def main(_):
''' Trains model from data '''
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
print('Created training directory', FLAGS.train_dir)
word_vocab, word_tensors, max_doc_length, label_tensors = \
load_data(FLAGS.data_dir, FLAGS.max_doc_length, FLAGS.max_sen_length)
train_reader = DataReader(word_tensors['train'], label_tensors['train'],
FLAGS.batch_size)
valid_reader = DataReader(word_tensors['valid'], label_tensors['valid'],
FLAGS.batch_size)
test_reader = DataReader(word_tensors['test'], label_tensors['test'],
FLAGS.batch_size)
print('initialized all dataset readers')
with tf.Graph().as_default(), tf.Session() as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.param_init,
FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = build_model(word_vocab, max_doc_length, train=True)
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=50)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = build_model(word_vocab, max_doc_length, train=False)
if FLAGS.load_model:
saver.restore(session, FLAGS.load_model)
print('Loaded model from', FLAGS.load_model,
'saved at global step', train_model.global_step.eval())
else:
tf.global_variables_initializer().run()
session.run(train_model.clear_word_embedding_padding)
print('Created and initialized fresh model. Size:',
model.model_size())
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate,
FLAGS.learning_rate), )
''' training starts here '''
best_valid_loss = None
#rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(FLAGS.max_epochs):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
loss, _, gradient_norm, step, _ = session.run(
[
train_model.loss, train_model.train_op,
train_model.global_norm, train_model.global_step,
train_model.clear_word_embedding_padding
], {
train_model.input: x,
train_model.targets: y,
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
print(
'%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), time_elapsed, gradient_norm))
print('Epoch training time:', time.time() - epoch_start_time)
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
#rnn_state = session.run(valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss = session.run(valid_model.loss, {
valid_model.input: x,
valid_model.targets: y,
})
if count % FLAGS.print_every == 0:
print("\t> validation loss = %6.8f, perplexity = %6.8f" %
(loss, np.exp(loss)))
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:", epoch)
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss", simple_value=avg_valid_loss)
])
summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(
best_valid_loss) - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
def main(_):
''' Trains model from data '''
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
print('Created training directory', FLAGS.train_dir)
word_vocab, char_vocab, word_tensors, char_tensors, max_word_length = \
load_data(FLAGS.data_dir, FLAGS.max_word_length, eos=FLAGS.EOS)
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
valid_reader = DataReader(word_tensors['valid'], char_tensors['valid'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
test_reader = DataReader(word_tensors['test'], char_tensors['test'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
print('initialized all dataset readers')
minimum_valid_ppl = 1000000
minimum_vl_epoch = 0
text_file = open("train_log.txt", "w")
# text_file.write("Purchase Amount: %s" % TotalAmount)
with tf.Graph().as_default(), tf.Session() as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.param_init,
FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=FLAGS.dropout)
train_model.update(
model.loss_graph(train_model.logits, FLAGS.batch_size,
FLAGS.num_unroll_steps))
# scaling loss by FLAGS.num_unroll_steps effectively scales gradients by the same factor.
# we need it to reproduce how the original Torch code optimizes. Without this, our gradients will be
# much smaller (i.e. 35 times smaller) and to get system to learn we'd have to scale learning rate and max_grad_norm appropriately.
# Thus, scaling gradients so that this trainer is exactly compatible with the original
train_model.update(
model.training_graph(train_model.loss * FLAGS.num_unroll_steps,
FLAGS.learning_rate, FLAGS.max_grad_norm))
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=10)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=0.0)
valid_model.update(
model.loss_graph(valid_model.logits, FLAGS.batch_size,
FLAGS.num_unroll_steps))
if FLAGS.load_model:
saver.restore(session, FLAGS.load_model)
print('Loaded model from', FLAGS.load_model,
'saved at global step', train_model.global_step.eval())
else:
tf.global_variables_initializer().run()
session.run(train_model.clear_char_embedding_padding)
print('Created and initialized fresh model. Size:',
model.model_size())
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate,
FLAGS.learning_rate), )
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(FLAGS.max_epochs):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
loss, _, rnn_state, gradient_norm, step, _ = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state, train_model.global_norm,
train_model.global_step,
train_model.clear_char_embedding_padding
], {
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
print(
'%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), time_elapsed, gradient_norm))
text_file.write(
'%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f \n'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), time_elapsed, gradient_norm))
print('Epoch training time:', time.time() - epoch_start_time)
# text_file.write('Epoch training time:'+str( time.time()-epoch_start_time)
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss, rnn_state = session.run(
[valid_model.loss, valid_model.final_rnn_state], {
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
if count % FLAGS.print_every == 0:
print("\t> validation loss = %6.8f, perplexity = %6.8f" %
(loss, np.exp(loss)))
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:", epoch)
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
text_file.write("at the end of epoch:" + str(epoch) + '\n')
text_file.write("train loss = %6.8f, perplexity = %6.8f \n" %
(avg_train_loss, np.exp(avg_train_loss)))
text_file.write("validation loss = %6.8f, perplexity = %6.8f \n" %
(avg_valid_loss, np.exp(avg_valid_loss)))
if (np.exp(avg_valid_loss) < minimum_valid_ppl):
minimum_valid_ppl = np.exp(avg_valid_loss)
minimum_vl_epoch = epoch
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
elif (epoch % 4 == 0):
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss", simple_value=avg_valid_loss)
])
summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(
best_valid_loss) - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
print("----------------------------------------------")
print(
"Minimum Valid PPL is attained in epoch:%d and Validation PPL is %6.8f"
% (minimum_vl_epoch, minimum_valid_ppl))
def main(file,
batch_size=20,
num_unroll_steps=35,
char_embed_size=15,
rnn_size=650,
kernels="[1,2,3,4,5,6,7]",
kernel_features="[50,100,150,200,200,200,200]",
max_grad_norm=5.0,
learning_rate=1.0,
learning_rate_decay=0.5,
decay_when=1.0,
seed=3435,
param_init=0.05,
max_epochs=25,
print_every=5):
''' Trains model from data '''
if not os.path.exists(TRAINING_DIR):
os.mkdir(TRAINING_DIR)
print('Created training directory', TRAINING_DIR)
word_vocab, char_vocab, word_tensors, char_tensors, max_word_length = \
load_dataset()
print('initialized all dataset readers')
with tf.Graph().as_default(), tf.Session() as session:
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
batch_size, num_unroll_steps, char_vocab)
valid_reader = DataReader(word_tensors['valid'], char_tensors['valid'],
batch_size, num_unroll_steps, char_vocab)
test_reader = DataReader(word_tensors['test'], char_tensors['test'],
batch_size, num_unroll_steps, char_vocab)
# tensorflow seed must be inside graph
tf.set_random_seed(seed)
np.random.seed(seed=seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(param_init, param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = model.inference_graph(
char_vocab_size=char_vocab.size(),
word_vocab_size=word_vocab.size(),
char_embed_size=char_embed_size,
batch_size=batch_size,
rnn_size=rnn_size,
max_word_length=max_word_length,
kernels=eval(kernels),
kernel_features=eval(kernel_features),
num_unroll_steps=num_unroll_steps)
train_model.update(
model.loss_graph(train_model.logits, batch_size,
num_unroll_steps))
# scaling loss by FLAGS.num_unroll_steps effectively scales gradients by the same factor.
# we need it to reproduce how the original Torch code optimizes. Without this, our gradients will be
# much smaller (i.e. 35 times smaller) and to get system to learn we'd have to scale learning rate and max_grad_norm appropriately.
# Thus, scaling gradients so that this trainer is exactly compatible with the original
train_model.update(
model.training_graph(train_model.loss * num_unroll_steps,
learning_rate, max_grad_norm))
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=50)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = model.inference_graph(
char_vocab_size=char_vocab.size(),
word_vocab_size=word_vocab.size(),
char_embed_size=char_embed_size,
batch_size=batch_size,
rnn_size=rnn_size,
max_word_length=max_word_length,
kernels=eval(kernels),
kernel_features=eval(kernel_features),
num_unroll_steps=num_unroll_steps)
valid_model.update(
model.loss_graph(valid_model.logits, batch_size,
num_unroll_steps))
'''if load_model:
saver.restore(session, load_model)
print('Loaded model from', load_model, 'saved at global step', train_model.global_step.eval())
else:'''
tf.global_variables_initializer().run()
session.run(train_model.clear_char_embedding_padding)
print('Created and initialized fresh model. Size:', model.model_size())
summary_writer = tf.summary.FileWriter(TRAINING_DIR,
graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate, learning_rate), )
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(max_epochs):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
loss, _, rnn_state, gradient_norm, step, _ = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state, train_model.global_norm,
train_model.global_step,
train_model.clear_char_embedding_padding
], {
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % print_every == 0:
print(
'%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), time_elapsed, gradient_norm))
print('Epoch training time:', time.time() - epoch_start_time)
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss, rnn_state = session.run(
[valid_model.loss, valid_model.final_rnn_state], {
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
if count % print_every == 0:
print("\t> validation loss = %6.8f, perplexity = %6.8f" %
(loss, np.exp(loss)))
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:", epoch)
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
save_as = '%s/epoch%03d_%.4f.model' % (TRAINING_DIR, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss", simple_value=avg_valid_loss)
])
summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(
avg_valid_loss) > np.exp(best_valid_loss) - decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
def fitness(self, word_vocab, char_vocab, word_tensors, char_tensors,
max_word_length, evo_epoch):
self._model, self._valid_model, self._saver = self.update_graph(
word_vocab, char_vocab, max_word_length, evo_epoch)
self._max_word_length = max_word_length
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
valid_reader = DataReader(word_tensors['valid'], char_tensors['valid'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
with tf.Session(
graph=self._graph,
config=tf.ConfigProto(
allow_soft_placement=True,
inter_op_parallelism_threads=FLAGS.num_cpus)) as session:
# initialize model
tf.global_variables_initializer().run()
session.run(self._model.clear_char_embedding_padding)
print('[EVOLUTION] Epoch_%d Individual_%d. Size: %d' %
(evo_epoch, self._id_number, model.model_size()))
# self._summary_writer = tf.summary.FileWriter(self._individual_dir, graph=session.graph)
session.run(
tf.assign(self._model.learning_rate, FLAGS.learning_rate), )
# scout train
best_valid_loss = None
rnn_state = session.run(self._model.initial_rnn_state)
# train a mini
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
loss, _, rnn_state, gradient_norm, step, _ = session.run(
[
self._model.loss, self._model.train_op,
self._model.final_rnn_state, self._model.global_norm,
self._model.global_step,
self._model.clear_char_embedding_padding
], {
self._model.input: x,
self._model.targets: y,
self._model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
print('training time:', time.time() - epoch_start_time)
# time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(self._valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss, rnn_state = session.run(
[
self._valid_model.loss,
self._valid_model.final_rnn_state
], {
self._valid_model.input: x,
self._valid_model.targets: y,
self._valid_model.initial_rnn_state: rnn_state,
})
avg_valid_loss += loss / valid_reader.length
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
# save_as = '%s/epoch%03d_%.4f.model' % (self._individual_dir, evo_epoch, avg_valid_loss)
save_as = '%s/epoch%03d.model' % (self._individual_dir, evo_epoch)
# self._saver.save(session, save_as)
# print('Saved model', save_as)
''' write out summary events '''
# summary = tf.Summary(value=[
# tf.Summary.Value(tag="train_loss", simple_value=avg_train_loss),
# tf.Summary.Value(tag="valid_loss", simple_value=avg_valid_loss)
# ])
# self._summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(
best_valid_loss) - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(self._model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
self._fitness = best_valid_loss
return self._fitness
session.run(
self._model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
self._fitness = best_valid_loss
return self._fitness
def main(print):
''' Trains model from data '''
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
print('Created training directory' + FLAGS.train_dir)
# CSV initialize
pd.DataFrame(FLAGS.flag_values_dict(),
index=range(1)).to_csv(FLAGS.train_dir +
'/train_parameters.csv')
epochs_results = initialize_epoch_data_dict()
fasttext_model_path = None
if FLAGS.fasttext_model_path:
fasttext_model_path = FLAGS.fasttext_model_path
word_vocab, char_vocab, word_tensors, char_tensors, max_word_length, words_list = \
load_data(FLAGS.data_dir, FLAGS.max_word_length, eos=FLAGS.EOS)
fasttext_model = None
if 'fasttext' in FLAGS.embedding:
fasttext_model = FasttextModel(
fasttext_path=fasttext_model_path).get_fasttext_model()
train_ft_reader = DataReaderFastText(
words_list=words_list,
batch_size=FLAGS.batch_size,
num_unroll_steps=FLAGS.num_unroll_steps,
model=fasttext_model,
data='train')
valid_ft_reader = DataReaderFastText(
words_list=words_list,
batch_size=FLAGS.batch_size,
num_unroll_steps=FLAGS.num_unroll_steps,
model=fasttext_model,
data='valid')
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
valid_reader = DataReader(word_tensors['valid'], char_tensors['valid'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
test_reader = DataReader(word_tensors['test'], char_tensors['test'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
print('initialized all dataset readers')
with tf.Graph().as_default(), tf.Session() as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.param_init,
FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=FLAGS.dropout,
embedding=FLAGS.embedding,
fasttext_word_dim=300,
acoustic_features_dim=4)
train_model.update(
model.loss_graph(train_model.logits, FLAGS.batch_size,
FLAGS.num_unroll_steps))
train_model.update(
model.training_graph(train_model.loss * FLAGS.num_unroll_steps,
FLAGS.learning_rate, FLAGS.max_grad_norm))
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=50)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=0.0,
embedding=FLAGS.embedding,
fasttext_word_dim=300,
acoustic_features_dim=4)
valid_model.update(
model.loss_graph(valid_model.logits, FLAGS.batch_size,
FLAGS.num_unroll_steps))
if FLAGS.load_model_for_training:
saver.restore(session, FLAGS.load_model_for_training)
string = str('Loaded model from' +
str(FLAGS.load_model_for_training) +
'saved at global step' +
str(train_model.global_step.eval()))
print(string)
else:
tf.global_variables_initializer().run()
session.run(train_model.clear_char_embedding_padding)
string = str('Created and initialized fresh model. Size:' +
str(model.model_size()))
print(string)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate,
FLAGS.learning_rate), )
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(FLAGS.max_epochs):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
if fasttext_model:
iter_over = zip(train_reader.iter(), train_ft_reader.iter())
else:
iter_over = train_reader.iter()
for batch_kim, batch_ft in iter_over:
if fasttext_model:
x, y = batch_kim
else:
x, y = batch_kim, batch_ft
count += 1
start_time = time.time()
if fasttext_model:
ft_vectors = fasttext_model.wv[
words_list['train'][count]].reshape(
fasttext_model.wv.vector_size, 1)
loss, _, rnn_state, gradient_norm, step, _, probas = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state,
train_model.global_norm, train_model.global_step,
train_model.clear_char_embedding_padding
], {
train_model.input2: batch_ft,
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
else:
loss, _, rnn_state, gradient_norm, step, _ = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state,
train_model.global_norm, train_model.global_step,
train_model.clear_char_embedding_padding
], {
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
string = str(
'%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), time_elapsed, gradient_norm))
print(string)
string = str('Epoch training time:' +
str(time.time() - epoch_start_time))
print(string)
epochs_results['epoch_training_time'].append(
str(time.time() - epoch_start_time))
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for batch_kim, batch_ft in zip(valid_reader.iter(),
valid_ft_reader.iter()):
x, y = batch_kim
count += 1
start_time = time.time()
loss, rnn_state = session.run(
[valid_model.loss, valid_model.final_rnn_state], {
valid_model.input2: batch_ft,
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
if count % FLAGS.print_every == 0:
string = str(
"\t> validation loss = %6.8f, perplexity = %6.8f" %
(loss, np.exp(loss)))
print(string)
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:" + str(epoch))
epochs_results['epoch_number'].append(str(epoch))
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
epochs_results['train_loss'].append(avg_train_loss)
epochs_results['train_perplexity'].append(np.exp(avg_train_loss))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
epochs_results['validation_loss'].append(avg_valid_loss)
epochs_results['valid_perplexity'].append(np.exp(avg_valid_loss))
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model' + str(save_as))
epochs_results['model_name'].append(str(save_as))
epochs_results['learning_rate'].append(
str(session.run(train_model.learning_rate)))
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="train_perplexity",
simple_value=np.exp(avg_train_loss)),
tf.Summary.Value(tag="valid_loss",
simple_value=avg_valid_loss),
tf.Summary.Value(tag="valid_perplexity",
simple_value=np.exp(avg_valid_loss)),
])
summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(
best_valid_loss) - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
string = str('learning rate was:' + str(current_learning_rate))
print(string)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-3:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
string = str('new learning rate is:' +
str(current_learning_rate))
print(string)
else:
best_valid_loss = avg_valid_loss
# Save model performance data
pd.DataFrame(epochs_results).to_csv(FLAGS.train_dir + '/train_results.csv')
def main(_):
''' Trains model from data '''
print("we in main")
print(sys.argv[2])
print(FLAGS)
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
print('Created training directory', FLAGS.train_dir)
word_vocab, char_vocab, word_tensors, char_tensors, max_word_length = \
load_data(FLAGS.data_dir, FLAGS.max_word_length, eos=FLAGS.EOS)
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
valid_reader = DataReader(word_tensors['valid'], char_tensors['valid'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
test_reader = DataReader(word_tensors['test'], char_tensors['test'],
FLAGS.batch_size, FLAGS.num_unroll_steps)
print('initialized all dataset readers')
with tf.Graph().as_default(), tf.Session() as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.param_init, FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=FLAGS.dropout)
train_model.update(model.loss_graph(train_model.logits, FLAGS.batch_size, FLAGS.num_unroll_steps))
# scaling loss by FLAGS.num_unroll_steps effectively scales gradients by the same factor.
# we need it to reproduce how the original Torch code optimizes. Without this, our gradients will be
# much smaller (i.e. 35 times smaller) and to get system to learn we'd have to scale learning rate and max_grad_norm appropriately.
# Thus, scaling gradients so that this trainer is exactly compatible with the original
train_model.update(model.training_graph(train_model.loss * FLAGS.num_unroll_steps, FLAGS.learning_rate, FLAGS.max_grad_norm))
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=50)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=0.0)
valid_model.update(model.loss_graph(valid_model.logits, FLAGS.batch_size, FLAGS.num_unroll_steps))
with tf.variable_scope("Model", reuse=True):
test_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=1,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=1,
dropout=0.0)
test_model.update(model.loss_graph(test_model.logits, 1, 1))
if FLAGS.load_model:
saver.restore(session, FLAGS.load_model)
print('Loaded model from', FLAGS.load_model, 'saved at global step', train_model.global_step.eval())
else:
tf.initialize_all_variables().run()
print('Created and initialized fresh model. Size:', model.model_size())
summary_writer = tf.train.SummaryWriter(FLAGS.train_dir, graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(
tf.assign(train_model.learning_rate, FLAGS.learning_rate),
)
def clear_char_embedding_padding():
char_embedding = session.run(train_model.char_embedding)
char_embedding[0,:] = 0.0
session.run(tf.assign(train_model.char_embedding, char_embedding))
char_embedding = session.run(train_model.char_embedding)
clear_char_embedding_padding()
run_test2(session, test_model, train_reader)
#exit(1)
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(FLAGS.max_epochs):
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
print (x)
exit(1)
loss, _, rnn_state, gradient_norm, step = session.run([
train_model.loss,
train_model.train_op,
train_model.final_rnn_state,
train_model.global_norm,
train_model.global_step,
], {
train_model.input : x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
clear_char_embedding_padding()
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
print('%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f' % (step,
epoch, count,
train_reader.length,
loss, np.exp(loss),
time_elapsed,
gradient_norm))
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss, rnn_state = session.run([
valid_model.loss,
valid_model.final_rnn_state
], {
valid_model.input : x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
if count % FLAGS.print_every == 0:
print("\t> validation loss = %6.8f, perplexity = %6.8f" % (loss, np.exp(loss)))
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:", epoch)
print("train loss = %6.8f, perplexity = %6.8f" % (avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" % (avg_valid_loss, np.exp(avg_valid_loss)))
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch, avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss", simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss", simple_value=avg_valid_loss)
])
summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(best_valid_loss) - FLAGS.decay_when:
print('** validation perplexity did not improve enough, decay learning rate')
current_learning_rate = session.run(train_model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
break
session.run(train_model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
run_test2(session, test_model, train_reader)
print ("AGAIN")
run_test2(session, test_model, train_reader)
def main(_):
''' Trains model from data '''
date = time.strftime("%Y-%m-%d %H-%M-%S", time.localtime())
directory = FLAGS.train_dir.format(date)
if not os.path.exists(directory):
os.makedirs(directory)
print('Created training directory', directory)
word_vocab, char_vocab, word_tensors, char_tensors, max_word_length = \
load_data(FLAGS.data_dir, FLAGS.max_word_length, eos=FLAGS.EOS)
char_embedding_metadata = os.path.join(directory +
"characters_embeddings.tsv")
with open(char_embedding_metadata, "w", encoding="utf-8") as metadata_file:
metadata_file.write('padding\n')
for i in range(1, char_vocab.size):
metadata_file.write('%s\n' % (char_vocab.token(i)))
word_embedding_metadata = os.path.join(directory + "words_embeddings.tsv")
with open(word_embedding_metadata, "w", encoding="utf-8") as metadata_file:
metadata_file.write('padding\n')
for i in range(1, word_vocab.size):
metadata_file.write('%s\n' % (word_vocab.token(i)))
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
FLAGS.batch_size, FLAGS.num_unroll_steps,
char_vocab)
valid_reader = DataReader(word_tensors['valid'], char_tensors['valid'],
FLAGS.batch_size, FLAGS.num_unroll_steps,
char_vocab)
test_reader = DataReader(word_tensors['test'], char_tensors['test'],
FLAGS.batch_size, FLAGS.num_unroll_steps,
char_vocab)
print('initialized all dataset readers')
config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC'
with tf.Graph().as_default(), tf.Session(config=config) as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
initializer = tf.random_uniform_initializer(-FLAGS.param_init,
FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = Model(flags=FLAGS,
char_vocab=char_vocab,
word_vocab=word_vocab,
max_word_length=max_word_length,
metadata=char_embedding_metadata)
with tf.variable_scope("Model", reuse=True):
valid_model = Model(FLAGS, char_vocab, word_vocab, max_word_length,
ModelUsage.VALIDATE, char_embedding_metadata)
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=50)
if FLAGS.load_model:
saver.restore(session, FLAGS.load_model)
print('Loaded model from', FLAFS.load_model,
'saved at global step', train_model.global_step.eval())
else:
tf.global_variables_initializer().run()
session.run(train_model.clear_char_embedding_padding)
print('Created and initialized fresh model. Size:',
model.model_size())
summary_writer = tf.summary.FileWriter(directory, graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate,
FLAGS.learning_rate), )
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(FLAGS.max_epochs):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
loss, _, rnn_state, gradient_norm, step, _ = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state, train_model.global_norm,
train_model.global_step,
train_model.clear_char_embedding_padding
], {
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
print(
'%6d: %d [%5d/%5d], train_loss/perplexity = %6.8f/%6.7f secs/batch = %.4fs, grad.norm=%6.8f'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), time_elapsed, gradient_norm))
print('Epoch training time:', time.time() - epoch_start_time)
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss, rnn_state = session.run(
[valid_model.loss, valid_model.final_rnn_state], {
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
if count % FLAGS.print_every == 0:
print("\t> validation loss = %6.8f, perplexity = %6.8f" %
(loss, np.exp(loss)))
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:", epoch)
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
save_as = '%s/epoch%03d_%.4f.model' % (directory, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss",
simple_value=avg_valid_loss),
tf.Summary.Value(tag="train_perplexity",
simple_value=np.exp(avg_train_loss)),
tf.Summary.Value(tag="valid_perplexity",
simple_value=np.exp(avg_valid_loss))
])
summary_writer.add_summary(summary, step)
projector.visualize_embeddings(summary_writer,
valid_model.projector_config)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(
best_valid_loss) - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
def main(print):
''' Trains model from data '''
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
print('Created training directory' + FLAGS.train_dir)
# CSV initialize
df_train_params = pd.DataFrame(FLAGS.flag_values_dict(), index=range(1))
df_train_params['comment'] = ''
df_train_params.to_csv(FLAGS.train_dir + '/train_parameters.csv')
epochs_results = initialize_epoch_data_dict()
fasttext_model_path = None
if FLAGS.fasttext_model_path:
fasttext_model_path = FLAGS.fasttext_model_path
word_vocab, char_vocab, word_tensors, char_tensors, max_word_length, words_list, wers, acoustics = \
load_data(FLAGS.data_dir, FLAGS.max_word_length, num_unroll_steps=FLAGS.num_unroll_steps, eos=FLAGS.EOS, batch_size=FLAGS.batch_size)
word_vocab_valid, char_vocab_valid, word_tensors_valid, char_tensors_valid, max_word_length_valid, words_list_valid, wers_valid,\
acoustics_valid, files_name_valid, kaldi_sents_index_valid = \
load_test_data(FLAGS.data_dir, FLAGS.max_word_length, num_unroll_steps=FLAGS.num_unroll_steps, eos=FLAGS.EOS,
datas=['valid'])
fasttext_model = None
if 'fasttext' in FLAGS.embedding:
fasttext_model = FasttextModel(
fasttext_path=fasttext_model_path).get_fasttext_model()
train_ft_reader = DataReaderFastText(
words_list=words_list,
batch_size=FLAGS.batch_size,
num_unroll_steps=FLAGS.num_unroll_steps,
model=fasttext_model,
data='train',
acoustics=acoustics)
valid_ft_reader = DataReaderFastText(
words_list=words_list,
batch_size=FLAGS.batch_size,
num_unroll_steps=FLAGS.num_unroll_steps,
model=fasttext_model,
data='valid',
acoustics=acoustics)
train_reader = DataReader(word_tensors['train'], char_tensors['train'],
FLAGS.batch_size, FLAGS.num_unroll_steps,
wers['train'])
valid_reader = TestDataReader(word_tensors_valid['valid'],
char_tensors_valid['valid'],
FLAGS.batch_size, FLAGS.num_unroll_steps,
wers_valid['valid'],
files_name_valid['valid'],
kaldi_sents_index_valid['valid'])
# test_reader = DataReader(word_tensors['test'], char_tensors['test'],
# FLAGS.batch_size, FLAGS.num_unroll_steps, wers['train'], word_vocab, char_vocab)
print('initialized all dataset readers')
with tf.Graph().as_default(), tf.Session() as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.param_init,
FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = model.inference_graph(
char_vocab_size=char_vocab.size,
word_vocab_size=word_vocab.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=FLAGS.dropout,
embedding=FLAGS.embedding,
fasttext_word_dim=300,
acoustic_features_dim=4)
train_model.update(
model.loss_graph(train_model.logits, FLAGS.batch_size))
# scaling loss by FLAGS.num_unroll_steps effectively scales gradients by the same factor.
# we need it to reproduce how the original Torch code optimizes. Without this, our gradients will be
# much smaller (i.e. 35 times smaller) and to get system to learn we'd have to scale learning rate and max_grad_norm appropriately.
# Thus, scaling gradients so that this trainer is exactly compatible with the original
train_model.update(
model.training_graph(train_model.loss * FLAGS.num_unroll_steps,
FLAGS.learning_rate, FLAGS.max_grad_norm))
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = model.inference_graph(
char_vocab_size=char_vocab_valid.size,
word_vocab_size=word_vocab_valid.size,
char_embed_size=FLAGS.char_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
max_word_length=max_word_length,
kernels=eval(FLAGS.kernels),
kernel_features=eval(FLAGS.kernel_features),
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=0.0,
embedding=FLAGS.embedding,
fasttext_word_dim=300,
acoustic_features_dim=4)
valid_model.update(
model.loss_graph(valid_model.logits, FLAGS.batch_size))
# create saver before creating more graph nodes, so that we do not save any vars defined below
if FLAGS.load_model_for_training:
# delete last layers (softmax) - SimpleLinear/Matrix + Bias
variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
subset_grpah_for_loading = variables[:29] + variables[31:]
loader = tf.train.Saver(max_to_keep=50,
var_list=subset_grpah_for_loading)
saver = tf.train.Saver(max_to_keep=50)
if FLAGS.load_model_for_training:
loader.restore(session, FLAGS.load_model_for_training)
string = str('Loaded model from' +
str(FLAGS.load_model_for_training) +
'saved at global step' +
str(train_model.global_step.eval()))
print(string)
session.run(tf.variables_initializer(var_list=variables[29:31]))
string = str('initialized specific scope for fresh model. Size:' +
str(model.model_size()))
print(string)
else:
tf.global_variables_initializer().run()
session.run(train_model.clear_char_embedding_padding)
string = str('Created and initialized fresh model. Size:' +
str(model.model_size()))
print(string)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate,
FLAGS.learning_rate), )
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(FLAGS.max_epochs):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for batch_kim, batch_ft in zip(train_reader.iter(),
train_ft_reader.iter()):
x, y = batch_kim
count += 1
start_time = time.time()
if fasttext_model:
ft_vectors = fasttext_model.wv[
words_list['train'][count]].reshape(
fasttext_model.wv.vector_size, 1)
loss, _, rnn_state, gradient_norm, step, _, logits = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state,
train_model.global_norm, train_model.global_step,
train_model.clear_char_embedding_padding,
train_model.logits
], {
train_model.input2: batch_ft,
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
else:
loss, _, rnn_state, gradient_norm, step, _, logits = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state,
train_model.global_norm, train_model.global_step,
train_model.clear_char_embedding_padding,
train_model.logits
], {
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
string = str(
'%6d: %d [%5d/%5d], train_loss = %6.8f secs/batch = %.4fs'
% (step, epoch, count, train_reader.length, loss,
time_elapsed))
print(string)
string = str('Epoch training time:' +
str(time.time() - epoch_start_time))
print(string)
epochs_results['epoch_training_time'].append(
str(time.time() - epoch_start_time))
# epoch done: time to evaluate
avg_valid_loss = 0.
labels = []
predictions = []
files_name_list = []
kaldi_sents_index_list = []
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for batch_kim, batch_ft in zip(valid_reader.iter(),
valid_ft_reader.iter()):
x, y, files_name_batch, kaldi_sents_index_batch = batch_kim
count += 1
start_time = time.time()
loss, logits = session.run(
[valid_model.loss, valid_model.logits], {
valid_model.input2: batch_ft,
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
labels.append(y)
predictions.append(logits)
files_name_list.append(files_name_batch)
kaldi_sents_index_list.append(kaldi_sents_index_batch)
if count % FLAGS.print_every == 0:
string = str("\t> validation loss = %6.8f" % (loss))
print(string)
avg_valid_loss = get_valid_rescore_loss(labels, predictions,
files_name_list,
kaldi_sents_index_list)
print("at the end of epoch:" + str(epoch))
epochs_results['epoch_number'].append(str(epoch))
print("train loss = %6.8f" % (avg_train_loss))
epochs_results['train_loss'].append(avg_train_loss)
print("validation loss = %6.8f" % (avg_valid_loss))
epochs_results['validation_loss'].append(avg_valid_loss)
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model' + str(save_as))
epochs_results['model_name'].append(str(save_as))
epochs_results['learning_rate'].append(
str(session.run(train_model.learning_rate)))
current_learning_rate = session.run(train_model.learning_rate)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and avg_valid_loss > best_valid_loss - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
string = str('learning rate was:' + str(current_learning_rate))
print(string)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-6:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
string = str('new learning rate is:' +
str(current_learning_rate))
print(string)
else:
best_valid_loss = avg_valid_loss
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss",
simple_value=avg_valid_loss),
tf.Summary.Value(tag="learning_rate",
simple_value=current_learning_rate)
])
summary_writer.add_summary(summary, step)
# Save model performance data
pd.DataFrame(epochs_results).to_csv(FLAGS.train_dir + '/train_results.csv')
def main(_):
''' Trains model from data '''
min = [1000, 1000, 1000, 1000] # [t_loss, t_ppl, v_loss, v_ppl]
total_time = 0.
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
print('Created training directory', FLAGS.train_dir)
word_vocab, \
char_vocab, \
word_tensors, \
char_tensors, \
max_word_length = load_data(FLAGS.data_dir, FLAGS.max_word_length, flist = FILE_NAME_LIST, eos=FLAGS.EOS)
train_reader = DataReader(word_tensors[FILE_NAME_LIST[0]],
FLAGS.batch_size, FLAGS.num_unroll_steps)
valid_reader = DataReader(word_tensors[FILE_NAME_LIST[1]],
FLAGS.batch_size, FLAGS.num_unroll_steps)
test_reader = DataReader(word_tensors[FILE_NAME_LIST[2]], FLAGS.batch_size,
FLAGS.num_unroll_steps)
print('initialized all dataset readers')
with tf.Graph().as_default(), tf.Session() as session:
# tensorflow seed must be inside graph
tf.set_random_seed(FLAGS.seed)
np.random.seed(seed=FLAGS.seed)
''' build training graph '''
initializer = tf.random_uniform_initializer(-FLAGS.param_init,
FLAGS.param_init)
with tf.variable_scope("Model", initializer=initializer):
train_model = model.inference_graph(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=FLAGS.dropout)
train_model.update(
model.loss_graph(train_model.logits, FLAGS.batch_size,
FLAGS.num_unroll_steps))
# scaling loss by FLAGS.num_unroll_steps effectively scales gradients by the same factor.
# we need it to reproduce how the original Torch code optimizes. Without this, our gradients will be
# much smaller (i.e. 35 times smaller) and to get system to learn we'd have to scale learning rate and max_grad_norm appropriately.
# Thus, scaling gradients so that this trainer is exactly compatible with the original
train_model.update(
model.training_graph(train_model.loss * FLAGS.num_unroll_steps,
FLAGS.learning_rate, FLAGS.max_grad_norm))
# create saver before creating more graph nodes, so that we do not save any vars defined below
saver = tf.train.Saver(max_to_keep=5)
''' build graph for validation and testing (shares parameters with the training graph!) '''
with tf.variable_scope("Model", reuse=True):
valid_model = model.inference_graph(
word_vocab_size=word_vocab.size,
word_embed_size=FLAGS.word_embed_size,
batch_size=FLAGS.batch_size,
num_highway_layers=FLAGS.highway_layers,
num_rnn_layers=FLAGS.rnn_layers,
rnn_size=FLAGS.rnn_size,
num_unroll_steps=FLAGS.num_unroll_steps,
dropout=0.0)
valid_model.update(
model.loss_graph(valid_model.logits, FLAGS.batch_size,
FLAGS.num_unroll_steps))
if FLAGS.load_model:
saver.restore(session, FLAGS.load_model)
print('Loaded model from', FLAGS.load_model,
'saved at global step', train_model.global_step.eval())
else:
tf.global_variables_initializer().run()
session.run(train_model.clear_char_embedding_padding)
print('Created and initialized fresh model. Size:',
model.model_size())
summary_writer = tf.summary.FileWriter(FLAGS.train_dir,
graph=session.graph)
''' take learning rate from CLI, not from saved graph '''
session.run(tf.assign(train_model.learning_rate, FLAGS.learning_rate))
print("=" * 89)
print("=" * 89)
all_weights = {v.name: v for v in tf.trainable_variables()}
total_size = 0
pi = 1 # 0 is for sum of grad_sses
for v_name in list(all_weights): # sorted()
v = all_weights[v_name]
v_size = int(np.prod(np.array(v.shape.as_list())))
print("%02d-Weight %s\tshape %s\ttsize %d" %
(pi, v.name[:-2].ljust(80), str(v.shape).ljust(20), v_size))
total_size += v_size
pi += 1
print("Total size %d, %.3fMiB" % (total_size,
(total_size * 4) / (1024 * 1024)))
print("-" * 89)
''' training starts here '''
best_valid_loss = None
rnn_state = session.run(train_model.initial_rnn_state)
for epoch in range(1, FLAGS.max_epochs + 1):
epoch_start_time = time.time()
avg_train_loss = 0.0
count = 0
for x, y in train_reader.iter():
count += 1
start_time = time.time()
loss, _, rnn_state, gradient_norm, step, _ = session.run(
[
train_model.loss, train_model.train_op,
train_model.final_rnn_state, train_model.global_norm,
train_model.global_step,
train_model.clear_char_embedding_padding
], {
train_model.input: x,
train_model.targets: y,
train_model.initial_rnn_state: rnn_state
})
avg_train_loss += 0.05 * (loss - avg_train_loss)
time_elapsed = time.time() - start_time
if count % FLAGS.print_every == 0:
cur_lr = session.run(train_model.learning_rate)
print(
'%6d: -%d- [%5d/%5d], train_loss/ppl = %6.8f/%6.7f batch/secs = %.1fb/s, cur_lr = %2.5f, grad.norm=%6.8f'
% (step, epoch, count, train_reader.length, loss,
np.exp(loss), FLAGS.print_every / time_elapsed,
cur_lr, gradient_norm))
print('Epoch training time:', time.time() - epoch_start_time)
total_time += (time.time() - epoch_start_time)
# epoch done: time to evaluate
avg_valid_loss = 0.0
count = 0
rnn_state = session.run(valid_model.initial_rnn_state)
for x, y in valid_reader.iter():
count += 1
start_time = time.time()
loss, rnn_state = session.run(
[valid_model.loss, valid_model.final_rnn_state], {
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: rnn_state,
})
if count % FLAGS.print_every == 0:
print("\t> validation loss = %6.8f, perplexity = %6.8f" %
(loss, np.exp(loss)))
avg_valid_loss += loss / valid_reader.length
print("at the end of epoch:", epoch)
print("train loss = %6.8f, perplexity = %6.8f" %
(avg_train_loss, np.exp(avg_train_loss)))
print("validation loss = %6.8f, perplexity = %6.8f" %
(avg_valid_loss, np.exp(avg_valid_loss)))
if min[2] > avg_valid_loss:
min[0] = avg_train_loss
min[1] = np.exp(avg_train_loss)
min[2] = avg_valid_loss
min[3] = np.exp(avg_valid_loss)
save_as = '%s/epoch%03d_%.4f.model' % (FLAGS.train_dir, epoch,
avg_valid_loss)
saver.save(session, save_as)
print('Saved model', save_as)
''' write out summary events '''
summary = tf.Summary(value=[
tf.Summary.Value(tag="train_loss",
simple_value=avg_train_loss),
tf.Summary.Value(tag="valid_loss", simple_value=avg_valid_loss)
])
summary_writer.add_summary(summary, step)
''' decide if need to decay learning rate '''
if best_valid_loss is not None and np.exp(avg_valid_loss) > np.exp(
best_valid_loss) - FLAGS.decay_when:
print(
'validation perplexity did not improve enough, decay learning rate'
)
current_learning_rate = session.run(train_model.learning_rate)
print('learning rate was:', current_learning_rate)
current_learning_rate *= FLAGS.learning_rate_decay
if current_learning_rate < 1.e-5:
print('learning rate too small - stopping now')
break
session.run(
train_model.learning_rate.assign(current_learning_rate))
print('new learning rate is:', current_learning_rate)
else:
best_valid_loss = avg_valid_loss
''' test on the test set '''
ave_test_loss = 0.
trnn_state = session.run(valid_model.initial_rnn_state)
for x, y in test_reader.iter():
loss, trnn_state = session.run(
[valid_model.loss, valid_model.final_rnn_state], {
valid_model.input: x,
valid_model.targets: y,
valid_model.initial_rnn_state: trnn_state
})
disp_loss = loss
ave_test_loss += disp_loss / test_reader.length
print("=" * 89)
print("=" * 89)
print("Total training time(not included the valid time): %f" %
total_time)
print("The best result:")
print("train loss = %.3f, ppl = %.4f" % (min[0], min[1]))
print("valid loss = %.3f, ppl = %.4f" % (min[2], min[3]))
print("test loss = %.3f, ppl = %.4f" %
(ave_test_loss, np.exp(ave_test_loss)))
print("=" * 89)
