_ids_filtered = [utt_id for utt_id in utt_ids if utt_id in utt_2_class]
#feats_filtered = [feat for feat,utt_id in zip(feats, utt_ids) if utt_id in utt_2_class]
assert (len(ground_truth_utt_2_class) == len(utt_ids_filtered))
#assert(len(utt_ids_filtered) == len(feats_filtered) )
dataset = {}
for feat, utt in zip(feats, utt_ids):
if utt in utt_2_class:
dataset[utt] = feat
myclass_2_utt = {}
myclass_2_samples = {}
for myclass in set(ground_truth_utt_2_class):
my_class_filtered_utts = [
utt_id for utt_id, gd_class in zip(utt_ids_filtered,
ground_truth_utt_2_class)
if gd_class == myclass
]
if len(my_class_filtered_utts) > 100:
myclass_2_utt[myclass] = my_class_filtered_utts
myclass_2_samples[myclass] = random.sample(
myclass_2_utt[myclass], min(1000, len(myclass_2_utt[myclass])))
feats_samples = []
feats_samples_classes = []
if class_mean_vector:
for myclass in myclass_2_samples:
feats_samples += [
np.vstack(dataset[utt]
for utt in myclass_2_samples[myclass]).mean(0)
]
feats_samples_classes += [myclass]
else:
for myclass in myclass_2_samples:
feats_samples += [
dataset[utt] for utt in myclass_2_samples[myclass]
]
feats_samples_classes += [myclass] * len(
myclass_2_samples[myclass])
print('Added', len(myclass_2_samples[myclass]), 'entries for',
myclass)
print([
utt.replace('train-sample', 'train/sample') + '.mp3'
for utt in myclass_2_samples[myclass]
])
class_2_num = dict([(a, b)
for b, a in enumerate(list(myclass_2_samples.keys()))])
print(class_2_num)
feats_samples_classes_num = [
class_2_num[myclass] for myclass in feats_samples_classes
]
#print(feats_samples_classes_num)
num_classes = max(feats_samples_classes_num)
print('Num classes=', num_classes)
print(feats_samples)
print('shape:', feats_samples[0].shape)
print('Calculating TSNE:')
model = TSNE(n_components=2, random_state=0, metric='euclidean')
tsne_data = model.fit_transform(np.vstack(feats_samples))
#model = TSNE(n_components=2, random_state=0, metric='cosine')
#tsne_data = model.fit_transform([feat[100:] for feat in feats])
colormap = plt.cm.gist_ncar #nipy_spectral, Set1,Paired
colorst = colormap(np.linspace(
0, 0.9, num_classes +
1)) #[colormap(i) for i in np.linspace(0, 0.9, num_speakers)]
cs = [
colorst[feats_samples_classes_num[i]]
for i in range(len(feats_samples_classes_num))
]
#print(tsne_data[:,0])
#print(tsne_data[:,1])
plt.scatter(tsne_data[:, 0], tsne_data[:, 1], color=cs)
#for i,elem in enumerate(tsne_data):
# print(cs[0])
# print(ground_truth_utt_2_spk[0])
# plt.scatter(elem[0], elem[1], color=cs[i], label=ground_truth_utt_2_spk[i])
plt.legend()
# for i in range(tsne_data.shape[0]):
# plt.text(tsne_data[i,0], tsne_data[i,1], uttids[i], fontsize=8, color=cs[i])
print('Now showing tsne plot:')
plt.show()
def convert(alignments, spectrograms):
input_dir = os.path.abspath(
os.path.dirname(os.path.abspath(__file__)) +
'/../dat/speech_tokenizer')
output_dir = os.path.abspath(
os.path.dirname(os.path.abspath(__file__)) + '/../dat/fast_load')
utterances_output_dir = output_dir + '/utterances'
if not os.path.exists(utterances_output_dir):
os.makedirs(utterances_output_dir)
if alignments:
print('Storing tag information...', end='', flush=True)
tags = []
with open(input_dir + '/new_alignments/phones.txt',
'r') as file_handle:
for i, line in enumerate(file_handle):
tag, id = line.strip('\n').split(' ')
assert i == int(id)
tag = tag.split('_')
tag = (tag[1] if len(tag) > 1 else None, tag[0])
tags.append(tag)
with open(output_dir + '/tags.json', 'w') as file_handle:
json.dump(tags, file_handle)
print(' DONE')
print('Converting alignments...', end='', flush=True)
tag_dict = dict()
with open(input_dir + '/new_alignments/merged_alignment.txt',
'r') as file_handle:
for line in file_handle:
# Get the data on one tag
line = line.strip('\n').split(' ')
# Convert starts and durations from seconds to numbers of frames without risking floating point errors
assert len(re.sub('.*\.', '', line[2])) == 3
assert len(re.sub('.*\.', '', line[3])) == 3
id, start, duration, tag = line[0], int(
re.sub('\.', '', line[2][:-1])), int(
re.sub('\.', '', line[3][:-1])), int(line[4])
if id not in tag_dict:
tag_dict[id] = []
tag_dict[id].append((start, duration, tag))
n_ids = len(tag_dict.keys())
print(' DONE')
start_time = time()
for i, (key, value) in enumerate(tag_dict.items()):
progress.print_bar(i, n_ids, 20, 'Storing alignment data... ┃',
'┃')
with open(utterances_output_dir + '/%s.json' % (key, ),
'w') as file_handle:
# file_dict = {'id' : key, 'alignments' : value}
json.dump(value, file_handle)
progress.print_bar(i + 1, n_ids, 20, 'Storing alignment data... ┃',
'┃ DONE %.4fs' % (time() - start_time))
print('Storing sequence IDs...', end='', flush=True)
with open(output_dir + '/utterances.json', 'w') as file_handle:
json.dump(
sorted(tag_dict), file_handle, indent=4
) # Asserts the same IDs in both alignments and spectrograms
print(' DONE')
if spectrograms:
# Convert spectrograms
print('Loading spectrogram data (this may take some time)...',
end='',
flush=True)
feats, ids = readArk(
input_dir + '/TEDLIUM_fbank_train_cleaned/unnormalized.feats.ark')
n_ids = len(ids)
print(' DONE')
start_time = time()
for i, (feat, id) in enumerate(zip(feats, ids)):
progress.print_bar(i, n_ids, 20, 'Storing spectrogram data... ┃',
'┃')
np.save(utterances_output_dir + '/%s.npy' % (id, ), feat)
progress.print_bar(i + 1, n_ids, 20, 'Storing spectrogram data... ┃',
'┃ DONE %.4fs' % (time() - start_time))
def cluster_speaker(ark_file,
cluster_algo='HDBSCAN',
half_index=-1,
dbscan_eps=0.0005,
dbscan_min_samples=3,
min_cluster_sizes_str="5",
min_samples_str="3",
utt_2_spk=None,
output_utt_2_spk=None,
fileset='dev',
tsne_viz=False,
n_jobs=4,
db_scan_range_search=False,
hdb_scan_range_search=False,
normalize=True,
do_save_result=True,
use_gpu=False):
postfix = ''
print('Loading feats now:')
feats, uttids = kaldi_io.readArk(ark_file.replace('%set', fileset))
print('feat[0] shape: ', feats[0].shape)
#feats = np.vstack([pairwise_normalize(feat[0]) for feat in feats])
print('Generating mean vector.')
feats = np.vstack([feat.mean(0) for utt, feat in zip(uttids, feats)])
if half_index != -1:
print('Cutting vectors at ', half_index,
'and normalize to unit length' if normalize else '')
feats = np.vstack([
feat[half_index:] /
(np.linalg.norm(feat[half_index:]) if normalize else 1.0)
for feat in feats
])
else:
if normalize:
print('Normalize to unit length.')
feats = np.vstack([feat / np.linalg.norm(feat) for feat in feats])
print('Done. feats shape:', feats.shape)
# feats = np.vstack([feat[0] for utt,feat in zip(uttids,feats) if 'AlGore' not in utt])
# uttids = [utt for utt in uttids if 'AlGore' not in utt]
print('feats shape:', feats.shape)
print('feat[0] shape: ', feats[0].shape)
print('halfindex:', half_index)
# print('some distances:')
# for a,b in [(random.randint(0, len(feats)-1), random.randint(0, len(feats)-1)) for i in range(10)] + [(0,0)]:
# dst = distance.euclidean(feats[a],feats[b])
# print('euc dst:', a,b,'=',dst)
# dst = distance.cosine(feats[a],feats[b])
# print('cos dst:', a,b,'=',dst)
# dst = np.dot(feats[a],feats[b])
# print('dot dst:', a,b,'=',dst)
#
# dst = pos_neg_dot_distance(feats[a],feats[b])
# print('pos_neg_dot_distance dst:', a,b,'=',dst)
#
#
# dst = pairwise_pos_neg_dot_distance(feats[a],feats[b])
# print('pairwise_pos_neg_dot_distance dst:', a,b,'=',dst)
#
# for a in range(10):
# for b in range(10):
# print('feats[a]:',feats[a])
# print('feats[b]:',feats[b])
# dst = pos_neg_dot_distance(feats[a],feats[b])
# print('pos_neg_dot_distance dst:', a,b,'=',dst)
# pairwise_pos_neg_dot_distance(feats[a],feats[b])
# print('pairwise_pos_neg_dot_distance dst:', a,b,'=',dst)
ground_truth_utt_2_spk, ground_truth_utt_2_spk_int = None, None
if utt_2_spk is not None and utt_2_spk.lower(
) != 'none' and utt_2_spk.strip() != '':
utt_2_spk = utils.loadUtt2Spk(utt_2_spk.replace('%set', fileset))
ground_truth_utt_2_spk = [utt_2_spk[utt_id] for utt_id in uttids]
le = preprocessing.LabelEncoder()
le.fit(ground_truth_utt_2_spk)
ground_truth_utt_2_spk_int = le.transform(ground_truth_utt_2_spk)
print("Ground truth speaker classes available:")
print(ground_truth_utt_2_spk_int)
print('Now running DBSCAN clustering on', len(uttids), 'entries.')
bestARI = 0.0
bestConf = {}
if db_scan_range_search:
eps_range = [x / 100.0 for x in range(1, 100)]
min_samples_range = [1, 2, 3, 4, 5, 6, 7, 8, 10, 20, 30, 50, 100]
result_mat = np.zeros((len(eps_range), len(min_samples_range)))
print('shape result mat:', result_mat.shape)
for i_eps, dbscan_eps in enumerate(eps_range):
for i_min_samples, dbscan_min_samples in enumerate(
min_samples_range):
dbscan_algo = DBSCAN(eps=dbscan_eps,
min_samples=dbscan_min_samples,
metric=pairwise_pos_neg_dot_distance,
n_jobs=1)
clustering = dbscan_algo.fit(feats)
clustering_labels = list(clustering.labels_)
print('dbscan_eps', dbscan_eps, 'dbscan_min_samples',
dbscan_min_samples)
print('num clusters:', len(set(clustering_labels)))
ARI = metrics.adjusted_rand_score(ground_truth_utt_2_spk_int,
clustering_labels)
result_mat[i_eps][i_min_samples] = float(ARI)
print('ARI:', ARI)
if ARI > bestARI:
print('Found new best conf:', ARI)
bestConf = {
'eps': dbscan_eps,
'min_samples': dbscan_min_samples
}
bestARI = ARI
plt.matshow(result_mat)
plt.show()
np.save(ark_file + '.dbrangescan_cluster_ARI' + postfix, result_mat)
print('bestARI:', bestARI)
print('bestConf:', bestConf)
min_cluster_sizes = [int(x) for x in min_cluster_sizes_str.split(',')]
min_samples = [int(x) for x in min_samples_str.split(',')]
result_mat = np.zeros((len(min_cluster_sizes), len(min_samples)))
result_mat_outliers = np.zeros_like(result_mat)
result_mat_n = np.zeros_like(result_mat)
best_pairwise_f1 = 0.0
bestConf = {}
# previous good config: min_cluster_size=5, min_samples=3
for i, min_cluster_size in enumerate(min_cluster_sizes):
for j, min_sample in enumerate(min_samples):
feat_key = ark_file.split('/')[-3] + '_' + str(
min_cluster_size) + '_' + str(min_sample)
if do_save_result:
save_result(feat_key, "cl_size", str(min_cluster_size))
save_result(feat_key, "min_s", str(min_sample))
if cluster_algo == 'HDBSCAN':
print('Running HDBSCAN with min_cluster_size',
min_cluster_size, 'min_samples', dbscan_min_samples)
cluster_algo = HDBSCAN(min_cluster_size=min_cluster_size,
min_samples=min_sample,
metric='euclidean',
algorithm='best',
core_dist_n_jobs=28)
elif cluster_algo == 'DBSCAN':
print('Running DBSCAN with dbscan_eps', dpscan_eps,
'dbscan_min_samples', dbscan_min_samples)
cluster_algo = DBSCAN(eps=dbscan_eps,
min_samples=dbscan_min_samples,
metric='euclidean',
n_jobs=28)
elif cluster_algo == 'kmeans':
print(
'kmeans clustering not available for speaker clustering yet. Exiting.'
)
sys.exit(-1)
else:
print('cluster_algo:', cluster_algo, 'not supported. Exiting.')
clustering = cluster_algo.fit(feats)
clustering_labels = list(clustering.labels_)
print('Num of clusters (as determined by density clustering):',
len(set(clustering_labels)))
print(clustering_labels)
sys.stdout.flush()
#print('Numpy bincount of the clustering:', np.bincount(clustering))
number_format = "%.4f"
clustering_labels1 = clustering_labels
clustering_labels2 = []
num_outliers = -1
for elem in clustering_labels1:
if elem == -1:
clustering_labels2.append(num_outliers)
num_outliers -= 1
else:
clustering_labels2.append(elem)
num_outliers = (num_outliers + 1) * -1
if utt_2_spk is not None and utt_2_spk.lower(
) != 'none' and utt_2_spk.strip() != '':
if do_save_result:
save_result(feat_key, 'outliers_' + fileset,
str(num_outliers))
save_result(feat_key, 'clusters_' + fileset,
str(len(set(clustering_labels))))
print(
'Number of outliers:', num_outliers, '(', number_format %
(float(num_outliers) * 100.0 / float(len(uttids))), '%)')
#This would compute scores with all outliers in the same cluster:
#ARI = metrics.adjusted_rand_score(ground_truth_utt_2_spk_int, clustering_labels)
#print('ARI score:', number_format % ARI)
#vmeasure = metrics.v_measure_score(ground_truth_utt_2_spk_int, clustering_labels)
#print('V-measure:', number_format % vmeasure)
ARI2 = metrics.adjusted_rand_score(ground_truth_utt_2_spk_int,
clustering_labels2)
print('ARI score (each outlier its own cluster):',
number_format % ARI2)
vmeasure2 = metrics.v_measure_score(ground_truth_utt_2_spk_int,
clustering_labels2)
print('NMI / V-measure (each outlier its own cluster):',
number_format % vmeasure2)
if do_save_result:
save_result(feat_key, 'ARI_' + fileset,
number_format % ARI2)
save_result(feat_key, 'NMI_' + fileset,
number_format % vmeasure2)
print('Calculating pairwise recall:')
cluster_pairwise = pdist(
np.asarray(clustering_labels2)[:, np.newaxis],
metric='chebyshev') < 1
groundtruth_pairwise = pdist(
np.asarray(ground_truth_utt_2_spk_int)[:, np.newaxis],
metric='chebyshev') < 1
#scitkits recall_score and precision_score is slow unfortunatly
#pairwise_recall = metrics.recall_score(groundtruth_pairwise, cluster_pairwise , pos_label=True, average='binary')
#pairwise_precision = metrics.precision_score(groundtruth_pairwise, cluster_pairwise , pos_label=True, average='binary')
#print('scikit learn recall / precision:', pairwise_recall, pairwise_precision)
# efficient binary comparision, since the pairwise matrix can be huge for large n
tp = np.sum(
np.bitwise_and(groundtruth_pairwise, cluster_pairwise))
fp = np.sum(
np.bitwise_and(np.invert(groundtruth_pairwise),
cluster_pairwise))
fn = np.sum(
np.bitwise_and(groundtruth_pairwise,
np.invert(cluster_pairwise)))
pairwise_precision = tp / (tp + fp)
pairwise_recall = tp / (tp + fn)
pairwise_f1 = 2.0 * pairwise_recall * pairwise_precision / (
pairwise_recall + pairwise_precision)
print(
'pairwise recall / precision / f1-score (each outlier its own cluster):',
number_format % pairwise_recall,
number_format % pairwise_precision,
number_format % pairwise_f1)
if do_save_result:
save_result(feat_key, 'recall_' + fileset,
number_format % pairwise_recall)
save_result(feat_key, 'prec_' + fileset,
number_format % pairwise_precision)
save_result(feat_key, 'f1_' + fileset,
number_format % pairwise_f1)
if pairwise_f1 > best_pairwise_f1:
print('Found new best pairwise f1:', pairwise_f1)
bestConf = {
'min_cluster_size': min_cluster_size,
'min_sample': min_sample,
'n': len(set(clustering_labels)),
'outliers': num_outliers
}
best_pairwise_f1 = pairwise_f1
result_mat[i][j] = float(pairwise_f1)
result_mat_outliers[i][j] = num_outliers
result_mat_n[i][j] = len(set(clustering_labels))
#print('pairwise recall / precision / f1-score:', number_format % pairwise_recall, number_format % pairwise_precision, number_format % pairwise_f1)
print('Clustering predicted classes:',
len(set(clustering_labels)))
print('Ground truth classes',
len(set(ground_truth_utt_2_spk_int)))
# if len(min_cluster_sizes) > 1 or len(min_samples) > 1:
# np.save(ark_file + '.hdbrangescan_cluster_f1' + postfix, result_mat)
#
# print('best f1:', best_pairwise_f1)
# print(bestConf)
#
# print('f1 scores:')
# plt.matshow(result_mat)
# plt.show()
#
# print('num outliers')
# plt.matshow(result_mat_outliers)
# plt.show()
#
# print('n')
# plt.matshow(result_mat_n)
# plt.show()
if tsne_viz:
print('Calculating TSNE:')
model = TSNE(n_components=2, random_state=0, metric='euclidean')
tsne_data = model.fit_transform(feats)
#model = TSNE(n_components=2, random_state=0, metric='cosine')
#tsne_data = model.fit_transform([feat[100:] for feat in feats])
if utt_2_spk is not None and utt_2_spk.lower(
) != 'none' and utt_2_spk.strip() != '':
num_speakers = max(ground_truth_utt_2_spk_int) + 1
else:
num_speakers = len(set(clustering_labels))
colormap = plt.cm.gist_ncar #nipy_spectral, Set1,Paired
colorst = colormap(
np.linspace(0, 0.9, num_speakers)
) #[colormap(i) for i in np.linspace(0, 0.9, num_speakers)]
if utt_2_spk is not None and utt_2_spk.lower(
) != 'none' and utt_2_spk.strip() != '':
cs = [
colorst[ground_truth_utt_2_spk_int[i]]
for i in range(len(clustering_labels))
]
else:
cs = [
colorst[clustering_labels[i]]
for i in range(len(clustering_labels))
]
#print(tsne_data[:,0])
#print(tsne_data[:,1])
plt.scatter(tsne_data[:, 0], tsne_data[:, 1], color=cs)
#for i,elem in enumerate(tsne_data):
# print(cs[0])
# print(ground_truth_utt_2_spk[0])
# plt.scatter(elem[0], elem[1], color=cs[i], label=ground_truth_utt_2_spk[i])
plt.legend()
# for i in range(tsne_data.shape[0]):
# plt.text(tsne_data[i,0], tsne_data[i,1], uttids[i], fontsize=8, color=cs[i])
print('Now showing tsne plot:')
plt.show()
if output_utt_2_spk is not None and output_utt_2_spk.lower(
) != 'none' and output_utt_2_spk.strip() != '':
if len(min_cluster_sizes) > 1 or len(min_samples) > 1:
print(
'Not saving clustering result, since we searched a full range. Rerun with a single min_cluster_size and min_samples parameter.'
)
else:
output_utt_2_spk = output_utt_2_spk.replace(
'%minclustersize', str(min_cluster_size))
output_utt_2_spk = output_utt_2_spk.replace(
'%minsample', str(min_sample))
output_utt_2_spk = output_utt_2_spk.replace('%set', fileset)
featstr = ark_file.split('/')[-3]
featstr = featstr.replace(
'featinput_unnormalized.feats.ark_dot_combine_tied_embs',
'std_end_conf').replace('feats_', '')
print('featstr:', featstr)
output_utt_2_spk = output_utt_2_spk.replace('%feat', featstr)
output_utt_2_spk += ('_l2norm' if normalize else '')
#output_utt_2_spk += postfix
print('Saving result to:', output_utt_2_spk)
with open(output_utt_2_spk, 'w') as output_utt_2_spk_out:
for utt, label in zip(uttids, clustering_labels2):
output_utt_2_spk_out.write(utt +
(' spk%07d' %
label).replace('-', 'o') +
'\n')
def corr_phn(ark_file,
alignment_dir,
fileset='train',
limit=5000,
cmudict_sort=True):
print('len cmudict arpabet:',
len(cmudict_silence + cmudict_vowels + cmudict_consonants))
feats, uttids = kaldi_io.readArk(ark_file.replace('%set', fileset),
limit=limit)
phones_txt = alignment_dir + 'phones.txt'
num2phone_orig = load_phones_txt(phones_txt)
#phone2num_orig =
print(num2phone_orig)
phoneset = list(set(num2phone_orig.values()))
not_vowels = []
for elem in phoneset:
if elem not in cmudict_vowels:
not_vowels += [elem]
for elem in cmudict_vowels:
if elem not in phoneset:
print('vowel NA:', elem)
print(not_vowels)
print('Loaded num2phone_orig for: ', len(set(num2phone_orig.values())),
'phones')
if cmudict_sort:
num2phone = dict(
enumerate(cmudict_silence + cmudict_vowels + cmudict_consonants))
phone2num = dict([
(y, x) for x, y in enumerate(cmudict_silence + cmudict_vowels +
cmudict_consonants)
])
else:
num2phone = dict(enumerate(sorted(list(set(num2phone_orig.values())))))
phone2num = dict([
(y, x)
for x, y in enumerate(sorted(list(set(num2phone_orig.values()))))
])
num2phoneid = dict([(i, phone2num[x]) for i, x in num2phone_orig.items()])
print(num2phoneid)
ctm_file = alignment_dir + 'all.ctm'
all_ctm_db = load_all_ctm_pos(ctm_file, num2phoneid, limit=limit)
phone_len = len(num2phone.keys())
feat_dim = len(feats[0][0])
print(feat_dim)
phone_mutual_info_classifs = []
do_2d = True
parallel = True
if do_2d:
if not parallel:
for phone in sorted(num2phone.keys()):
#for feat_num in range(feat_dim):
phone_mutual_info_classif = get_phone_mutual_info(
phone, feats, uttids, all_ctm_db)
phone_mutual_info_classifs += [phone_mutual_info_classif]
phone_mutual_info_classifs = np.vstack(phone_mutual_info_classifs)
else:
pool = multiprocessing.Pool(28)
get_phone_mutual_info_partial = partial(get_phone_mutual_info,
feats=feats,
uttids=uttids,
all_ctm_db=all_ctm_db)
phone_mutual_info_classifs = pool.map(
get_phone_mutual_info_partial, sorted(num2phone.keys()))
#for feat_num in range(feat_dim):
else:
feat_vars = []
phone_vars = []
for feat, uttid in zip(feats, uttids):
ctm = all_ctm_db[uttid]
#print('lens (ctm|feat):',len(ctm), len(feat))
if len(ctm) > len(feat):
print(uttid,
'warning, ctm is slightly longer than feature len:',
len(ctm), 'vs.', len(feat))
ctm = ctm[:len(feat)]
elif len(ctm) < len(feat):
print(uttid,
'warning, ctm is slightly shorter than feature len:',
len(ctm), 'vs.', len(feat))
assert (len(feat) == len(ctm))
feat_vars += [feat]
#phone_var = #(np.array(ctm) == phone) * 1.0
phone_vars += [ctm]
feat_vars = np.vstack(feat_vars)
phone_vars = np.hstack(phone_vars)
print(feat_vars.shape)
print(phone_vars.shape)
phone_mutual_info_classifs = sklearn.feature_selection.mutual_info_classif(
feat_vars, phone_vars)
print(phone_mutual_info_classifs)
print(plt.rcParams["figure.figsize"])
fig = plt.figure(figsize=(11, 7))
ax = fig.add_subplot(111)
phone_mutual_info_classifs[0] *= 0.7
ax.matshow(phone_mutual_info_classifs, aspect='auto')
plt.yticks(sorted(num2phone.keys()),
[num2phone[x] for x in sorted(num2phone.keys())])
plt.show()
def tsne(ark_file,
alignment_dir,
fileset='train',
limit=4000,
subsample=1000,
only_common_bigrams=True,
cmudict_sort=True,
normalize=True):
feats, uttids = kaldi_io.readArk(ark_file.replace('%set', fileset),
limit=limit)
if normalize:
print('Normalize to unit length.')
feats = [feat / np.linalg.norm(feat) for feat in feats]
phones_txt = alignment_dir + 'phones.txt'
num2phone_orig = load_phones_txt(phones_txt)
#phone2num_orig =
print(num2phone_orig)
phoneset = list(set(num2phone_orig.values()))
if cmudict_sort:
num2phone = dict(
enumerate(cmudict_silence + cmudict_vowels + cmudict_consonants))
phone2num = dict([
(y, x) for x, y in enumerate(cmudict_silence + cmudict_vowels +
cmudict_consonants)
])
else:
num2phone = dict(enumerate(sorted(list(set(num2phone_orig.values())))))
phone2num = dict([
(y, x)
for x, y in enumerate(sorted(list(set(num2phone_orig.values()))))
])
num2phoneid = dict([(i, phone2num[x]) for i, x in num2phone_orig.items()])
print(num2phoneid)
ctm_file = alignment_dir + 'all.ctm'
all_ctm_db = load_all_ctm_pos(ctm_file, num2phoneid, limit=limit)
phn_ngram_counts = Counter()
ngram_func = fivegrams
feat_vars = []
phone_vars = []
for feat, uttid in zip(feats, uttids):
pos_array, phone_array = all_ctm_db[uttid]
assert (len(pos_array) == len(phone_array))
feat_vars += [feat[list(pos_array)]]
phone_vars += [phone_array]
phn_ngram_counts.update(ngram_func(phone_array))
feat_vars = np.vstack(feat_vars)
phone_vars = np.hstack(phone_vars)
#print(phn_counts)
most_common_ngrams = []
print('Most common phone ngrams:')
for phone_ngram, count in phn_ngram_counts.most_common(10):
print(phone_ngram, [num2phone[elem] for elem in phone_ngram], count)
most_common_ngrams.append(phone_ngram)
#most_common_ngrams = [(30, 12, 30, 4, 26), (32, 4, 27, 35, 11), (4, 37, 9, 26, 19)]
#print('Manually selected:', most_common_ngrams)
most_common_bigrams_to_num = dict([
(b, a) for a, b in enumerate(most_common_ngrams)
])
print('Loaded', len(phone_vars), 'phoneme examples')
if only_common_bigrams:
select_pos = [
i for i, elem in enumerate(ngram_func(phone_vars))
if elem in most_common_ngrams
]
feat_vars = feat_vars[select_pos]
phone_vars = np.asarray([
most_common_bigrams_to_num[elem]
for i, elem in enumerate(ngram_func(phone_vars))
if elem in most_common_ngrams
]) #phone_vars[select_pos]
print(phone_vars)
print(feat_vars)
assert (len(feat_vars) == len(phone_vars))
if subsample != -1:
feat_phone_vars_sampled_idx = np.random.choice(np.arange(
len(feat_vars)),
subsample,
replace=False)
feat_vars = feat_vars[feat_phone_vars_sampled_idx]
phone_vars = phone_vars[feat_phone_vars_sampled_idx]
print('Subsampled to:', len(phone_vars), 'phoneme examples')
model = TSNE(n_components=2, random_state=0, metric='cosine')
tsne_data = model.fit_transform(feat_vars)
num_classes = len(phoneset)
if only_common_bigrams:
num_classes = len(list(set(phone_vars)))
colormap = plt.cm.gist_ncar #nipy_spectral, Set1,Paired
colorst = colormap(np.linspace(
0, 1.0, num_classes +
1)) #[colormap(i) for i in np.linspace(0, 0.9, num_speakers)]
cs = [colorst[phone_vars[i]] for i in range(len(phone_vars))]
#print(tsne_data[:,0])
#print(tsne_data[:,1])
for i, color in enumerate(colorst):
plt.text(30 + float(i) * 3, 30, str(i), color=color)
plt.scatter(tsne_data[:, 0], tsne_data[:, 1], color=cs)
#for i,elem in enumerate(tsne_data):
# print(cs[0])
# print(ground_truth_utt_2_spk[0])
# plt.scatter(elem[0], elem[1], color=cs[i], label=ground_truth_utt_2_spk[i])
#plt.legend(most_common_ngrams)
plt.show()
def same_different_experiment(ark_file,
utt_2_spk,
half_index=-1,
normalize=False,
fileset='',
use_metric='cosine',
max_spks=-1,
random_seed=42):
results_file = 'samedifferent_results.csv'
from pyannote.metrics.plot.binary_classification import plot_det_curve, plot_distributions
print('Loading feats now:')
feats, uttids = kaldi_io.readArk(ark_file.replace('%set', fileset))
print('loaded: ' + str(len(feats)) + ' feats')
print('feat[0] shape: ', feats[0].shape)
#feats = np.vstack([pairwise_normalize(feat[0]) for feat in feats])
print('Generating mean vector.')
feats = np.vstack([feat.mean(0) for utt, feat in zip(uttids, feats)])
if half_index != -1:
print('Cutting vectors at ', half_index,
'and normalize to unit length' if normalize else '')
feats = np.vstack([
feat[:half_index] /
(np.linalg.norm(feat[:half_index]) if normalize else 1.0)
for feat in feats
])
else:
if normalize:
print('Normalize to unit length.')
feats = np.vstack([feat / np.linalg.norm(feat) for feat in feats])
#print(type(feats))
#print(feats)
if utt_2_spk is not None and utt_2_spk.lower(
) != 'none' and utt_2_spk.strip() != '':
utt_2_spk = utils.loadUtt2Spk(utt_2_spk.replace('%set', fileset))
if max_spks != -1:
# sample subset of speakers
spk_set = list(set(utt_2_spk.values()))
# make speaker selection reproducable
spk_set.sort()
#print('Selecting from these speakers:', spk_set)
random.seed(random_seed)
#np.random.seed(42)
selected_spks = random.sample(spk_set, min(len(spk_set), max_spks))
print('Selecting random subset of speakers with random seed',
random_seed, ':', len(selected_spks), 'speakers')
print(selected_spks)
# make new utt2spk dictionary on subset
utt_2_spk_new = dict([(key, utt_2_spk[key]) for key in utt_2_spk
if utt_2_spk[key] in selected_spks])
#filter feats and uttids
feats = [
feat for feat, uttid in zip(feats, uttids)
if uttid in utt_2_spk_new
]
uttids = [uttid for uttid in uttids if uttid in utt_2_spk_new]
print('Reduced feats to: ' + str(len(feats)) + ' feats')
print('Reduced uttids to: ' + str(len(feats)) + ' uttids')
utt_2_spk = utt_2_spk_new
else:
print('Using all speakers:', len(set(utt_2_spk.values())))
ground_truth_utt_2_spk = [utt_2_spk[utt_id] for utt_id in uttids]
le = preprocessing.LabelEncoder()
le.fit(ground_truth_utt_2_spk)
ground_truth_utt_2_spk_int = le.transform(ground_truth_utt_2_spk)
print("Ground truth speaker classes available:")
print(ground_truth_utt_2_spk_int)
print('Calculating', use_metric, 'distance matrix...')
#print('feats shape:', feats.shape)
distances = pdist(feats, metric=use_metric)
print('Calculating ground thruth distance matrix...')
y_true = pdist(np.asarray(ground_truth_utt_2_spk_int)[:, np.newaxis],
metric='chebyshev') < 1
result_key = ark_file.split('/')[-3] + ('.' + fileset if fileset != '' else
'') + '.' + use_metric
prefix = 'plots/plot.' + ark_file.split('/')[-3] + '.' + use_metric + (
'.' + fileset if fileset != '' else '') + '.seed_' + str(random_seed)
plot_distributions(y_true,
distances,
prefix,
xlim=(0, 2),
ymax=3,
nbins=100)
eer = plot_det_curve(y_true, -distances, prefix)
print('EER = {eer:.2f}%'.format(eer=100 * eer))
with open(results_file, 'a') as outfile:
outfile.write(result_key + ' ' + '{eer:.2f}%'.format(eer=100 * eer) +
'\n')