def _sample_decode(self, model, global_step, sess,
iterator_src_placeholder,
iterator_batch_size_placeholder, eval_data,
summary_writer):
"""Pick a sentence and decode."""
decode_id = random.randint(0, len(eval_data) - 1)
log.print_out(" # {}".format(decode_id))
iterator_feed_dict = {
iterator_src_placeholder: [eval_data[decode_id]],
iterator_batch_size_placeholder: 1,
}
sess.run(model.iterator.initializer, feed_dict=iterator_feed_dict)
ncm_outputs, attention_summary = model.decode(sess)
if self.config.beam_width > 0:
# get the top translation.
ncm_outputs = ncm_outputs[0]
translation = ncm_utils.get_translation(ncm_outputs, sent_id=0)
log.print_out(" sources:")
utterances = eval_data[decode_id].split("\t")
sources, target = utterances[:-1], utterances[-1]
for t, src in enumerate(sources):
log.print_out(" @{} {}".format(t + 1, src))
log.print_out(" resp: {}".format(target))
log.print_out(b" generated: " + translation)
# Summary
if attention_summary is not None:
summary_writer.add_summary(attention_summary, global_step)
def decode_and_evaluate(name,
model,
sess,
out_file,
ref_file,
metrics,
beam_width,
num_translations_per_input=1,
decode=True):
"""Decode a test set and compute a score according to the evaluation task."""
# Decode
if decode:
log.print_out(" decoding to output '{}'".format(out_file))
start_time = time.time()
num_sentences = 0
with codecs.getwriter("utf-8")(
tf.gfile.GFile(out_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)
i = 0
while True:
i += 1
try:
if i % 1000 == 0:
log.print_out(" decoding step {}, num sentences {}".format(i, num_sentences))
ncm_outputs, _ = model.decode(sess)
if beam_width == 0:
ncm_outputs = np.expand_dims(ncm_outputs, 0)
batch_size = ncm_outputs.shape[1]
num_sentences += batch_size
for sent_id in range(batch_size):
translations = [get_translation(ncm_outputs[beam_id], sent_id)
for beam_id in range(num_translations_per_input)]
trans_f.write(b"\t".join(translations).decode("utf-8") + "\n")
except tf.errors.OutOfRangeError:
log.print_time(
" Done, num sentences {}, num translations per input {}".format(
num_sentences, num_translations_per_input), start_time)
break
# Evaluation
evaluation_scores = {}
# if ref_file and tf.gfile.Exists(out_file):
# for metric in metrics:
# score = evaluate(ref_file, out_file, metric)
# evaluation_scores[metric] = score
# log.print_out(" %s %s: %.1f" % (metric, name, score))
return evaluation_scores
def test(self):
start_test_time = time.time()
assert self.config.n_responses >= 1
if self.config.beam_width > 0:
assert self.config.n_responses <= self.config.beam_width
else:
assert self.config.n_responses == 1
self._pre_model_creation()
infer_model = vanilla_helper.create_infer_model(self.config)
config_proto = model_helper.get_config_proto(self.config.log_device)
ckpt = tf.train.latest_checkpoint(self.config.get_infer_model_dir())
with tf.Session(graph=infer_model.graph,
config=config_proto) as infer_sess:
loaded_infer_model = model_helper.load_model(
infer_model.model, ckpt, infer_sess, "infer")
log.print_out("# Start decoding")
log.print_out(" beam width: {}".format(self.config.beam_width))
log.print_out(" length penalty: {}".format(
self.config.length_penalty_weight))
log.print_out(" sampling temperature: {}".format(
self.config.sampling_temperature))
log.print_out(" num responses per tests instance: {}".format(
self.config.n_responses))
feed_dict = {
infer_model.src_placeholder:
self._load_data(self.config.test_data),
infer_model.batch_size_placeholder:
self.config.infer_batch_size,
}
infer_sess.run(infer_model.iterator.initializer,
feed_dict=feed_dict)
if self.config.sampling_temperature > 0:
label = "%s_t%.1f" % (fs.file_name(
self.config.test_data), self.config.sampling_temperature)
else:
label = "%s_bw%d_lp%.1f" % (fs.file_name(
self.config.test_data), self.config.beam_width,
self.config.length_penalty_weight)
self._decode_and_evaluate(
loaded_infer_model,
infer_sess,
feed_dict,
label=label,
num_responses_per_input=self.config.n_responses)
log.print_time("# Decoding done", start_test_time)
eval_model = vanilla_helper.create_eval_model(self.config)
with tf.Session(config=model_helper.get_config_proto(
self.config.log_device),
graph=eval_model.graph) as eval_sess:
loaded_eval_model = model_helper.load_model(
eval_model.model, ckpt, eval_sess, "eval")
log.print_out("# Compute Perplexity")
feed_dict = {
eval_model.eval_file_placeholder: self.config.test_data
}
eval_sess.run(eval_model.iterator.initializer, feed_dict=feed_dict)
model_helper.compute_perplexity(loaded_eval_model, eval_sess,
"test")
log.print_time("# Test finished", start_test_time)
def train(self, target_session="", scope=None):
out_dir = self.config.model_dir
model_dir = out_dir
steps_per_stats = self.config.steps_per_stats
steps_per_eval = self.config.steps_per_eval
self._pre_model_creation()
train_model = vanilla_helper.create_train_model(self.config, scope)
eval_model = vanilla_helper.create_eval_model(self.config, scope)
infer_model = vanilla_helper.create_infer_model(self.config, scope)
# Preload data for sample decoding.
eval_data = self._load_data(self.config.dev_data)
self.config.dev_size = math.ceil(
len(eval_data) / self.config.batch_size)
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")
log.print_out("# log_file=%s" % log_file, log_f)
avg_step_time = 0.0
# TensorFlow model
config_proto = model_helper.get_config_proto(self.config.log_device)
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
# self.run_full_eval(
# model_dir, infer_model, infer_sess,
# eval_model, eval_sess, summary_writer, eval_data)
last_stats_step = global_step
last_eval_step = global_step
# last_external_eval_step = global_step
patience = self.config.patience
# This is the training loop.
stats = self.init_stats()
speed, train_ppl = 0.0, 0.0
start_train_time = time.time()
self.config.save()
# Initialize all of the iterators
skip_count = self.config.batch_size * self.config.epoch_step
lr = loaded_train_model.learning_rate.eval(session=train_sess)
log.print_out(
"# Starting step {}/{} (skipping {} elements), epoch {}/{}, lr {:f}, {}"
.format(global_step, self.config.num_train_steps, skip_count,
self.config.epoch, self.config.num_train_epochs, lr,
time.ctime()), log_f)
train_sess.run(
train_model.iterator.initializer,
feed_dict={train_model.skip_count_placeholder: skip_count})
pbar = trange(self.config.num_train_steps, initial=global_step)
pbar.set_postfix(lr=lr,
wps='0K',
ppl='inf',
gN='inf',
best_dev_ppl=self.config.best_dev_ppl)
pbar.set_description("Ep {}/{}".format(self.config.epoch,
self.config.num_train_epochs))
while self.config.epoch < self.config.num_train_epochs and patience > 0:
### Run a step ###
start_time = time.time()
try:
step_result = loaded_train_model.train(train_sess)
self.config.epoch_step += 1
except tf.errors.OutOfRangeError:
# Finished going through the training dataset. Go to next epoch.
sw = Stopwatch()
self.run_sample_decode(infer_model, infer_sess, model_dir,
summary_writer, eval_data)
log.print_out(
"## Done epoch {} in {} steps. step {} @ eval time: {}s".
format(self.config.epoch, self.config.epoch_step,
global_step, sw.elapsed()))
self.config.epoch += 1
self.config.epoch_step = 0
self.config.save()
pbar.set_description("Ep {}/{}".format(
self.config.epoch, self.config.num_train_epochs))
# dev_scores, test_scores, _ = self.run_external_eval(infer_model, infer_sess, model_dir, 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 = self.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:
train_ppl, speed, is_overflow = self.check_stats(
stats, global_step, steps_per_stats, log_f, pbar)
pbar.update(global_step - last_stats_step)
last_stats_step = global_step
if is_overflow:
break
# Reset statistics
stats = self.init_stats()
if global_step - last_eval_step >= steps_per_eval:
last_eval_step = global_step
log.print_out(
"# Save eval, global step {}".format(global_step))
log.add_summary(summary_writer, global_step, "train_ppl",
train_ppl)
# Save checkpoint
loaded_train_model.saver.save(train_sess,
os.path.join(
out_dir, "vanilla.ckpt"),
global_step=global_step)
# Evaluate on dev
self.run_sample_decode(infer_model, infer_sess, model_dir,
summary_writer, eval_data)
dev_ppl = self.run_internal_eval(eval_model, eval_sess,
model_dir, summary_writer)
if dev_ppl < self.config.best_dev_ppl:
self.config.best_dev_ppl = dev_ppl
patience = self.config.patience
log.print_out(
' **** Best model so far @Ep {} @step {} (global {}) dev_ppl {:.3f}'
.format(self.config.epoch, self.config.epoch_step,
global_step, dev_ppl))
elif dev_ppl > self.config.degrade_threshold * self.config.best_dev_ppl:
patience -= 1
log.print_out(
' PPL got worse @Ep {} @step {} (global {}) patience {} '
'dev_ppl {:.3f} best_dev_ppl {:.3f}'.format(
self.config.epoch, self.config.epoch_step,
global_step, patience, dev_ppl,
self.config.best_dev_ppl))
# Save config parameters
self.config.save()
pbar.close()
# Done training
loaded_train_model.saver.save(train_sess,
os.path.join(out_dir, "vanilla.ckpt"),
global_step=global_step)
dev_scores, test_scores, dev_ppl, test_ppl = None, None, None, None
log.print_out(
"# Final, step {} ep {}/{} lr {:f} "
"step-time {:.2f} wps {:.2f}K train_ppl {:.3f} best_dev_ppl {:.3f}, {}"
.format(global_step, self.config.epoch, self.config.epoch_step,
loaded_train_model.learning_rate.eval(session=train_sess),
avg_step_time, speed, train_ppl, self.config.best_dev_ppl,
time.ctime()), log_f)
log.print_time("# Done training!", start_train_time)
if self.config.eval_best_model:
log.print_out("Evaluating the best model begins...")
test_ppl = self.run_infer_best_model(
infer_model, eval_model, infer_sess, eval_sess,
self.config.best_dev_ppl_dir,
fs.file_name(self.config.test_data) + '_final', summary_writer)
log.print_out(
"# test_ppl {:.3f} w. beam_width: {} | length_penalty: {:.1f} | sampling_temperature: {:.1f}"
.format(test_ppl, self.config.beam_width,
self.config.length_penalty_weight,
self.config.sampling_temperature), log_f)
summary_writer.close()
return (dev_scores, test_scores, dev_ppl, test_ppl, global_step)
def __init__(self,
mode,
num_turns,
iterator,
params,
rev_vocab_table=None,
scope=None,
log_trainables=True):
log.print_out("# creating %s graph ..." % mode)
self.dtype = tf.float32
self.mode = mode
self.num_turns = num_turns - 1
self.device_manager = DeviceManager()
self.round_robin = RoundRobin(self.device_manager)
self.num_gpus = min(params.num_gpus,
self.device_manager.num_available_gpus())
log.print_out("# number of gpus %d" % self.num_gpus)
self.iterator = iterator
with tf.variable_scope(scope or 'thred_graph', dtype=self.dtype):
self.init_embeddings(params.vocab_file,
params.vocab_pkl,
scope=scope)
encoder_keep_prob, decoder_keep_prob = self.get_keep_probs(
mode, params)
if mode == tf.contrib.learn.ModeKeys.TRAIN:
context_keep_prob = 1.0 - params.context_dropout_rate
else:
context_keep_prob = 1.0
with tf.variable_scope(scope or "build_network"):
with tf.variable_scope(
"decoder/output_projection") as output_scope:
if params.boost_topic_gen_prob:
self.output_layer = taware_layer.JointDenseLayer(
params.vocab_size,
params.topic_vocab_size,
scope=output_scope,
name="output_projection")
else:
self.output_layer = layers_core.Dense(
params.vocab_size,
use_bias=False,
name="output_projection")
self.batch_size = tf.size(self.iterator.source_sequence_lengths[0])
devices = self.round_robin.assign(2, base=self.num_gpus - 1)
encoder_results, encoder_state = self.__build_encoder(
params, encoder_keep_prob)
context_outputs, context_state = self.__build_context(
params, encoder_results, encoder_state, context_keep_prob,
devices[0])
self.global_step = tf.Variable(0, trainable=False)
self.use_scheduled_sampling = False
if mode == tf.contrib.learn.ModeKeys.TRAIN:
self.sampling_probability = tf.constant(
params.scheduled_sampling_prob)
self.sampling_probability = self._get_sampling_probability(
params, self.global_step, self.sampling_probability)
self.use_scheduled_sampling = params.scheduled_sampling_prob > 0
elif mode == tf.contrib.learn.ModeKeys.EVAL:
self.sampling_probability = tf.constant(0.0)
logits, sample_ids, _ = self.__build_decoder(
params, context_outputs, context_state, decoder_keep_prob,
devices[1])
if mode != tf.contrib.learn.ModeKeys.INFER:
with tf.device(self.device_manager.tail_gpu()):
loss = self.__compute_loss(logits)
else:
loss, losses = None, None
if mode == tf.contrib.learn.ModeKeys.TRAIN:
self.train_loss = loss
self.word_count = sum(
[tf.reduce_sum(self.iterator.source_sequence_lengths[t]) for t in range(self.num_turns)]) + \
tf.reduce_sum(self.iterator.target_sequence_length)
elif mode == tf.contrib.learn.ModeKeys.EVAL:
self.eval_loss = loss
elif mode == tf.contrib.learn.ModeKeys.INFER:
self.sample_words = rev_vocab_table.lookup(
tf.to_int64(sample_ids))
if mode != tf.contrib.learn.ModeKeys.INFER:
## Count the number of predicted words for compute ppl.
self.predict_count = tf.reduce_sum(
self.iterator.target_sequence_length)
trainables = tf.trainable_variables()
if mode == tf.contrib.learn.ModeKeys.TRAIN:
self.learning_rate = tf.constant(params.learning_rate)
# decay
self.learning_rate = self._get_learning_rate_decay(
params, self.global_step, self.learning_rate)
# Optimizer
if params.optimizer.lower() == "sgd":
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
tf.summary.scalar("lr", self.learning_rate)
elif params.optimizer.lower() == "adam":
opt = tf.train.AdamOptimizer(self.learning_rate)
tf.summary.scalar("lr", self.learning_rate)
else:
raise ValueError('Unknown optimizer: ' + params.optimizer)
# Gradients
gradients = tf.gradients(self.train_loss,
trainables,
colocate_gradients_with_ops=True)
clipped_grads, grad_norm = tf.clip_by_global_norm(
gradients, params.max_gradient_norm)
grad_norm_summary = [tf.summary.scalar("grad_norm", grad_norm)]
grad_norm_summary.append(
tf.summary.scalar("clipped_gradient",
tf.global_norm(clipped_grads)))
self.grad_norm = grad_norm
self.update = opt.apply_gradients(zip(clipped_grads,
trainables),
global_step=self.global_step)
# Summary
self.train_summary = tf.summary.merge([
tf.summary.scalar("lr", self.learning_rate),
tf.summary.scalar("train_loss", self.train_loss),
] + grad_norm_summary)
if mode == tf.contrib.learn.ModeKeys.INFER:
self.infer_logits, self.sample_id = logits, sample_ids
self.infer_summary = tf.no_op()
# Saver
self.saver = tf.train.Saver(tf.global_variables(), max_to_keep=2)
# Print trainable variables
if log_trainables:
log.print_out("# Trainable variables")
for trainable in trainables:
log.print_out(" %s, %s, %s" %
(trainable.name, str(trainable.get_shape()),
trainable.op.device))
def _sample_decode(self, model, global_step, sess, src_placeholder,
batch_size_placeholder, eval_data, summary_writer):
"""Pick a sentence and decode."""
decode_ids = np.random.randint(low=0, high=len(eval_data) - 1, size=1)
sample_data = []
for decode_id in decode_ids:
sample_data.append(eval_data[decode_id])
iterator_feed_dict = {
src_placeholder: sample_data,
batch_size_placeholder: len(decode_ids),
}
sess.run(model.iterator.initializer, feed_dict=iterator_feed_dict)
ncm_outputs, infer_summary = model.decode(sess)
for i, decode_id in enumerate(decode_ids):
log.print_out(" # {}".format(decode_id))
output = ncm_outputs[i]
if self.config.beam_width > 0 and self._consider_beam():
# get the top translation.
output = output[0]
translation = ncm_utils.get_translation(output, sent_id=0)
delimited_sample = eval_data[decode_id].split("\t")
utterances, topic = delimited_sample[:-1], delimited_sample[-1]
sources, target = utterances[:-1], utterances[-1]
log.print_out(" sources:")
for t, src in enumerate(sources):
log.print_out(" @{} {}".format(t + 1, src))
log.print_out(" topic: {}".format(topic))
log.print_out(" resp: {}".format(target))
log.print_out(b" generated: " + translation)
# Summary
if infer_summary is not None:
summary_writer.add_summary(infer_summary, global_step)