def load_data(self, args, mode, type):
if mode == 'train':
return load_batched_data(train_mfcc_dir, train_label_dir,
batch_size, mode, type)
elif mode == 'test':
return load_batched_data(test_mfcc_dir, test_label_dir, batch_size,
mode, type)
else:
raise TypeError('mode should be train or test.')
TARGET_PATH = '../TRAIN/All/phone_y/' #directory of nPhonemes 1-D array .npy files
####Learning Parameters
learningRate = 0.001
momentum = 0.9
nEpochs = 120
batchSize = 128
####Network Parameters
nFeatures = 26 #12 MFCC coefficients + energy, and derivatives
nHidden = 128
nClasses = 40 #39 phonemes, plus the "blank" for CTC
####Load data
print('Loading data')
batchedData, maxTimeSteps, totalN = load_batched_data(INPUT_PATH, TARGET_PATH,
batchSize)
####Define graph
print('Defining graph')
graph = tf.Graph()
with graph.as_default():
####NOTE: try variable-steps inputs and dynamic bidirectional rnn, when it's implemented in tensorflow
####Graph input
inputX = tf.placeholder(tf.float32,
shape=(maxTimeSteps, batchSize, nFeatures))
#Prep input data to fit requirements of rnn.bidirectional_rnn
# Reshape to 2-D tensor (nTimeSteps*batchSize, nfeatures)
inputXrs = tf.reshape(inputX, [-1, nFeatures])
# Split to get a list of 'n_steps' tensors of shape (batch_size, n_hidden)
def main(_):
train_mfcc_dir = os.path.join(FLAGS.input_data_dir, FLAGS.level, 'TRAIN',
'mfcc')
train_label_dir = os.path.join(FLAGS.input_data_dir, FLAGS.level, 'TRAIN',
'label')
test_mfcc_dir = os.path.join(FLAGS.input_data_dir, FLAGS.level, 'TEST',
'mfcc')
test_label_dir = os.path.join(FLAGS.input_data_dir, FLAGS.level, 'TEST',
'label')
savedir = os.path.join(FLAGS.exp_dir, FLAGS.level, 'save')
resultdir = os.path.join(FLAGS.exp_dir, FLAGS.level, 'result')
if FLAGS.is_training:
batched_data, max_time_steps, total_n = load_batched_data(
train_mfcc_dir, train_label_dir, FLAGS.batch_size, FLAGS.level)
else:
batched_data, max_time_steps, total_n = load_batched_data(
test_mfcc_dir, test_label_dir, FLAGS.batch_size, FLAGS.level)
hparams = {}
hparams['level'] = FLAGS.level
hparams['batch_size'] = FLAGS.batch_size
hparams['partition_size'] = FLAGS.partition_size
hparams['num_hidden'] = FLAGS.num_hidden
hparams['feature_length'] = FLAGS.feature_length
hparams['num_classes'] = FLAGS.num_classes
hparams['num_proj'] = FLAGS.num_proj
hparams['learning_rate'] = FLAGS.learning_rate
hparams['keep_prob'] = FLAGS.keep_prob
hparams['clip_gradient_norm'] = FLAGS.clip_gradient_norm
hparams['use_peepholes'] = FLAGS.use_peepholes
if FLAGS.activation == 'tanh':
hparams['activation'] = tf.tanh
elif FLAGS.activation == 'relu':
hparams['activation'] = tf.nn.relu
hparams['max_time_steps'] = max_time_steps
with tf.Graph().as_default():
model = DRNN(FLAGS.cell, hparams, FLAGS.is_training)
train_writer = tf.summary.FileWriter(resultdir + '/train')
test_writer = tf.summary.FileWriter(resultdir + '/test')
with tf.Session(FLAGS.master) as sess:
# restore from stored model
if FLAGS.restore:
ckpt = tf.train.get_checkpoint_state(savedir)
if ckpt and ckpt.model_checkpoint_path:
model.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model restored from:' + ckpt.model_checkpoint_path)
else:
print('Initializing')
sess.run(model.initial_op)
train_writer.add_graph(sess.graph)
for epoch in range(FLAGS.num_epochs):
## training
start = time.time()
if FLAGS.is_training:
print('Epoch', epoch + 1, '...')
batch_errors = np.zeros(len(batched_data))
batched_random_idx = np.random.permutation(len(batched_data))
for batch, batch_original_idx in enumerate(batched_random_idx):
batch_inputs, batch_target_sparse, batch_seq_length = batched_data[
batch_original_idx]
batch_tgt_idx, batch_tgt_vals, batch_tgt_shape = batch_target_sparse
feeddict = {
model.x: batch_inputs,
model.tgt_idx: batch_tgt_idx,
model.tgt_vals: batch_tgt_vals,
model.tgt_shape: batch_tgt_shape,
model.seq_length: batch_seq_length
}
if FLAGS.is_training and (
(epoch * len(batched_random_idx) + batch + 1) % 20 == 0
or (epoch == FLAGS.num_epochs - 1
and batch == len(batched_random_idx) - 1)):
checkpoint_path = os.path.join(savedir, 'model.ckpt')
model.saver.save(sess,
checkpoint_path,
global_step=model.global_step)
print('Model has been saved in {}'.format(savedir))
if FLAGS.level == 'cha':
if FLAGS.is_training:
_, l, pre, y, er, global_step = sess.run(
[
model.train_op, model.loss,
model.predictions, model.y,
model.error_rate, model.global_step
],
feed_dict=feeddict)
batch_errors[batch] = er
if global_step % 10 == 0:
log_scalar(train_writer, 'CER',
er / FLAGS.batch_size, global_step)
print(
'{} mode, global_step:{}, lr:{}, total:{}, '
'batch:{}/{},epoch:{}/{},train loss={:.3f},mean train '
'CER={:.3f}'.format(
FLAGS.level, global_step,
FLAGS.learning_rate,
total_n, batch + 1,
len(batched_random_idx), epoch + 1,
FLAGS.num_epochs, l,
er / FLAGS.batch_size))
elif not FLAGS.is_training:
l, pre, y, er, global_step = sess.run(
[
model.loss, model.predictions, model.y,
model.error_rate, model.global_step
],
feed_dict=feeddict)
batch_errors[batch] = er
log_scalar(test_writer, 'CER',
er / FLAGS.batch_size, global_step)
print(
'{} mode, global_step:{}, total:{}, batch:{}/{},test '
'loss={:.3f},mean test CER={:.3f}'.format(
FLAGS.level, global_step, total_n,
batch + 1, len(batched_random_idx), l,
er / FLAGS.batch_size))
elif FLAGS.level == 'phn':
if FLAGS.is_training:
_, l, pre, y, global_step = sess.run(
[
model.train_op, model.loss,
model.predictions, model.y,
model.global_step
],
feed_dict=feeddict)
er = get_edit_distance([pre.values], [y.values],
True, FLAGS.level)
if global_step % 10 == 0:
log_scalar(train_writer, 'PER', er,
global_step)
print(
'{} mode, global_step:{}, lr:{}, total:{}, '
'batch:{}/{},epoch:{}/{},train loss={:.3f},mean train '
'PER={:.3f}'.format(
FLAGS.level, global_step,
FLAGS.learning_rate,
total_n, batch + 1,
len(batched_random_idx), epoch + 1,
FLAGS.num_epochs, l, er))
batch_errors[batch] = er * len(batch_seq_length)
elif not FLAGS.is_training:
l, pre, y, global_step = sess.run(
[
model.loss, model.predictions, model.y,
model.global_step
],
feed_dict=feeddict)
er = get_edit_distance([pre.values], [y.values],
True, FLAGS.level)
log_scalar(test_writer, 'PER', er, global_step)
print(
'{} mode, global_step:{}, total:{}, batch:{}/{},test '
'loss={:.3f},mean test PER={:.3f}'.format(
FLAGS.level, global_step, total_n,
batch + 1, len(batched_random_idx), l, er))
batch_errors[batch] = er * len(batch_seq_length)
# NOTE:
if er / FLAGS.batch_size == 1.0:
break
if batch % 100 == 0:
print('Truth:\n' +
output_to_sequence(y, level=FLAGS.level))
print('Output:\n' +
output_to_sequence(pre, level=FLAGS.level))
end = time.time()
delta_time = end - start
print('Epoch ' + str(epoch + 1) + ' needs time:' +
str(delta_time) + ' s')
if FLAGS.is_training:
if (epoch + 1) % 1 == 0:
checkpoint_path = os.path.join(savedir, 'model.ckpt')
model.saver.save(sess,
checkpoint_path,
global_step=model.global_step)
print('Model has been saved in {}'.format(savedir))
epoch_er = batch_errors.sum() / total_n
print('Epoch', epoch + 1, 'mean train error rate:',
epoch_er)
if not FLAGS.is_training:
with tf.gfile.GFile(
os.path.join(resultdir,
FLAGS.level + '_result.txt'),
'a') as result:
result.write(
output_to_sequence(y, level=FLAGS.level) + '\n')
result.write(
output_to_sequence(pre, level=FLAGS.level) + '\n')
result.write('\n')
epoch_er = batch_errors.sum() / total_n
print(' test error rate:', epoch_er)
TARGET_PATH = './sample_data/char_y/' #directory of nCharacters 1-D array .npy files
####Learning Parameters
learningRate = 0.001
momentum = 0.9
nEpochs = 300
batchSize = 4
####Network Parameters
nFeatures = 26 #12 MFCC coefficients + energy, and derivatives
nHidden = 128
nClasses = 28#27 characters, plus the "blank" for CTC
####Load data
print('Loading data')
batchedData, maxTimeSteps, totalN = load_batched_data(INPUT_PATH, TARGET_PATH, batchSize)
####Define graph
print('Defining graph')
graph = tf.Graph()
with graph.as_default():
####NOTE: try variable-steps inputs and dynamic bidirectional rnn, when it's implemented in tensorflow
####Graph input
inputX = tf.placeholder(tf.float32, shape=(maxTimeSteps, batchSize, nFeatures))
#Prep input data to fit requirements of rnn.bidirectional_rnn
# Reshape to 2-D tensor (nTimeSteps*batchSize, nfeatures)
inputXrs = tf.reshape(inputX, [-1, nFeatures])
# Split to get a list of 'n_steps' tensors of shape (batch_size, n_hidden)
inputList = tf.split(0, maxTimeSteps, inputXrs)
def load_data(self, feature_dir, label_dir, mode, level):
return load_batched_data(feature_dir, label_dir, batch_size, mode,
level)
def train(self):
print("[INFO]: Data Loaded")
model = SimpleModel(cfg)
print("[INFO]: Building Graph ...")
model.build_graph()
num_batches = int(len(self.data[0]) / cfg.batch_size)
print("[INFO]: Number of batches are %d " % num_batches)
with tf.Session(graph=model.graph) as sess:
writer = tf.summary.FileWriter("logging", graph=model.graph)
if (cfg.restore == 1):
model.saver.restore(
sess, tf.train.latest_checkpoint(cfg.checkpoint_dir))
else:
sess.run(model.initial_op)
total_loss = 0.
avg_loss = 0.
for iteration in range(cfg.num_iters):
X_tuple = next(
load_batched_data(self.data, cfg.batch_size, cfg))
_, loss, summary = sess.run(
[model.opt, model.loss, model.summary_op],
feed_dict={
model.name: X_tuple[0],
model.item_condition_id: X_tuple[1],
model.category_id: X_tuple[2],
model.brand_id: X_tuple[3],
model.shipping: X_tuple[4],
model.item_description: X_tuple[5],
model.target_price: X_tuple[-1]
})
writer.add_summary(summary, iteration)
total_loss = total_loss + loss
avg_loss = total_loss / (iteration + 1)
if (iteration != 0 & iteration % 50 == 0):
print("[Info]: Iter:%d/%d , avg_loss: %f " %
(iteration + 1, cfg.num_iters, avg_loss))
if (iteration % 100 == 0):
model.saver.save(sess,
cfg.checkpoint_dir + cfg.dataset +
"_SimpleModel.ckpt",
global_step=model.global_step)
# Test after every 500 iterations
if (iteration % 500 == 0):
total_val_loss = 0.
for iteration in range(
len(self.val_data[0]) // cfg.batch_size):
test_tuple = next(
load_batched_data(self.val_data, cfg.batch_size,
cfg))
_, val_loss = sess.run(
[model.pred_price, model.loss],
feed_dict={
model.name: test_tuple[0],
model.item_condition_id: test_tuple[1],
model.category_id: test_tuple[2],
model.brand_id: test_tuple[3],
model.shipping: test_tuple[4],
model.item_description: test_tuple[5],
model.target_price: test_tuple[-1]
})
total_val_loss = total_val_loss + val_loss
avg_val_loss = total_val_loss / (iteration + 1)
print("Test RMSE: %f" % avg_val_loss)
def main(_):
print('%s mode...' % str(FLAGS.mode))
savedir = os.path.join(FLAGS.exp_dir, FLAGS.level, 'save')
resultdir = os.path.join(FLAGS.exp_dir, FLAGS.level, 'result')
check_path_exists([savedir, resultdir])
# load data
hparams = {}
hparams['level'] = FLAGS.level
hparams['batch_size'] = FLAGS.batch_size
hparams['partition_size'] = FLAGS.partition_size
hparams['num_hidden'] = FLAGS.num_hidden
hparams['feature_length'] = FLAGS.feature_length
hparams['num_classes'] = FLAGS.num_classes
hparams['num_proj'] = FLAGS.num_proj
hparams['learning_rate'] = FLAGS.learning_rate
hparams['keep_prob'] = FLAGS.keep_prob
hparams['clip_gradient_norm'] = FLAGS.clip_gradient_norm
hparams['use_peepholes'] = FLAGS.use_peepholes
if FLAGS.activation == 'tanh':
hparams['activation'] = tf.tanh
elif FLAGS.activation == 'relu':
hparams['activation'] = tf.nn.relu
feature_dirs, label_dirs = get_data(FLAGS.input_data_dir, FLAGS.level,
FLAGS.train_dataset, FLAGS.dev_dataset,
FLAGS.test_dataset, FLAGS.mode)
batched_data, max_time_steps, total_n = load_batched_data(
feature_dirs[0], label_dirs[0], FLAGS.batch_size, FLAGS.level)
hparams['max_time_steps'] = max_time_steps
## shuffle feature_dir and label_dir by same order
FL_pair = list(zip(feature_dirs, label_dirs))
random.shuffle(FL_pair)
feature_dirs, label_dirs = zip(*FL_pair)
train_writer = tf.summary.FileWriter(resultdir + '/train')
test_writer = tf.summary.FileWriter(resultdir + '/test')
for feature_dir, label_dir in zip(feature_dirs, label_dirs):
id_dir = feature_dirs.index(feature_dir)
print('dir id:{}'.format(id_dir))
batched_data, max_time_steps, total_n = load_batched_data(
feature_dir, label_dir, FLAGS.batch_size, FLAGS.level)
hparams['max_time_steps'] = max_time_steps
model = DRNN(FLAGS.cell, hparams, FLAGS.mode == 'train')
with tf.Session(FLAGS.master) as sess:
# restore from stored model
if FLAGS.restore:
ckpt = tf.train.get_checkpoint_state(savedir)
if ckpt and ckpt.model_checkpoint_path:
model.saver.restore(sess, ckpt.model_checkpoint_path)
print('Model restored from:' + savedir)
else:
print('Initializing')
sess.run(model.initial_op)
for epoch in range(FLAGS.num_epochs):
## training
start = time.time()
if FLAGS.mode == 'train':
print('Epoch {} ...'.format(epoch + 1))
batch_errors = np.zeros(len(batched_data))
batched_random_idx = np.random.permutation(len(batched_data))
for batch, batch_original_idx in enumerate(batched_random_idx):
batch_inputs, batch_target_sparse, batch_seq_length = batched_data[
batch_original_idx]
batch_tgt_idx, batch_tgt_vals, batch_tgt_shape = batch_target_sparse
feedDict = {
model.x: batch_inputs,
model.tgt_idx: batch_tgt_idx,
model.tgt_vals: batch_tgt_vals,
model.tgt_shape: batch_tgt_shape,
model.seq_length: batch_seq_length
}
if FLAGS.level == 'cha':
if FLAGS.mode == 'train':
_, l, pre, y, er = sess.run([
model.train_op, model.loss, model.predictions,
model.y, model.error_rate
],
feed_dict=feedDict)
batch_errors[batch] = er
print(
'\n{} mode, total:{},subdir:{}/{},batch:{}/{},epoch:{}/{},train loss={:.3f},mean train CER={:.3f}\n'
.format(FLAGS.level, total_n, id_dir + 1,
len(feature_dirs), batch + 1,
len(batched_random_idx), epoch + 1,
FLAGS.num_epochs, l,
er / FLAGS.batch_size))
elif FLAGS.mode == 'dev':
l, pre, y, er = sess.run([
model.loss, model.predictions, model.y,
model.error_rate
],
feed_dict=feedDict)
batch_errors[batch] = er
print(
'\n{} mode, total:{},subdir:{}/{},batch:{}/{},dev loss={:.3f},mean dev CER={:.3f}\n'
.format(FLAGS.level, total_n, id_dir + 1,
len(feature_dirs), batch + 1,
len(batched_random_idx), l,
er / FLAGS.batch_size))
elif FLAGS.mode == 'test':
l, pre, y, er = sess.run([
model.loss, model.predictions, model.y,
model.error_rate
],
feed_dict=feedDict)
batch_errors[batch] = er
print(
'\n{} mode, total:{},subdir:{}/{},batch:{}/{},test loss={:.3f},mean test CER={:.3f}\n'
.format(FLAGS.level, total_n, id_dir + 1,
len(feature_dirs), batch + 1,
len(batched_random_idx), l,
er / FLAGS.batch_size))
elif FLAGS.level == 'seq2seq':
raise ValueError('level %s is not supported now' %
str(FLAGS.level))
# NOTE:
if er / FLAGS.batch_size == 1.0:
break
if batch % 20 == 0:
print('Truth:\n' +
output_to_sequence(y, level=FLAGS.level))
print('Output:\n' +
output_to_sequence(pre, level=FLAGS.level))
if FLAGS.mode == 'train' and (
(epoch * len(batched_random_idx) + batch + 1) % 20 == 0
or (epoch == FLAGS.num_epochs - 1
and batch == len(batched_random_idx) - 1)):
checkpoint_path = os.path.join(savedir, 'model.ckpt')
model.saver.save(sess,
checkpoint_path,
global_step=epoch)
print('Model has been saved in {}'.format(savedir))
end = time.time()
delta_time = end - start
print('Epoch ' + str(epoch + 1) + ' needs time:' +
str(delta_time) + ' s')
if FLAGS.mode == 'train':
if (epoch + 1) % 1 == 0:
checkpoint_path = os.path.join(savedir, 'model.ckpt')
model.saver.save(sess,
checkpoint_path,
global_step=epoch)
print('Model has been saved in {}'.format(savedir))
epoch_er = batch_errors.sum() / total_n
print('Epoch', epoch + 1, 'mean train error rate:',
epoch_er)
if FLAGS.mode == 'test' or FLAGS.mode == 'dev':
with open(
os.path.join(resultdir,
FLAGS.level + '_result.txt'),
'a') as result:
result.write(
output_to_sequence(y, level=FLAGS.level) + '\n')
result.write(
output_to_sequence(pre, level=FLAGS.level) + '\n')
result.write('\n')
epoch_er = batch_errors.sum() / total_n
print(' test error rate:', epoch_er)
def load_data(self, batch_size, b, e, d, target_gaps):
return load_batched_data(batch_size, b, e, d, target_gaps)
def run(self, user):
args_dict = self._default_configs()
args = dotdict(args_dict)
#get data
#print("Loading data")
train_dataset = "../data/dataset_" + str(user) + "_train.pickle"
test_dataset = "../data/dataset_" + str(user) + "_test.pickle"
b, e, d, g, target_gaps = get_data(train_dataset)
b_t, e_t, d_t, g_t, target_gaps_t = get_data(test_dataset)
#print(len(b))
#print(type(g))
#g = np.vstack(g)
#print(b.shape)
totalN = len(b)
#print("b :",len(b))
num_batches = len(b) / batch_size
maxLength = 0
for x in b:
maxLength = max(maxLength, x.shape[0])
maxSessLen = maxLength
print("Building Model")
recom = Model(args, maxSessLen)
recom.build_graph(args, maxSessLen)
print("Starting Session")
#print(recom.config)
test_Err = []
test_MAE = []
with tf.Session(graph=recom.graph) as sess:
if (mode == 'train'):
writer = tf.summary.FileWriter("loggingdir", graph=recom.graph)
sess.run(recom.initial_op)
for epoch in range(num_epochs):
# training
start = time.time()
print('Epoch {} ...'.format(epoch + 1))
batchLoss = np.zeros(num_batches)
batchErr = np.zeros(num_batches)
batchErrMAE = np.zeros(num_batches)
batchRandIxs = np.random.permutation(num_batches)
for batch, batchOrigI in enumerate(batchRandIxs):
batchInputs_b, batchInputs_e, batchInputs_d, batchInputs_g, batchTargetList, batchSeqLengths = next(
load_batched_data(batch_size, b, e, d, g,
target_gaps))
#print(type(batchInputs_g))
feedDict = {
recom.inputb: batchInputs_b,
recom.inpute: batchInputs_e,
recom.inputg: batchInputs_g,
recom.inputd: batchInputs_d,
recom.target_gaps: batchTargetList,
recom.sessLengths: batchSeqLengths
}
_, l, lamb = sess.run(
[recom.optimizer, recom.loss, recom.lamb],
feed_dict=feedDict)
#writer.add_summary(summary,epoch*num_batches+batch)
batchLoss[batch] = l
predicted_gaps = self.predict_gaps(
lamb, batchSeqLengths, maxSessLen)
error = self.MSE(predicted_gaps, batchTargetList)
error_MAE = self.MAE(predicted_gaps, batchTargetList)
batchErr[batch] = error
batchErrMAE[batch] = error_MAE
#print(len(batchTargetList))
print(
'batch:{}/{},epoch:{}/{},train loss={:.3f},RMSE={:.3f},MAE={:.3f}'
.format(batch + 1, len(batchRandIxs), epoch + 1,
num_epochs, l, error, error_MAE))
#print(predicted_gaps)
#print(batchTargetList)
print("MAE after %d epoch is %.2f" %
(epoch + 1, np.mean(batchErrMAE)))
end = time.time()
delta_time = end - start
print('Average loss of the epoch is %.2f' %
np.mean(batchLoss))
print('Epoch ' + str(epoch + 1) + ' needs time:' +
str(delta_time) + ' s')
#Testing after ever epoch
num_b = len(b_t) / batch_size
batchErr = np.zeros(num_b)
batchErrMAE = np.zeros(num_b)
batchRandIxs = np.random.permutation(num_b)
for batch, batchOrigI in enumerate(batchRandIxs):
batchInputs_b, batchInputs_e, batchInputs_d, batchInputs_g, batchTargetList, batchSeqLengths = next(
load_batched_data_test(maxSessLen, batch_size, b_t,
e_t, d_t, g_t,
target_gaps_t))
feedDict = {
recom.inputb: batchInputs_b,
recom.inpute: batchInputs_e,
recom.inputg: batchInputs_g,
recom.inputd: batchInputs_d,
recom.target_gaps: batchTargetList,
recom.sessLengths: batchSeqLengths
}
__, l, lamb = sess.run(
[recom.optimizer, recom.loss, recom.lamb],
feed_dict=feedDict)
predicted_gaps = self.predict_gaps(
lamb, batchSeqLengths, maxSessLen)
error = self.MSE(predicted_gaps, batchTargetList)
error_MAE = self.MAE(predicted_gaps, batchTargetList)
batchErr[batch] = error
batchErrMAE[batch] = error_MAE
test_Err.append(np.mean(batchErr))
test_MAE.append(np.mean(batchErrMAE))
#print("RMSE error of test set is %.2f"% np.mean(batchErr))
print("MAE error of test set is %.2f\n" %
np.mean(batchErrMAE))
return test_Err, test_MAE
