def inference_embeddings(audio_reader, speaker_id):
speaker_feat = generate_features_for_unseen_speakers(audio_reader, target_speaker=speaker_id)
# batch_size => None (for inference).
m = triplet_softmax_model(num_speakers_softmax=len(c.AUDIO.SPEAKERS_TRAINING_SET),
emb_trainable=False,
normalize_embeddings=True,
batch_size=None)
checkpoints = natsorted(glob('checkpoints/*.h5'))
print(m.summary())
if len(checkpoints) != 0:
checkpoint_file = checkpoints[-1]
initial_epoch = int(checkpoint_file.split('/')[-1].split('.')[0].split('_')[-1])
logger.info('Initial epoch is {}.'.format(initial_epoch))
logger.info('Loading checkpoint: {}.'.format(checkpoint_file))
m.load_weights(checkpoint_file) # latest one.
emb_sp1 = m.predict(np.vstack(speaker_feat))[0]
logger.info('Emb1.shape = {}'.format(emb_sp1.shape))
np.set_printoptions(suppress=True)
emb1 = np.mean(emb_sp1, axis=0)
print('*' * 80)
print(emb1)
print('*' * 80)
def inference_unseen_speakers(audio_reader, sp1, sp2):
sp1_feat = generate_features_for_unseen_speakers(audio_reader,
target_speaker=sp1)
sp2_feat = generate_features_for_unseen_speakers(audio_reader,
target_speaker=sp2)
# batch_size => None (for inference).
m = triplet_softmax_model(num_speakers_softmax=len(
c.AUDIO.SPEAKERS_TRAINING_SET),
emb_trainable=False,
normalize_embeddings=True,
batch_size=None)
checkpoints = natsorted(glob('checkpoints/*.h5'))
# compile_triplet_softmax_model(m, loss_on_softmax=False, loss_on_embeddings=False)
print(m.summary())
if len(checkpoints) != 0:
checkpoint_file = checkpoints[-1]
initial_epoch = int(
checkpoint_file.split('/')[-1].split('.')[0].split('_')[-1])
logger.info('Initial epoch is {}.'.format(initial_epoch))
logger.info('Loading checkpoint: {}.'.format(checkpoint_file))
m.load_weights(checkpoint_file) # latest one.
emb_sp1 = m.predict(np.vstack(sp1_feat))[0]
emb_sp2 = m.predict(np.vstack(sp2_feat))[0]
logger.info('Checking that L2 norm is 1.')
logger.info(np.mean(np.linalg.norm(emb_sp1, axis=1)))
logger.info(np.mean(np.linalg.norm(emb_sp2, axis=1)))
from scipy.spatial.distance import cosine
# note to myself:
# embeddings are sigmoid-ed.
# so they are between 0 and 1.
# A hypersphere is defined on tanh.
logger.info('Emb1.shape = {}'.format(emb_sp1.shape))
logger.info('Emb2.shape = {}'.format(emb_sp2.shape))
emb1 = np.mean(emb_sp1, axis=0)
emb2 = np.mean(emb_sp2, axis=0)
logger.info('Cosine = {}'.format(cosine(emb1, emb2)))
def load_model():
m = triplet_softmax_model(num_speakers_softmax=len(
c.AUDIO.SPEAKERS_TRAINING_SET),
emb_trainable=False,
normalize_embeddings=True,
batch_size=None)
checkpoints = natsorted(glob('checkpoints/*.h5'))
print(m.summary())
if len(checkpoints) != 0:
checkpoint_file = checkpoints[-1]
initial_epoch = int(
checkpoint_file.split('/')[-1].split('.')[0].split('_')[-1])
logger.info('Initial epoch is {}.'.format(initial_epoch))
logger.info('Loading checkpoint: {}.'.format(checkpoint_file))
m.load_weights(checkpoint_file) # latest one.
return m