def _decode_inference_indices(model, sess, output_infer,
output_infer_summary_prefix, inference_indices,
tgt_sos, tgt_eos, bpe_delimiter):
"""Decoding only a specific set of sentences."""
utils.print_out(" decoding to output %s , num sents %d." %
(output_infer, len(inference_indices)))
start_time = time.time()
with codecs.getwriter("utf-8")(tf.gfile.GFile(output_infer,
mode="wb")) as trans_f:
trans_f.write("") # Write empty string to ensure file is created.
for decode_id in inference_indices:
nmt_outputs, infer_summary = model.decode(sess)
# get text translation
assert nmt_outputs.shape[0] == 1
translation = nmt_utils.get_translation(
nmt_outputs,
sent_id=0,
tgt_sos=tgt_sos,
tgt_eos=tgt_eos,
bpe_delimiter=bpe_delimiter)
if infer_summary is not None: # Attention models
image_file = output_infer_summary_prefix + str(
decode_id) + ".png"
utils.print_out(" save attention image to %s*" % image_file)
image_summ = tf.Summary()
image_summ.ParseFromString(infer_summary)
with tf.gfile.GFile(image_file, mode="w") as img_f:
img_f.write(image_summ.value[0].image.encoded_image_string)
trans_f.write("%s\n" % translation)
utils.print_out(b"%s\n" % translation)
utils.print_time(" done", start_time)
def compute_perplexity(model, sess, name):
"""Compute perplexity of the output of the model.
Args:
model: model for compute perplexity.
sess: tensorflow session to use.
name: name of the batch.
Returns:
The perplexity of the eval outputs.
"""
total_loss = 0
total_predict_count = 0
start_time = time.time()
while True:
try:
loss, predict_count, batch_size = model.eval(sess)
total_loss += loss * batch_size
total_predict_count += predict_count
except tf.errors.OutOfRangeError:
break
perplexity = utils.safe_exp(total_loss / total_predict_count)
utils.print_time(" eval %s: perplexity %.2f" % (name, perplexity),
start_time)
return perplexity
def decode_and_evaluate(name,
model,
sess,
trans_file,
ref_file,
metrics,
subword_option,
beam_width,
tgt_eos,
num_translations_per_input=1,
decode=True):
"""Decode a test set and compute a score according to the evaluation task."""
# Decode
if decode:
utils.print_out(" decoding to output %s." % trans_file)
start_time = time.time()
num_sentences = 0
with codecs.getwriter("utf-8")(
tf.gfile.GFile(trans_file, mode="wb")) as trans_f:
trans_f.write("") # Write empty string to ensure file is created.
num_translations_per_input = max(
min(num_translations_per_input, beam_width), 1)
while True:
try:
nmt_outputs, _ = model.decode(sess)
if beam_width == 0:
nmt_outputs = np.expand_dims(nmt_outputs, 0)
batch_size = nmt_outputs.shape[1]
num_sentences += batch_size
for sent_id in range(batch_size):
for beam_id in range(num_translations_per_input):
translation = get_translation(
nmt_outputs[beam_id],
sent_id,
tgt_eos=tgt_eos,
subword_option=subword_option)
trans_f.write((translation + b"\n").decode("utf-8"))
except tf.errors.OutOfRangeError:
utils.print_time(
" done, num sentences %d, num translations per input %d" %
(num_sentences, num_translations_per_input), start_time)
break
# Evaluation
evaluation_scores = {}
if ref_file and tf.gfile.Exists(trans_file):
for metric in metrics:
score = evaluation_utils.evaluate(
ref_file,
trans_file,
metric,
subword_option=subword_option)
evaluation_scores[metric] = score
utils.print_out(" %s %s: %.1f" % (metric, name, score))
return evaluation_scores
def decode_and_evaluate(name,
model,
sess,
trans_file,
ref_file,
metrics,
bpe_delimiter,
beam_width,
tgt_sos,
tgt_eos,
decode=True):
"""Decode a test set and compute a score according to the evaluation task."""
# Decode
if decode:
utils.print_out(" decoding to output %s." % trans_file)
start_time = time.time()
num_sentences = 0
with codecs.getwriter("utf-8")(tf.gfile.GFile(trans_file,
mode="wb")) as trans_f:
trans_f.write("") # Write empty string to ensure file is created.
while True:
try:
nmt_outputs, _, _ = model.decode(sess)
if beam_width > 0:
# get the top translation.
nmt_outputs = nmt_outputs[0]
num_sentences += len(nmt_outputs)
for sent_id in range(len(nmt_outputs)):
translation = get_translation(
nmt_outputs,
sent_id,
tgt_sos=tgt_sos,
tgt_eos=tgt_eos,
bpe_delimiter=bpe_delimiter)
trans_f.write((translation + b"\n").decode("utf-8"))
except tf.errors.OutOfRangeError:
utils.print_time(
" done, num sentences %d" % num_sentences, start_time)
break
# Evaluation
evaluation_scores = {}
if ref_file and tf.gfile.Exists(trans_file):
for metric in metrics:
score = evaluation_utils.evaluate(ref_file,
trans_file,
metric,
bpe_delimiter=bpe_delimiter)
evaluation_scores[metric] = score
utils.print_out(" %s %s: %.1f" % (metric, name, score))
return evaluation_scores
def decode_and_evaluate(name,
model,
sess,
output_file,
reference_file,
metrics,
bpe_delimiter,
beam_width,
eos,
number_token=None,
name_token=None,
decode=True):
"""Decode a test set and compute a score according to the evaluation task."""
# Decode
if decode:
utils.print_out(" decoding to output %s." % output_file)
start_time = time.time()
num_sentences = 0
with tf.gfile.GFile(output_file, mode="w+") as out_f:
out_f.write("") # Write empty string to ensure file is created.
while True:
try:
# Get the response(s) for each input in the batch (whole file in this case)
# ToDo: adapt for architectures
outputs, infer_summary = model.decode(sess)
if beam_width > 0:
# Get the top response if we used beam_search
outputs = outputs[0]
num_sentences += len(outputs)
# Iterate over the outputs an write them to file
for sent_id in range(len(outputs)):
response = postprocess_output(outputs, sent_id, eos,
bpe_delimiter,
number_token, name_token)
out_f.write("%s\n" % response)
except tf.errors.OutOfRangeError:
utils.print_time(
" done, num sentences %d" % num_sentences, start_time)
break
# Evaluation
evaluation_scores = {}
if reference_file and tf.gfile.Exists(output_file):
for metric in metrics:
score = evaluation_utils.evaluate(ref_file=reference_file,
trans_file=output_file,
metric=metric,
bpe_delimiter=bpe_delimiter)
evaluation_scores[metric] = score
utils.print_out(" %s %s: %.1f" % (metric, name, score))
return evaluation_scores
def _decode_inference_indices(model,
sess,
output_infer_file,
output_infer_summary_prefix,
inference_indices,
eos,
bpe_delimiter,
number_token=None,
name_token=None):
"""
Decoding only a specific set of sentences indicated by inference_indices
:param output_infer:
:param output_infer_summary_prefix:
:param inference_indices: A list of sentence indices
:param eos: the eos token
:param bpe_delimiter: delimiter used for byte-pair entries
:return:
"""
utils.print_out(" decoding to output %s , num sents %d." %
(output_infer_file, len(inference_indices)))
start_time = time.time()
with codecs.getwriter("utf-8")(tf.gfile.GFile(output_infer_file,
'wb')) as f:
f.write("") # Write empty string to ensure that the file is created
# Get the outputs
outputs, infer_summary = model.decode(sess)
# Iterate over the sentences we want to process. Use the index to process sentences and the
# decode_id to create logs
for sentence_id, decode_id in enumerate(inference_indices):
# Get the response
response = chatbot_utils.postprocess_output(
outputs,
sentence_id=sentence_id,
eos=eos,
bpe_delimiter=bpe_delimiter,
number_token=number_token,
name_token=name_token)
# TODO: add inference_summary if deciding to use attention
# Write the response to file
f.write("%s\n" % response)
utils.print_out("%s\n" % response)
utils.print_time(" done", start_time)
def decode_and_evaluate(mode,
sess,
out_tensor,
trans_file,
ref_file,
metric='bleu',
beam_width=10,
num_translations_per_input=1,
iterations=1):
"""Decode a test set and compute a score according to the evaluation task."""
utils.print_out(" Decoding to output %s" % trans_file)
with codecs.getwriter("utf-8")(tf.io.gfile.GFile(trans_file,
mode="wb")) as trans_f:
trans_f.write("") # Write empty string to ensure file is created.
num_translations_per_input = min(num_translations_per_input,
beam_width)
print(" Running inference with beam_width %g, num translations per input %d. " \
% (beam_width, num_translations_per_input))
print(" Total iterations count %d." % iterations)
# Warmup for the first batch to take out the very first runtime
# session overhead.
nmt_outputs = sess.run(out_tensor) # time x batch_size x beam_width
nmt_outputs = nmt_outputs.transpose() # beam_width x batch_size x time
batch_size = nmt_outputs.shape[1]
for sent_id in range(batch_size):
translation = get_translation(nmt_outputs[0],
sent_id,
tgt_eos='</s>')
if mode == 'accuracy':
trans_f.write((translation + b"\n").decode("utf-8"))
# prediction time is the time for the model prediction only
# overall time is the time for data pre-processing and data post-processing
prediction_times = list()
overall_start = time.time()
num_sentences = 0
n = 0
while n < iterations:
n += 1
while True:
try:
start = time.time()
nmt_outputs = sess.run(
out_tensor) # time x batch_size x beam_width
nmt_outputs = nmt_outputs.transpose(
) # beam_width x batch_size x time
prediction_times.append(time.time() - start)
batch_size = nmt_outputs.shape[1]
num_sentences += batch_size
for sent_id in range(batch_size):
for beam_id in range(num_translations_per_input):
translation = get_translation(nmt_outputs[beam_id],
sent_id,
tgt_eos='</s>')
if mode == 'accuracy':
trans_f.write(
(translation + b"\n").decode("utf-8"))
except tf.errors.OutOfRangeError:
utils.print_time(
" Done, num sentences %d, num translations per input %d"
% (num_sentences, num_translations_per_input),
overall_start)
break
overall_time = (time.time() - overall_start)
print("\nAverage Prediction Latency: {:.5f} sec per batch.".format(
sum(prediction_times) / float(len(prediction_times))))
print("Overall Latency: {:.5f} sec for the entire test "
"dataset.".format(overall_time / float(iterations)))
print("Overall Throughput : {:.3f} sentences per sec.".format(
num_sentences / float(overall_time)))
# Evaluation
if mode == 'accuracy':
if ref_file and tf.io.gfile.exists(trans_file):
score = evaluation_utils.evaluate(ref_file, trans_file, metric)
utils.print_out(" Accuracy metric %s: %.1f" % (metric, score))
def train(hparams, scope=None, target_session=''):
"""Train the chatbot"""
# Initialize some local hyperparameters
log_device_placement = hparams.log_device_placement
out_dir = hparams.out_dir
num_train_steps = hparams.num_train_steps
steps_per_stats = hparams.steps_per_stats
steps_per_external_eval = hparams.steps_per_external_eval
steps_per_eval = 10 * steps_per_stats
if not steps_per_external_eval:
steps_per_external_eval = 5 * steps_per_eval
if hparams.architecture == "simple":
model_creator = SimpleModel
get_infer_iterator = iterator_utils.get_infer_iterator
get_iterator = iterator_utils.get_iterator
elif hparams.architecture == "hier":
model_creator = HierarchicalModel
# Parse some of the arguments now
def curry_get_infer_iterator(dataset, vocab_table, batch_size, src_reverse,
eos, src_max_len):
return end2end_iterator_utils.get_infer_iterator(dataset, vocab_table, batch_size, src_reverse, eos,
src_max_len=src_max_len, eou=hparams.eou,
dialogue_max_len=hparams.dialogue_max_len)
get_infer_iterator = curry_get_infer_iterator
def curry_get_iterator(src_dataset,
tgt_dataset,
vocab_table,
batch_size,
sos,
eos,
src_reverse,
random_seed,
num_buckets,
src_max_len=None,
tgt_max_len=None,
num_threads=4,
output_buffer_size=None,
skip_count=None):
return end2end_iterator_utils.get_iterator(src_dataset, tgt_dataset, vocab_table, batch_size, sos, eos,
eou=hparams.eou, src_reverse=src_reverse, random_seed=random_seed,
num_dialogue_buckets=num_buckets, src_max_len=src_max_len,
tgt_max_len=tgt_max_len, num_threads=num_threads,
output_buffer_size=output_buffer_size, skip_count=skip_count)
get_iterator = curry_get_iterator
else:
raise ValueError("Unkown architecture", hparams.architecture)
# Create three models which share parameters through the use of checkpoints
train_model = create_train_model(model_creator, get_iterator, hparams, scope)
eval_model = create_eval_model(model_creator, get_iterator, hparams, scope)
infer_model = inference.create_infer_model(model_creator, get_infer_iterator, hparams, scope)
# ToDo: adapt for architectures
# Preload the data to use for sample decoding
dev_src_file = "%s.%s" % (hparams.dev_prefix, hparams.src)
dev_tgt_file = "%s.%s" % (hparams.dev_prefix, hparams.tgt)
sample_src_data = inference.load_data(dev_src_file)
sample_tgt_data = inference.load_data(dev_tgt_file)
summary_name = "train_log"
model_dir = hparams.out_dir
# Log and output files
log_file = os.path.join(out_dir, "log_%d" % time.time())
log_f = tf.gfile.GFile(log_file, mode="a")
utils.print_out("# log_file=%s" % log_file, log_f)
avg_step_time = 0.0
# Create the configurations for the sessions
config_proto = utils.get_config_proto(log_device_placement=log_device_placement)
# Create three sessions, one for each model
train_sess = tf.Session(target=target_session, config=config_proto, graph=train_model.graph)
eval_sess = tf.Session(target=target_session, config=config_proto, graph=eval_model.graph)
infer_sess = tf.Session(target=target_session, config=config_proto, graph=infer_model.graph)
# Load the train model from checkpoint or create a new one
with train_model.graph.as_default():
loaded_train_model, global_step = model_helper.create_or_load_model(train_model.model, model_dir,
train_sess, name="train")
# Summary writer
summary_writer = tf.summary.FileWriter(
os.path.join(out_dir, summary_name), train_model.graph)
# First evaluation
run_full_eval(
model_dir, infer_model, infer_sess,
eval_model, eval_sess, hparams,
summary_writer, sample_src_data,
sample_tgt_data)
last_stats_step = global_step
last_eval_step = global_step
last_external_eval_step = global_step
# This is the training loop.
# Initialize the hyperparameters for the loop.
step_time, checkpoint_loss, checkpoint_predict_count = 0.0, 0.0, 0.0
checkpoint_total_count = 0.0
speed, train_ppl = 0.0, 0.0
start_train_time = time.time()
utils.print_out(
"# Start step %d, lr %g, %s" %
(global_step, loaded_train_model.learning_rate.eval(session=train_sess),
time.ctime()),
log_f)
# epoch_step records where we were within an epoch. Used to skip trained on examples
skip_count = hparams.batch_size * hparams.epoch_step
utils.print_out("# Init train iterator, skipping %d elements" % skip_count)
# Initialize the training iterator
train_sess.run(
train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: skip_count})
# Train until we reach num_steps.
while global_step < num_train_steps:
# Run a step
start_step_time = time.time()
try:
step_result = loaded_train_model.train(train_sess)
(_, step_loss, step_predict_count, step_summary, global_step, # The _ is the output of the update op
step_word_count, batch_size) = step_result
hparams.epoch_step += 1
except tf.errors.OutOfRangeError:
# Finished going through the training dataset. Go to next epoch.
hparams.epoch_step = 0
utils.print_out(
"# Finished an epoch, step %d. Perform external evaluation" %
global_step)
# Decode and print a random sentence
run_sample_decode(infer_model, infer_sess, model_dir, hparams, summary_writer,
sample_src_data, sample_tgt_data)
# Perform external evaluation to save checkpoints if this is the best for some metric
dev_scores, test_scores, _ = run_external_evaluation(infer_model, infer_sess, model_dir, hparams,
summary_writer, save_on_best_dev=True)
# Reinitialize the iterator from the beginning
train_sess.run(train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: 0})
continue
# Write step summary.
summary_writer.add_summary(step_summary, global_step)
# update statistics
step_time += (time.time() - start_step_time)
checkpoint_loss += (step_loss * batch_size)
checkpoint_predict_count += step_predict_count
checkpoint_total_count += float(step_word_count)
# Once in a while, we print statistics.
if global_step - last_stats_step >= steps_per_stats:
last_stats_step = global_step
# Print statistics for the previous epoch.
avg_step_time = step_time / steps_per_stats
train_ppl = utils.safe_exp(checkpoint_loss / checkpoint_predict_count)
speed = checkpoint_total_count / (1000 * step_time)
utils.print_out(
" global step %d lr %g "
"step-time %.2fs wps %.2fK ppl %.2f %s" %
(global_step,
loaded_train_model.learning_rate.eval(session=train_sess),
avg_step_time, speed, train_ppl, _get_best_results(hparams)),
log_f)
if math.isnan(train_ppl):
# The model has screwed up
break
# Reset timer and loss.
step_time, checkpoint_loss, checkpoint_predict_count = 0.0, 0.0, 0.0
checkpoint_total_count = 0.0
if global_step - last_eval_step >= steps_per_eval:
# Perform evaluation. Start by reassigning the last_eval_step variable to the current step
last_eval_step = global_step
# Print the progress and add summary
utils.print_out("# Save eval, global step %d" % global_step)
utils.add_summary(summary_writer, global_step, "train_ppl", train_ppl)
# Save checkpoint
loaded_train_model.saver.save(train_sess, os.path.join(out_dir, "chatbot.ckpt"), global_step=global_step)
# Decode and print a random sample
run_sample_decode(infer_model, infer_sess, model_dir, hparams, summary_writer,
sample_src_data, sample_tgt_data)
# Run internal evaluation, and update the ppl variables. The data iterator is instantieted in the method.
dev_ppl, test_ppl = run_internal_eval(eval_model, eval_sess, model_dir, hparams, summary_writer)
if global_step - last_external_eval_step >= steps_per_external_eval:
# Run the external evaluation
last_external_eval_step = global_step
# Save checkpoint
loaded_train_model.saver.save(train_sess, os.path.join(out_dir, "chatbot.ckpt"), global_step=global_step)
# Decode and print a random sample
run_sample_decode(infer_model, infer_sess, model_dir, hparams, summary_writer,
sample_src_data, sample_tgt_data)
# Run external evaluation, updating metric scores in the meanwhile. The unneeded output is the global step.
dev_scores, test_scores, _ = run_external_evaluation(infer_model, infer_sess, model_dir, hparams,
summary_writer, save_on_best_dev=True)
# Done training. Save the model
loaded_train_model.saver.save(
train_sess,
os.path.join(out_dir, "chatbot.ckpt"),
global_step=global_step)
result_summary, _, dev_scores, test_scores, dev_ppl, test_ppl = run_full_eval(
model_dir, infer_model, infer_sess,
eval_model, eval_sess, hparams,
summary_writer, sample_src_data,
sample_tgt_data)
utils.print_out(
"# Final, step %d lr %g "
"step-time %.2f wps %.2fK ppl %.2f, %s, %s" %
(global_step, loaded_train_model.learning_rate.eval(session=train_sess),
avg_step_time, speed, train_ppl, result_summary, time.ctime()),
log_f)
utils.print_time("# Done training!", start_train_time)
utils.print_out("# Start evaluating saved best models.")
for metric in hparams.metrics:
best_model_dir = getattr(hparams, "best_" + metric + "_dir")
result_summary, best_global_step, _, _, _, _ = run_full_eval(
best_model_dir, infer_model, infer_sess, eval_model, eval_sess, hparams,
summary_writer, sample_src_data, sample_tgt_data)
utils.print_out("# Best %s, step %d "
"step-time %.2f wps %.2fK, %s, %s" %
(metric, best_global_step, avg_step_time, speed,
result_summary, time.ctime()), log_f)
summary_writer.close()
return (dev_scores, test_scores, dev_ppl, test_ppl, global_step)
def train(hparams, scope=None, target_session=""):
"""Train a translation model."""
log_device_placement = hparams.log_device_placement
out_dir = hparams.out_dir
num_train_steps = hparams.num_train_steps
steps_per_stats = hparams.steps_per_stats
steps_per_external_eval = hparams.steps_per_external_eval
steps_per_eval = 10 * steps_per_stats
if not steps_per_external_eval:
steps_per_external_eval = 5 * steps_per_eval
if not hparams.attention:
model_creator = nmt_model.Model
elif hparams.attention_architecture == "standard":
model_creator = attention_model.AttentionModel
elif hparams.attention_architecture in ["gnmt", "gnmt_v2"]:
model_creator = gnmt_model.GNMTModel
else:
raise ValueError("Unknown model architecture")
train_model = model_helper.create_train_model(model_creator, hparams,
scope)
eval_model = model_helper.create_eval_model(model_creator, hparams, scope)
infer_model = model_helper.create_infer_model(model_creator, hparams,
scope)
# Preload data for sample decoding.
dev_src_file = "%s.%s" % (hparams.dev_prefix, hparams.src)
dev_tgt_file = "%s.%s" % (hparams.dev_prefix, hparams.tgt)
sample_src_data = inference.load_data(dev_src_file)
sample_tgt_data = inference.load_data(dev_tgt_file)
summary_name = "train_log"
model_dir = hparams.out_dir
# Log and output files
log_file = os.path.join(out_dir, "log_%d" % time.time())
log_f = tf.gfile.GFile(log_file, mode="w")
utils.print_out("# log_file=%s" % log_file, log_f)
avg_step_time = 0.0
# TensorFlow model
config_proto = utils.get_config_proto(
log_device_placement=log_device_placement)
train_sess = tf.Session(target=target_session,
config=config_proto,
graph=train_model.graph)
eval_sess = tf.Session(target=target_session,
config=config_proto,
graph=eval_model.graph)
infer_sess = tf.Session(target=target_session,
config=config_proto,
graph=infer_model.graph)
with train_model.graph.as_default():
loaded_train_model, global_step = model_helper.create_or_load_model(
train_model.model, model_dir, train_sess, "train")
# Summary writer
summary_writer = tf.summary.FileWriter(os.path.join(out_dir, summary_name),
train_model.graph)
# First evaluation
run_full_eval(model_dir, infer_model, infer_sess, eval_model, eval_sess,
hparams, summary_writer, sample_src_data, sample_tgt_data)
last_stats_step = global_step
last_eval_step = global_step
last_external_eval_step = global_step
# This is the training loop.
step_time, checkpoint_loss, checkpoint_predict_count = 0.0, 0.0, 0.0
checkpoint_total_count = 0.0
speed, train_ppl = 0.0, 0.0
start_train_time = time.time()
utils.print_out(
"# Start step %d, lr %g, %s" %
(global_step, loaded_train_model.learning_rate.eval(
session=train_sess), time.ctime()), log_f)
# Initialize all of the iterators
skip_count = hparams.batch_size * hparams.epoch_step
utils.print_out("# Init train iterator, skipping %d elements" % skip_count)
if hparams.curriculum == 'none':
train_sess.run(
train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: skip_count})
else:
if hparams.curriculum == 'predictive_gain':
exp3s = Exp3S(hparams.num_curriculum_buckets, 0.001, 0, 0.05)
elif hparams.curriculum == 'look_back_and_forward':
curriculum_point = 0
handle = train_model.iterator.handle
for i in range(hparams.num_curriculum_buckets):
train_sess.run(
train_model.iterator.initializer[i].initializer,
feed_dict={train_model.skip_count_placeholder: skip_count})
iterator_handles = [
train_sess.run(
train_model.iterator.initializer[i].string_handle(),
feed_dict={train_model.skip_count_placeholder: skip_count})
for i in range(hparams.num_curriculum_buckets)
]
utils.print_out("Starting training")
while global_step < num_train_steps:
### Run a step ###
start_time = time.time()
try:
if hparams.curriculum != 'none':
if hparams.curriculum == 'predictive_gain':
lesson = exp3s.draw_task()
elif hparams.curriculum == 'look_back_and_forward':
if curriculum_point == hparams.num_curriculum_buckets:
lesson = np.random.randint(
low=0, high=hparams.num_curriculum_buckets)
else:
lesson = curriculum_point if np.random.random_sample(
) < 0.8 else np.random.randint(
low=0, high=hparams.num_curriculum_buckets)
step_result = loaded_train_model.train(
hparams,
train_sess,
handle=handle,
iterator_handle=iterator_handles[lesson],
use_fed_source_placeholder=loaded_train_model.
use_fed_source,
fed_source_placeholder=loaded_train_model.fed_source)
(_, step_loss, step_predict_count, step_summary, global_step,
step_word_count, batch_size, source) = step_result
if hparams.curriculum == 'predictive_gain':
new_loss = train_sess.run(
[loaded_train_model.train_loss],
feed_dict={
handle: iterator_handles[lesson],
loaded_train_model.use_fed_source: True,
loaded_train_model.fed_source: source
})
# new_loss = loaded_train_model.train_loss.eval(
# session=train_sess,
# feed_dict={
# handle: iterator_handles[lesson],
# loaded_train_model.use_fed_source: True,
# loaded_train_model.fed_source: source
# })
# utils.print_out("lesson: %s, step loss: %s, new_loss: %s" % (lesson, step_loss, new_loss))
# utils.print_out("exp3s dist: %s" % (exp3s.pi, ))
curriculum_point_a = lesson * (
hparams.src_max_len //
hparams.num_curriculum_buckets) + 1
curriculum_point_b = (
lesson + 1) * (hparams.src_max_len //
hparams.num_curriculum_buckets) + 1
v = step_loss - new_loss
exp3s.update_w(
v,
float(curriculum_point_a + curriculum_point_b) / 2.0)
elif hparams.curriculum == 'look_back_and_forward':
utils.print_out("step loss: %s, lesson: %s" %
(step_loss, lesson))
curriculum_point_a = curriculum_point * (
hparams.src_max_len //
hparams.num_curriculum_buckets) + 1
curriculum_point_b = (curriculum_point + 1) * (
hparams.src_max_len //
hparams.num_curriculum_buckets) + 1
if step_loss < (hparams.curriculum_progress_loss *
(float(curriculum_point_a +
curriculum_point_b) / 2.0)):
curriculum_point += 1
else:
step_result = loaded_train_model.train(hparams, train_sess)
(_, step_loss, step_predict_count, step_summary, global_step,
step_word_count, batch_size) = step_result
hparams.epoch_step += 1
except tf.errors.OutOfRangeError:
# Finished going through the training dataset. Go to next epoch.
hparams.epoch_step = 0
# utils.print_out(
# "# Finished an epoch, step %d. Perform external evaluation" %
# global_step)
# run_sample_decode(infer_model, infer_sess,
# model_dir, hparams, summary_writer, sample_src_data,
# sample_tgt_data)
# dev_scores, test_scores, _ = run_external_eval(
# infer_model, infer_sess, model_dir,
# hparams, summary_writer)
if hparams.curriculum == 'none':
train_sess.run(
train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: 0})
else:
train_sess.run(
train_model.iterator.initializer[lesson].initializer,
feed_dict={train_model.skip_count_placeholder: 0})
continue
# Write step summary.
summary_writer.add_summary(step_summary, global_step)
# update statistics
step_time += (time.time() - start_time)
checkpoint_loss += (step_loss * batch_size)
checkpoint_predict_count += step_predict_count
checkpoint_total_count += float(step_word_count)
# Once in a while, we print statistics.
if global_step - last_stats_step >= steps_per_stats:
if hparams.curriculum == 'predictive_gain':
utils.print_out("lesson: %s, step loss: %s, new_loss: %s" %
(lesson, step_loss, new_loss))
utils.print_out("exp3s dist: %s" % (exp3s.pi, ))
last_stats_step = global_step
# Print statistics for the previous epoch.
avg_step_time = step_time / steps_per_stats
train_ppl = utils.safe_exp(checkpoint_loss /
checkpoint_predict_count)
speed = checkpoint_total_count / (1000 * step_time)
utils.print_out(
" global step %d lr %g "
"step-time %.2fs wps %.2fK ppl %.2f %s" %
(global_step,
loaded_train_model.learning_rate.eval(session=train_sess),
avg_step_time, speed, train_ppl, _get_best_results(hparams)),
log_f)
if math.isnan(train_ppl):
break
# Reset timer and loss.
step_time, checkpoint_loss, checkpoint_predict_count = 0.0, 0.0, 0.0
checkpoint_total_count = 0.0
if global_step - last_eval_step >= steps_per_eval:
last_eval_step = global_step
utils.print_out("# Save eval, global step %d" % global_step)
utils.add_summary(summary_writer, global_step, "train_ppl",
train_ppl)
# Save checkpoint
loaded_train_model.saver.save(train_sess,
os.path.join(out_dir,
"translate.ckpt"),
global_step=global_step)
# Evaluate on dev/test
run_sample_decode(infer_model, infer_sess, model_dir, hparams,
summary_writer, sample_src_data, sample_tgt_data)
dev_ppl, test_ppl = run_internal_eval(eval_model, eval_sess,
model_dir, hparams,
summary_writer)
dev_scores, test_scores, _ = run_external_eval(
infer_model, infer_sess, model_dir, hparams, summary_writer)
# if global_step - last_external_eval_step >= steps_per_external_eval:
# last_external_eval_step = global_step
# # Save checkpoint
# loaded_train_model.saver.save(
# train_sess,
# os.path.join(out_dir, "translate.ckpt"),
# global_step=global_step)
# run_sample_decode(infer_model, infer_sess,
# model_dir, hparams, summary_writer, sample_src_data,
# sample_tgt_data)
# dev_scores, test_scores, _ = run_external_eval(
# infer_model, infer_sess, model_dir,
# hparams, summary_writer)
# Done training
loaded_train_model.saver.save(train_sess,
os.path.join(out_dir, "translate.ckpt"),
global_step=global_step)
result_summary, _, dev_scores, test_scores, dev_ppl, test_ppl = run_full_eval(
model_dir, infer_model, infer_sess, eval_model, eval_sess, hparams,
summary_writer, sample_src_data, sample_tgt_data)
utils.print_out(
"# Final, step %d lr %g "
"step-time %.2f wps %.2fK ppl %.2f, %s, %s" %
(global_step,
loaded_train_model.learning_rate.eval(session=train_sess),
avg_step_time, speed, train_ppl, result_summary, time.ctime()), log_f)
utils.print_time("# Done training!", start_train_time)
utils.print_out("# Start evaluating saved best models.")
for metric in hparams.metrics:
best_model_dir = getattr(hparams, "best_" + metric + "_dir")
result_summary, best_global_step, _, _, _, _ = run_full_eval(
best_model_dir, infer_model, infer_sess, eval_model, eval_sess,
hparams, summary_writer, sample_src_data, sample_tgt_data)
utils.print_out(
"# Best %s, step %d "
"step-time %.2f wps %.2fK, %s, %s" %
(metric, best_global_step, avg_step_time, speed, result_summary,
time.ctime()), log_f)
summary_writer.close()
return (dev_scores, test_scores, dev_ppl, test_ppl, global_step)
def train(hparams, scope=None, target_session=""):
"""Train a translation model."""
log_device_placement = hparams.log_device_placement
out_dir = hparams.out_dir
num_train_steps = hparams.num_train_steps
steps_per_stats = hparams.steps_per_stats
steps_per_external_eval = hparams.steps_per_external_eval
steps_per_eval = 10 * steps_per_stats
if not steps_per_external_eval:
steps_per_external_eval = 5 * steps_per_eval
print(hparams.attention)
if hparams.attention.strip() == "":
model_creator = nmt_model.Model
print("using nmt model " + hparams.attention + "done")
elif hparams.attention_architecture == "standard":
model_creator = attention_model.AttentionModel
print("using attention model " + hparams.attention + "done")
elif hparams.attention_architecture in ["gnmt", "gnmt_v2"]:
model_creator = gnmt_model.GNMTModel
print("using gnmt model")
else:
raise ValueError("Unknown model architecture")
train_model = model_helper.create_train_model(model_creator, hparams,
scope)
eval_model = model_helper.create_eval_model(model_creator, hparams, scope)
infer_model = model_helper.create_infer_model(model_creator, hparams,
scope)
# Preload data for sample decoding.
dev_src_file = hparams.dev_src
dev_tgt_file = hparams.dev_tgt
sample_src_data = inference.load_data(dev_src_file)
sample_tgt_data = inference.load_data(dev_tgt_file)
summary_name = "train_log"
#
# Log and output files
log_file = os.path.join(out_dir, "log_%d" % time.time())
log_f = tf.gfile.GFile(log_file, mode="a")
utils.print_out("# log_file=%s" % log_file, log_f)
avg_step_time = 0.0
model_dir = hparams.out_dir
# TensorFlow model
config_proto = utils.get_config_proto(
log_device_placement=log_device_placement)
train_sess = tf.Session(target=target_session,
config=config_proto,
graph=train_model.graph)
eval_sess = tf.Session(target=target_session,
config=config_proto,
graph=eval_model.graph)
infer_sess = tf.Session(target=target_session,
config=config_proto,
graph=infer_model.graph)
with train_model.graph.as_default():
loaded_train_model, global_step = model_helper.create_model_Alveo(
train_model.model, train_sess, "train")
# Summary writer
summary_writer = tf.summary.FileWriter(os.path.join(out_dir, summary_name),
train_model.graph)
# First evaluation
run_full_eval(model_dir, infer_model, infer_sess, eval_model, eval_sess,
hparams, summary_writer, sample_src_data, sample_tgt_data)
last_stats_step = global_step
last_eval_step = global_step
last_external_eval_step = global_step
# This is the training loop.
stats = init_stats()
speed, train_ppl = 0.0, 0.0
start_train_time = time.time()
utils.print_out(
"# Start step %d, lr %g, %s" %
(global_step, loaded_train_model.learning_rate.eval(
session=train_sess), time.ctime()), log_f)
# Initialize all of the iterators
skip_count = hparams.batch_size * hparams.epoch_step
utils.print_out("# Init train iterator, skipping %d elements" % skip_count)
train_sess.run(train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: skip_count})
while global_step < num_train_steps:
### Run a step ###
start_time = time.time()
try:
step_result = loaded_train_model.train(train_sess)
hparams.epoch_step += 1
except tf.errors.OutOfRangeError:
# Finished going through the training dataset. Go to next epoch.
hparams.epoch_step = 0
utils.print_out(
"# Finished an epoch, step %d. Perform external evaluation" %
global_step)
run_sample_decode(infer_model, infer_sess, model_dir, hparams,
summary_writer, sample_src_data, sample_tgt_data)
dev_scores, test_scores, _ = run_external_eval(
infer_model, infer_sess, model_dir, hparams, summary_writer)
train_sess.run(train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: 0})
continue
# Write step summary and accumulate statistics
global_step = update_stats(stats, summary_writer, start_time,
step_result)
# Once in a while, we print statistics.
if global_step - last_stats_step >= steps_per_stats:
last_stats_step = global_step
is_overflow = check_stats(stats, global_step, steps_per_stats,
hparams, log_f)
if is_overflow:
break
# Reset statistics
stats = init_stats()
if global_step - last_eval_step >= steps_per_eval:
last_eval_step = global_step
utils.print_out("# Save eval, global step %d" % global_step)
utils.add_summary(summary_writer, global_step, "train_ppl",
train_ppl)
# Save checkpoint
loaded_train_model.saver.save(train_sess,
os.path.join(out_dir,
"translate.ckpt"),
global_step=global_step)
# Evaluate on dev/test
run_sample_decode(infer_model, infer_sess, model_dir, hparams,
summary_writer, sample_src_data, sample_tgt_data)
dev_ppl, test_ppl = run_internal_eval(eval_model, eval_sess,
model_dir, hparams,
summary_writer)
if global_step - last_external_eval_step >= steps_per_external_eval:
last_external_eval_step = global_step
# Save checkpoint
loaded_train_model.saver.save(train_sess,
os.path.join(out_dir,
"translate.ckpt"),
global_step=global_step)
run_sample_decode(infer_model, infer_sess, model_dir, hparams,
summary_writer, sample_src_data, sample_tgt_data)
dev_scores, test_scores, _ = run_external_eval(
infer_model, infer_sess, model_dir, hparams, summary_writer)
# Done training
loaded_train_model.saver.save(train_sess,
os.path.join(out_dir, "translate.ckpt"),
global_step=global_step)
result_summary, _, dev_scores, test_scores, dev_ppl, test_ppl = run_full_eval(
model_dir, infer_model, infer_sess, eval_model, eval_sess, hparams,
summary_writer, sample_src_data, sample_tgt_data)
utils.print_out(
"# Final, step %d lr %g "
"step-time %.2f wps %.2fK ppl %.2f, %s, %s" %
(global_step,
loaded_train_model.learning_rate.eval(session=train_sess),
avg_step_time, speed, train_ppl, result_summary, time.ctime()), log_f)
utils.print_time("# Done training!", start_train_time)
summary_writer.close()
utils.print_out("# Start evaluating saved best models.")
for metric in hparams.metrics:
best_model_dir = getattr(hparams, "best_" + metric + "_dir")
summary_writer = tf.summary.FileWriter(
os.path.join(best_model_dir, summary_name), infer_model.graph)
result_summary, best_global_step, _, _, _, _ = run_full_eval(
best_model_dir, infer_model, infer_sess, eval_model, eval_sess,
hparams, summary_writer, sample_src_data, sample_tgt_data)
utils.print_out(
"# Best %s, step %d "
"step-time %.2f wps %.2fK, %s, %s" %
(metric, best_global_step, avg_step_time, speed, result_summary,
time.ctime()), log_f)
summary_writer.close()
return (dev_scores, test_scores, dev_ppl, test_ppl, global_step)