def train():
# setting parameters
batch_size = 32
n_epoch = 100
n_mfcc = 60
# load speech data
wav_path = os.path.join(os.getcwd(),'data','wav','train')
label_file = os.path.join(os.getcwd(),'data','doc','trans','train.word.txt')
speech_loader = SpeechLoader(wav_path, label_file, batch_size, n_mfcc)
n_out = speech_loader.vocab_size
# load model
model = Model(n_out, batch_size=batch_size, n_mfcc=n_mfcc)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
for epoch in range(n_epoch):
speech_loader.create_batches() # random shuffle data
speech_loader.reset_batch_pointer()
for batch in range(speech_loader.n_batches):
start = time.time()
batches_wav, batches_label = speech_loader.next_batch()
feed = {model.input_data: batches_wav, model.targets: batches_label}
train_loss, _ = sess.run([model.cost, model.optimizer_op], feed_dict=feed)
end = time.time()
print("epoch: %d/%d, batch: %d/%d, loss: %s, time: %.3f."%(epoch, n_epoch, batch, speech_loader.n_batches, train_loss, end-start))
# save models
if epoch % 5 ==0:
saver.save(sess, os.path.join(os.getcwd(), 'model','speech.module'), global_step=epoch)
def train():
# setting parameters
batch_size = 2
n_epoch = 100
n_mfcc = 60
# load speech data
wav_path = os.path.join(os.getcwd(), 'data', 'wav', 'train')
label_file = os.path.join(os.getcwd(), 'data', 'doc', 'trans',
'train.word.txt')
speech_loader = SpeechLoader(wav_path, label_file, batch_size, n_mfcc)
n_out = speech_loader.vocab_size
# load model
model = Model(n_out, batch_size=batch_size, n_mfcc=n_mfcc)
tf.summary.scalar('loss', model.cost)
merged = tf.summary.merge_all()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
if len(os.listdir('./model')) > 3:
print("loading model from checkpoint")
checkpoint = tf.train.latest_checkpoint('./model')
saver.restore(sess, checkpoint)
tf.train.write_graph(sess.graph_def, './model', 'model.pbtxt')
summary_writer = tf.summary.FileWriter('./model', graph=sess.graph)
for epoch in range(n_epoch):
speech_loader.create_batches() # random shuffle data
speech_loader.reset_batch_pointer()
for batch in range(speech_loader.n_batches):
start = time.time()
batches_wav, batches_label = speech_loader.next_batch()
feed = {
model.input_data: batches_wav,
model.targets: batches_label
}
result, train_loss, _ = sess.run(
[merged, model.cost, model.optimizer_op], feed_dict=feed)
end = time.time()
print("epoch: %d/%d, batch: %d/%d, loss: %s, time: %.3f." %
(epoch, n_epoch, batch, speech_loader.n_batches,
train_loss, end - start))
summary_writer.add_summary(result, epoch)
# save models
if epoch % 5 == 0:
saver.save(sess,
os.path.join(os.getcwd(), 'model', 'speech.module'),
global_step=epoch)
def train():
'''
:return:
'''
batch_size = 8
n_mfcc = 60
n_epoch = 100
source_file = '/home/ydf_micro/datasets/data_thchs30'
speech_loader = SpeechLoader(os.path.join(source_file, 'train'),
batch_size, n_mfcc)
n_out = speech_loader.vocab_size
# load model
model = WaveNet(n_out, batch_size=batch_size, n_mfcc=n_mfcc)
saver = tf.train.Saver(tf.global_variables())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# sess.graph.finalize() # Graph is read-only after this statement
for epoch in range(n_epoch):
speech_loader.create_batches() # random shuffle data
speech_loader.reset_batch_pointer()
for batch in range(speech_loader.n_batches):
batch_start = time.time()
batches_wav, batches_label = speech_loader.next_batch()
feed = {
model.input_data: batches_wav,
model.targets: batches_label
}
train_loss, _ = sess.run([model.cost, model.optimizer_op],
feed_dict=feed)
batch_end = time.time()
print(
f'epoch: {epoch+1}/{n_epoch}, batch: {batch+1}/{speech_loader.n_batches}, '
f'loss: {train_loss:.2f}, time: {(batch_end-batch_start):.2f}s'
)
# save models
if epoch % 5 == 0:
saver.save(sess,
os.path.join(os.path.dirname(os.getcwd()), 'model',
'speech.module'),
global_step=epoch)