def train():
lr = 0.002
c = ClassificationModel()
cost, normVector, alfas_mean = c.buildModel()
gStep = tf.Variable(tf.constant(0))
learning_rate = tf.train.exponential_decay(float(lr), gStep, 300, 0.9, staircase=True)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
gradients, vars = zip(*optimizer.compute_gradients(cost)) #var_list=train_vars
gradients, _ = tf.clip_by_global_norm(gradients, 100)
train_op = optimizer.apply_gradients(zip(gradients, vars))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# saver = tf.train.Saver(var_list=restore_vars) #var_list=restore_vars
# saver.restore(sess, 'step/model.ckpt-'+ckpoint)
saver = tf.train.Saver()
for step in range(10000):
mels, ids = c.getBatch_data_label(10)
feed_dict = {c.melInputs:mels, c.labelInputs: ids, gStep: step}
_, _lr, _loss, _alfas_mean = sess.run([train_op, learning_rate, cost, alfas_mean], feed_dict=feed_dict)
print('{}: lr = {:.6f}, loss = {}, alfas_mean = {}'.format(step, _lr, _loss, _alfas_mean))
if(step%300==0 and step!=0):
saver.save(sess, 'step/model.ckpt', global_step=step)
def test():
c = ClassificationModel()
_, normVector, _ = c.buildModel(is_training=False)
restore_vars = []
for var in tf.global_variables():
if('temp' not in var.name):
restore_vars.append(var)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(var_list=restore_vars)
saver.restore(sess, 'step/model.ckpt-2700')
speaker_path = '/home/logview/workspace/projects/TI-SV/samples'
utterance_specs = []
files = os.listdir(speaker_path)
files.sort()
for utter_name in files:
utter_path = os.path.join(speaker_path, utter_name) # path of each utterance
utter, sr = librosa.core.load(utter_path, 16000) # load utterance audio
utter_trim, index = librosa.effects.trim(utter, top_db=20) # voice activity detection, only trim
S = librosa.feature.mfcc(y=utter_trim, sr=sr, n_mfcc=40)
inputs = S.transpose((1,0))[:160]
print(inputs.shape)
utterance_specs.append(inputs)
utterance_specs = np.array(utterance_specs)
print(utterance_specs.shape)
vectors = sess.run(normVector, feed_dict={c.melInputs:utterance_specs})
similar(vectors)