def train(config, sess):
assert (config.prior_model != None and (tf.train.checkpoint_exists(os.path.abspath(config.prior_model))) or (config.map_decay_c==0.0)), \
"MAP training requires a prior model file: Use command-line option --prior_model"
logging.info('Building model...')
model = StandardModel(config)
if config.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=config.learning_rate)
else:
logging.error('No valid optimizer defined: {}'.format(
config.optimizer))
sys.exit(1)
init = tf.zeros_initializer(dtype=tf.int32)
global_step = tf.get_variable('time', [],
initializer=init,
trainable=False)
if config.summaryFreq:
summary_dir = (config.summary_dir if config.summary_dir is not None
else os.path.abspath(os.path.dirname(config.saveto)))
writer = tf.summary.FileWriter(summary_dir, sess.graph)
else:
writer = None
updater = ModelUpdater(config, model, optimizer, global_step, writer)
saver, progress = init_or_restore_variables(config, sess, train=True)
global_step.load(progress.uidx, sess)
#save model options
config_as_dict = OrderedDict(sorted(vars(config).items()))
json.dump(config_as_dict, open('%s.json' % config.saveto, 'wb'), indent=2)
text_iterator, valid_text_iterator = load_data(config)
_, _, num_to_source, num_to_target = load_dictionaries(config)
total_loss = 0.
n_sents, n_words = 0, 0
last_time = time.time()
logging.info("Initial uidx={}".format(progress.uidx))
for progress.eidx in xrange(progress.eidx, config.max_epochs):
logging.info('Starting epoch {0}'.format(progress.eidx))
for source_sents, target_sents in text_iterator:
print("")
print("")
print("")
print("########## Source Sents ############")
print(source_sents)
print("")
print("")
print("")
print("########## Target Sents ############")
print(target_sents)
if len(source_sents[0][0]) != config.factors:
logging.error(
'Mismatch between number of factors in settings ({0}), and number in training corpus ({1})\n'
.format(config.factors, len(source_sents[0][0])))
sys.exit(1)
x_in, x_mask_in, y_in, y_mask_in = util.prepare_data(
source_sents, target_sents, config.factors, maxlen=None)
if x_in is None:
logging.info(
'Minibatch with zero sample under length {0}'.format(
config.maxlen))
continue
write_summary_for_this_batch = config.summaryFreq and (
(progress.uidx % config.summaryFreq == 0) or
(config.finish_after
and progress.uidx % config.finish_after == 0))
(factors, seqLen, batch_size) = x_in.shape
loss = updater.update(sess, x_in, x_mask_in, y_in, y_mask_in,
write_summary_for_this_batch)
total_loss += loss
n_sents += batch_size
n_words += int(numpy.sum(y_mask_in))
progress.uidx += 1
if config.dispFreq and progress.uidx % config.dispFreq == 0:
duration = time.time() - last_time
disp_time = datetime.now().strftime('[%Y-%m-%d %H:%M:%S]')
logging.info(
'{0} Epoch: {1} Update: {2} Loss/word: {3} Words/sec: {4} Sents/sec: {5}'
.format(disp_time, progress.eidx, progress.uidx,
total_loss / n_words, n_words / duration,
n_sents / duration))
last_time = time.time()
total_loss = 0.
n_sents = 0
n_words = 0
if config.sampleFreq and progress.uidx % config.sampleFreq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :
10], x_mask_in[:, :
10], y_in[:, :
10]
samples = model.sample(sess, x_small, x_mask_small)
assert len(samples) == len(x_small.T) == len(
y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
sample = util.seq2words(ss, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
logging.info('SAMPLE: {}'.format(sample))
if config.beamFreq and progress.uidx % config.beamFreq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :
10], x_mask_in[:, :
10], y_in[:, :
10]
samples = model.beam_search(sess, x_small, x_mask_small,
config.beam_size)
# samples is a list with shape batch x beam x len
assert len(samples) == len(x_small.T) == len(
y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
for i, (sample_seq, cost) in enumerate(ss):
sample = util.seq2words(sample_seq, num_to_target)
msg = 'SAMPLE {}: {} Cost/Len/Avg {}/{}/{}'.format(
i, sample, cost, len(sample), cost / len(sample))
logging.info(msg)
if config.validFreq and progress.uidx % config.validFreq == 0:
costs = validate(config, sess, valid_text_iterator, model)
# validation loss is mean of normalized sentence log probs
valid_loss = sum(costs) / len(costs)
if (len(progress.history_errs) == 0
or valid_loss < min(progress.history_errs)):
progress.history_errs.append(valid_loss)
progress.bad_counter = 0
saver.save(sess, save_path=config.saveto)
progress_path = '{0}.progress.json'.format(config.saveto)
progress.save_to_json(progress_path)
else:
progress.history_errs.append(valid_loss)
progress.bad_counter += 1
if progress.bad_counter > config.patience:
logging.info('Early Stop!')
progress.estop = True
break
if config.valid_script is not None:
score = validate_with_script(sess, model, config,
valid_text_iterator)
need_to_save = (
score is not None
and (len(progress.valid_script_scores) == 0
or score > max(progress.valid_script_scores)))
if score is None:
score = 0.0 # ensure a valid value is written
progress.valid_script_scores.append(score)
if need_to_save:
save_path = config.saveto + ".best-valid-script"
saver.save(sess, save_path=save_path)
progress_path = '{}.progress.json'.format(save_path)
progress.save_to_json(progress_path)
if config.saveFreq and progress.uidx % config.saveFreq == 0:
saver.save(sess,
save_path=config.saveto,
global_step=progress.uidx)
progress_path = '{0}-{1}.progress.json'.format(
config.saveto, progress.uidx)
progress.save_to_json(progress_path)
if config.finish_after and progress.uidx % config.finish_after == 0:
logging.info("Maximum number of updates reached")
saver.save(sess,
save_path=config.saveto,
global_step=progress.uidx)
progress.estop = True
progress_path = '{0}-{1}.progress.json'.format(
config.saveto, progress.uidx)
progress.save_to_json(progress_path)
break
if progress.estop:
break
def train(config, sess):
assert (config.prior_model != None and (tf.train.checkpoint_exists(os.path.abspath(config.prior_model))) or (config.map_decay_c==0.0)), \
"MAP training requires a prior model file: Use command-line option --prior_model"
# Construct the graph, with one model replica per GPU
num_gpus = len(tf_utils.get_available_gpus())
num_replicas = max(1, num_gpus)
if config.loss_function == 'MRT':
assert config.gradient_aggregation_steps == 1
assert config.max_sentences_per_device == 0, "MRT mode does not support sentence-based split"
if config.max_tokens_per_device != 0:
assert (config.samplesN * config.maxlen <= config.max_tokens_per_device), "need to make sure candidates of a sentence could be " \
"feed into the model"
else:
assert num_replicas == 1, "MRT mode does not support sentence-based split"
assert (config.samplesN * config.maxlen <= config.token_batch_size), "need to make sure candidates of a sentence could be " \
"feed into the model"
logging.info('Building model...')
replicas = []
for i in range(num_replicas):
device_type = "GPU" if num_gpus > 0 else "CPU"
device_spec = tf.DeviceSpec(device_type=device_type, device_index=i)
with tf.device(device_spec):
with tf.variable_scope(tf.get_variable_scope(), reuse=(i>0)):
if config.model_type == "transformer":
model = TransformerModel(config)
else:
model = rnn_model.RNNModel(config)
replicas.append(model)
init = tf.zeros_initializer(dtype=tf.int32)
global_step = tf.get_variable('time', [], initializer=init, trainable=False)
if config.learning_schedule == "constant":
schedule = learning_schedule.ConstantSchedule(config.learning_rate)
elif config.learning_schedule == "transformer":
schedule = learning_schedule.TransformerSchedule(
global_step=global_step,
dim=config.state_size,
warmup_steps=config.warmup_steps)
elif config.learning_schedule == "warmup-plateau-decay":
schedule = learning_schedule.WarmupPlateauDecaySchedule(
global_step=global_step,
peak_learning_rate=config.learning_rate,
warmup_steps=config.warmup_steps,
plateau_steps=config.plateau_steps)
else:
logging.error('Learning schedule type is not valid: {}'.format(
config.learning_schedule))
sys.exit(1)
if config.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=schedule.learning_rate,
beta1=config.adam_beta1,
beta2=config.adam_beta2,
epsilon=config.adam_epsilon)
else:
logging.error('No valid optimizer defined: {}'.format(config.optimizer))
sys.exit(1)
if config.summary_freq:
summary_dir = (config.summary_dir if config.summary_dir is not None
else os.path.abspath(os.path.dirname(config.saveto)))
writer = tf.summary.FileWriter(summary_dir, sess.graph)
else:
writer = None
updater = ModelUpdater(config, num_gpus, replicas, optimizer, global_step,
writer)
if config.exponential_smoothing > 0.0:
smoothing = ExponentialSmoothing(config.exponential_smoothing)
saver, progress = model_loader.init_or_restore_variables(
config, sess, train=True)
global_step.load(progress.uidx, sess)
if config.sample_freq:
random_sampler = RandomSampler(
models=[replicas[0]],
configs=[config],
beam_size=1)
if config.beam_freq or config.valid_script is not None:
beam_search_sampler = BeamSearchSampler(
models=[replicas[0]],
configs=[config],
beam_size=config.beam_size)
#save model options
write_config_to_json_file(config, config.saveto)
text_iterator, valid_text_iterator = load_data(config)
_, _, num_to_source, num_to_target = util.load_dictionaries(config)
total_loss = 0.
n_sents, n_words = 0, 0
last_time = time.time()
logging.info("Initial uidx={}".format(progress.uidx))
# set epoch = 1 if print per-token-probability
if config.print_per_token_pro:
config.max_epochs = progress.eidx+1
for progress.eidx in range(progress.eidx, config.max_epochs):
logging.info('Starting epoch {0}'.format(progress.eidx))
for source_sents, target_sents in text_iterator:
if len(source_sents[0][0]) != config.factors:
logging.error('Mismatch between number of factors in settings ({0}), and number in training corpus ({1})\n'.format(config.factors, len(source_sents[0][0])))
sys.exit(1)
x_in, x_mask_in, y_in, y_mask_in = util.prepare_data(
source_sents, target_sents, config.factors, maxlen=None)
if x_in is None:
logging.info('Minibatch with zero sample under length {0}'.format(config.maxlen))
continue
write_summary_for_this_batch = config.summary_freq and ((progress.uidx % config.summary_freq == 0) or (config.finish_after and progress.uidx % config.finish_after == 0))
(factors, seqLen, batch_size) = x_in.shape
output = updater.update(
sess, x_in, x_mask_in, y_in, y_mask_in, num_to_target,
write_summary_for_this_batch)
if config.print_per_token_pro == False:
total_loss += output
else:
# write per-token probability into the file
f = open(config.print_per_token_pro, 'a')
for pro in output:
pro = str(pro) + '\n'
f.write(pro)
f.close()
n_sents += batch_size
n_words += int(numpy.sum(y_mask_in))
progress.uidx += 1
# Update the smoothed version of the model variables.
# To reduce the performance overhead, we only do this once every
# N steps (the smoothing factor is adjusted accordingly).
if config.exponential_smoothing > 0.0 and progress.uidx % smoothing.update_frequency == 0:
sess.run(fetches=smoothing.update_ops)
if config.disp_freq and progress.uidx % config.disp_freq == 0:
duration = time.time() - last_time
disp_time = datetime.now().strftime('[%Y-%m-%d %H:%M:%S]')
logging.info('{0} Epoch: {1} Update: {2} Loss/word: {3} Words/sec: {4} Sents/sec: {5}'.format(disp_time, progress.eidx, progress.uidx, total_loss/n_words, n_words/duration, n_sents/duration))
last_time = time.time()
total_loss = 0.
n_sents = 0
n_words = 0
if config.sample_freq and progress.uidx % config.sample_freq == 0:
x_small = x_in[:, :, :10]
x_mask_small = x_mask_in[:, :10]
y_small = y_in[:, :10]
samples = translate_utils.translate_batch(
sess, random_sampler, x_small, x_mask_small,
config.translation_maxlen, 0.0)
assert len(samples) == len(x_small.T) == len(y_small.T), \
(len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
sample = util.seq2words(ss[0][0], num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
logging.info('SAMPLE: {}'.format(sample))
if config.beam_freq and progress.uidx % config.beam_freq == 0:
x_small = x_in[:, :, :10]
x_mask_small = x_mask_in[:, :10]
y_small = y_in[:,:10]
samples = translate_utils.translate_batch(
sess, beam_search_sampler, x_small, x_mask_small,
config.translation_maxlen, config.normalization_alpha)
assert len(samples) == len(x_small.T) == len(y_small.T), \
(len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
for i, (sample_seq, cost) in enumerate(ss):
sample = util.seq2words(sample_seq, num_to_target)
msg = 'SAMPLE {}: {} Cost/Len/Avg {}/{}/{}'.format(
i, sample, cost, len(sample), cost/len(sample))
logging.info(msg)
if config.valid_freq and progress.uidx % config.valid_freq == 0:
if config.exponential_smoothing > 0.0:
sess.run(fetches=smoothing.swap_ops)
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
sess.run(fetches=smoothing.swap_ops)
else:
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
if (len(progress.history_errs) == 0 or
valid_ce < min(progress.history_errs)):
progress.history_errs.append(valid_ce)
progress.bad_counter = 0
save_non_checkpoint(sess, saver, config.saveto)
progress_path = '{0}.progress.json'.format(config.saveto)
progress.save_to_json(progress_path)
else:
progress.history_errs.append(valid_ce)
progress.bad_counter += 1
if progress.bad_counter > config.patience:
logging.info('Early Stop!')
progress.estop = True
break
if config.valid_script is not None:
if config.exponential_smoothing > 0.0:
sess.run(fetches=smoothing.swap_ops)
score = validate_with_script(sess, beam_search_sampler)
sess.run(fetches=smoothing.swap_ops)
else:
score = validate_with_script(sess, beam_search_sampler)
need_to_save = (score is not None and
(len(progress.valid_script_scores) == 0 or
score > max(progress.valid_script_scores)))
if score is None:
score = 0.0 # ensure a valid value is written
progress.valid_script_scores.append(score)
if need_to_save:
progress.bad_counter = 0
save_path = config.saveto + ".best-valid-script"
save_non_checkpoint(sess, saver, save_path)
write_config_to_json_file(config, save_path)
progress_path = '{}.progress.json'.format(save_path)
progress.save_to_json(progress_path)
if config.save_freq and progress.uidx % config.save_freq == 0:
saver.save(sess, save_path=config.saveto, global_step=progress.uidx)
write_config_to_json_file(config, "%s-%s" % (config.saveto, progress.uidx))
progress_path = '{0}-{1}.progress.json'.format(config.saveto, progress.uidx)
progress.save_to_json(progress_path)
if config.finish_after and progress.uidx % config.finish_after == 0:
logging.info("Maximum number of updates reached")
saver.save(sess, save_path=config.saveto, global_step=progress.uidx)
write_config_to_json_file(config, "%s-%s" % (config.saveto, progress.uidx))
progress.estop=True
progress_path = '{0}-{1}.progress.json'.format(config.saveto, progress.uidx)
progress.save_to_json(progress_path)
break
if progress.estop:
break
def train(config, sess):
####################################################
assert (config.prior_model != None and (tf.train.checkpoint_exists(os.path.abspath(config.prior_model))) or (config.map_decay_c==0.0)), \
"MAP training requires a prior model file: Use command-line option --prior_model"
# Construct the graph, with one model replica per GPU
num_gpus = len(util.get_available_gpus())
num_replicas = max(1, num_gpus)
logging.info('Building model...')
replicas = []
for i in range(num_replicas):
device_type = "GPU" if num_gpus > 0 else "CPU"
device_spec = tf.DeviceSpec(device_type=device_type, device_index=i)
with tf.device(device_spec):
with tf.variable_scope(tf.get_variable_scope(), reuse=(i > 0)):
if config.model_type == "transformer":
model = TransformerModel(config)
else:
model = rnn_model.RNNModel(config)
replicas.append(model)
init = tf.zeros_initializer(dtype=tf.int32)
global_step = tf.get_variable('time', [],
initializer=init,
trainable=False)
if config.learning_schedule == "constant":
schedule = ConstantSchedule(config.learning_rate)
elif config.learning_schedule == "transformer":
schedule = TransformerSchedule(global_step=global_step,
dim=config.state_size,
warmup_steps=config.warmup_steps)
else:
logging.error('Learning schedule type is not valid: {}'.format(
config.learning_schedule))
sys.exit(1)
if config.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(
learning_rate=schedule.learning_rate,
beta1=config.adam_beta1,
beta2=config.adam_beta2,
epsilon=config.adam_epsilon)
else:
logging.error('No valid optimizer defined: {}'.format(
config.optimizer))
sys.exit(1)
if config.summary_freq:
summary_dir = (config.summary_dir if config.summary_dir is not None
else os.path.abspath(os.path.dirname(config.saveto)))
writer = tf.summary.FileWriter(summary_dir, sess.graph)
else:
writer = None
updater = ModelUpdater(config, num_gpus, replicas, optimizer, global_step,
writer)
saver, progress = model_loader.init_or_restore_variables(config,
sess,
train=True)
############################################################
#add: pretrain
if config.pretrain:
logging.info("Start pre-training")
#预训练网络参数
pre_batch_size = 1000
epochs = 20
pre_learning_rate = 0.001
pre_optimizer = tf.train.GradientDescentOptimizer(
pre_learning_rate).minimize(replicas[0].loss_pre_train)
#加载预训练数据及相关字典
gvocab, gvectors = util.pre_load_data(config.pretrain_vocab,
config.pretrain_vectors)
pre_vocab_list = list(gvocab.keys())
#过采样
pre_train_list = []
with open('/media/ntfs-3/EXP/MULTI/mix/zh-en/data3/glove/vocab.txt',
'r',
encoding='utf-8') as f:
for line in f:
k, v = line.strip().split()
pre_train_list.extend([k] * int(v))
utf8_dict = json.load(
open(config.source_dicts[0], 'r', encoding='utf-8'))
embedding_list = []
#开始训练
for i in range(epochs):
logging.info("epoch:{}".format(i))
if i == epochs - 1:
source_x, source_y, _vocab = util.get_data(pre_vocab_list,
pre_batch_size,
gvocab,
gvectors,
utf8_dict,
shuffle=False)
else:
source_x, source_y, _vocab = util.get_data(pre_train_list,
pre_batch_size,
gvocab,
gvectors,
utf8_dict,
shuffle=True)
for idx, [s_x, s_y] in enumerate(zip(source_x, source_y)):
assert len(s_x) == len(s_y), "{}, {}".format(
len(s_x), len(s_y))
sx, sy = util.pre_prepare_data(s_x, s_y)
feed_dict = {}
feed_dict[replicas[0].pre_inputs.x] = sx
feed_dict[replicas[0].pre_inputs.y] = sy
_, loss, embedding = sess.run([
pre_optimizer, replicas[0].loss_pre_train,
replicas[0].pre_embedding
],
feed_dict=feed_dict)
if idx % 100 == 0:
logging.info("loss:{}".format(loss))
if i == epochs - 1:
embedding_list.append(embedding)
assert _vocab == pre_vocab_list
emb = embedding_list[0]
for e in embedding_list[1:]:
emb = numpy.concatenate((emb, e))
numpy.save("pre_emb/pre_emb.npy", emb)
with open("pre_emb/vocab", "w", encoding="utf-8") as f:
f.write("\n".join(pre_vocab_list))
#tsne可视化
tsne = util.get_tsne(emb, "pre_emb/tsne.npy")
gtsne = numpy.load(config.pretrain_tsne)
#util.plot_tsne(_vocab, tsne, gvocab, gtsne, top=20)
#exit(0)
##################################################################################
global_step.load(progress.uidx, sess)
# Use an InferenceModelSet to abstract over model types for sampling and
# beam search. Multi-GPU sampling and beam search are not currently
# supported, so we just use the first replica.
model_set = inference.InferenceModelSet([replicas[0]], [config])
#save model options
write_config_to_json_file(config, config.saveto)
text_iterator, valid_text_iterator = load_data(config)
_, _, num_to_source, num_to_target = util.load_dictionaries(config)
total_loss = 0.
n_sents, n_words = 0, 0
last_time = time.time()
logging.info("Initial uidx={}".format(progress.uidx))
for progress.eidx in range(progress.eidx, config.max_epochs):
logging.info('Starting epoch {0}'.format(progress.eidx))
for pre_source_sents, source_sents, target_sents in text_iterator:
#if len(source_sents[0][0]) != config.factors:
#logging.error('Mismatch between number of factors in settings ({0}), and number in training corpus ({1})\n'.format(config.factors, len(source_sents[0][0])))
#sys.exit(1)
px_in, x_in, x_mask_in, y_in, y_mask_in = util.prepare_data(
source_sents,
target_sents,
config.factors,
pre_source_sents,
maxlen=None)
if x_in is None:
logging.info(
'Minibatch with zero sample under length {0}'.format(
config.maxlen))
continue
write_summary_for_this_batch = config.summary_freq and (
(progress.uidx % config.summary_freq == 0) or
(config.finish_after
and progress.uidx % config.finish_after == 0))
(factors, seqLen, uLen, batch_size) = x_in.shape
loss = updater.update(sess, px_in, x_in, x_mask_in, y_in,
y_mask_in, write_summary_for_this_batch)
total_loss += loss
n_sents += batch_size
n_words += int(numpy.sum(y_mask_in))
progress.uidx += 1
if config.disp_freq and progress.uidx % config.disp_freq == 0:
duration = time.time() - last_time
disp_time = datetime.now().strftime('[%Y-%m-%d %H:%M:%S]')
logging.info(
'{0} Epoch: {1} Update: {2} Loss/word: {3} Words/sec: {4} Sents/sec: {5}'
.format(disp_time, progress.eidx, progress.uidx,
total_loss / n_words, n_words / duration,
n_sents / duration))
last_time = time.time()
total_loss = 0.
n_sents = 0
n_words = 0
if config.sample_freq and progress.uidx % config.sample_freq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :, :
10], x_mask_in[:, :, :
10], y_in[:, :
10]
samples = model_set.sample(sess, x_small, x_mask_small)
assert len(samples) == len(x_small.T) == len(
y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
#source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
sample = util.seq2words(ss, num_to_target)
#logging.info('SOURCE: {}'.format(source))
#logging.info('SOURCE: {}'.format(xx))
logging.info('TARGET: {}'.format(target))
logging.info('SAMPLE: {}'.format(sample))
if config.beam_freq and progress.uidx % config.beam_freq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :, :
10], x_mask_in[:, :, :
10], y_in[:, :
10]
samples = model_set.beam_search(
sess,
x_small,
x_mask_small,
config.beam_size,
normalization_alpha=config.normalization_alpha)
# samples is a list with shape batch x beam x len
assert len(samples) == len(x_small.T) == len(
y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
#source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
#logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
for i, (sample_seq, cost) in enumerate(ss):
sample = util.seq2words(sample_seq, num_to_target)
msg = 'SAMPLE {}: {} Cost/Len/Avg {}/{}/{}'.format(
i, sample, cost, len(sample), cost / len(sample))
logging.info(msg)
if config.valid_freq and progress.uidx % config.valid_freq == 0:
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
if (len(progress.history_errs) == 0
or valid_ce < min(progress.history_errs)):
progress.history_errs.append(valid_ce)
progress.bad_counter = 0
save_non_checkpoint(sess, saver, config.saveto)
progress_path = '{0}.progress.json'.format(config.saveto)
progress.save_to_json(progress_path)
else:
progress.history_errs.append(valid_ce)
progress.bad_counter += 1
if progress.bad_counter > config.patience:
logging.info('Early Stop!')
progress.estop = True
break
if config.valid_script is not None:
score = validate_with_script(sess, replicas[0], config)
need_to_save = (
score is not None
and (len(progress.valid_script_scores) == 0
or score > max(progress.valid_script_scores)))
if score is None:
score = 0.0 # ensure a valid value is written
progress.valid_script_scores.append(score)
if need_to_save:
progress.bad_counter = 0
save_path = config.saveto + ".best-valid-script"
save_non_checkpoint(sess, saver, save_path)
write_config_to_json_file(config, save_path)
progress_path = '{}.progress.json'.format(save_path)
progress.save_to_json(progress_path)
if config.save_freq and progress.uidx % config.save_freq == 0:
saver.save(sess,
save_path=config.saveto,
global_step=progress.uidx)
write_config_to_json_file(
config, "%s-%s" % (config.saveto, progress.uidx))
progress_path = '{0}-{1}.progress.json'.format(
config.saveto, progress.uidx)
progress.save_to_json(progress_path)
if config.finish_after and progress.uidx % config.finish_after == 0:
logging.info("Maximum number of updates reached")
saver.save(sess,
save_path=config.saveto,
global_step=progress.uidx)
write_config_to_json_file(
config, "%s-%s" % (config.saveto, progress.uidx))
progress.estop = True
progress_path = '{0}-{1}.progress.json'.format(
config.saveto, progress.uidx)
progress.save_to_json(progress_path)
break
if progress.estop:
break
def train(config, sess):
assert (config.prior_model != None and (tf.train.checkpoint_exists(os.path.abspath(config.prior_model))) or (config.map_decay_c==0.0)), \
"MAP training requires a prior model file: Use command-line option --prior_model"
# Construct the graph, with one model replica per GPU
num_gpus = len(util.get_available_gpus())
num_replicas = max(1, num_gpus)
logging.info('Building model...')
replicas = []
for i in range(num_replicas):
device_type = "GPU" if num_gpus > 0 else "CPU"
device_spec = tf.DeviceSpec(device_type=device_type, device_index=i)
with tf.device(device_spec):
with tf.variable_scope(tf.get_variable_scope(), reuse=(i > 0)):
if config.model_type == "transformer":
model = TransformerModel(config)
else:
model = rnn_model.RNNModel(config)
replicas.append(model)
init = tf.zeros_initializer(dtype=tf.int32)
global_step = tf.get_variable('time', [],
initializer=init,
trainable=False)
if config.learning_schedule == "constant":
schedule = ConstantSchedule(config.learning_rate)
elif config.learning_schedule == "transformer":
schedule = TransformerSchedule(global_step=global_step,
dim=config.state_size,
warmup_steps=config.warmup_steps)
else:
logging.error('Learning schedule type is not valid: {}'.format(
config.learning_schedule))
sys.exit(1)
if config.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(
learning_rate=schedule.learning_rate,
beta1=config.adam_beta1,
beta2=config.adam_beta2,
epsilon=config.adam_epsilon)
else:
logging.error('No valid optimizer defined: {}'.format(
config.optimizer))
sys.exit(1)
if config.summary_freq:
summary_dir = (config.summary_dir if config.summary_dir is not None
else os.path.abspath(os.path.dirname(config.saveto)))
writer = tf.summary.FileWriter(summary_dir, sess.graph)
else:
writer = None
updater = ModelUpdater(config, num_gpus, replicas, optimizer, global_step,
writer)
saver, progress = model_loader.init_or_restore_variables(config,
sess,
train=True)
global_step.load(progress.uidx, sess)
# Use an InferenceModelSet to abstract over model types for sampling and
# beam search. Multi-GPU sampling and beam search are not currently
# supported, so we just use the first replica.
model_set = inference.InferenceModelSet([replicas[0]], [config])
#save model options
write_config_to_json_file(config, config.saveto)
text_iterator, valid_text_iterator = load_data(config)
_, _, num_to_source, num_to_target = util.load_dictionaries(config)
total_loss = 0.
n_sents, n_words = 0, 0
last_time = time.time()
logging.info("Initial uidx={}".format(progress.uidx))
for progress.eidx in range(progress.eidx, config.max_epochs):
logging.info('Starting epoch {0}'.format(progress.eidx))
for source_sents, target_sents in text_iterator:
if len(source_sents[0][0]) != config.factors:
logging.error(
'Mismatch between number of factors in settings ({0}), and number in training corpus ({1})\n'
.format(config.factors, len(source_sents[0][0])))
sys.exit(1)
x_in, x_mask_in, y_in, y_mask_in = util.prepare_data(
source_sents, target_sents, config.factors, maxlen=None)
if x_in is None:
logging.info(
'Minibatch with zero sample under length {0}'.format(
config.maxlen))
continue
write_summary_for_this_batch = config.summary_freq and (
(progress.uidx % config.summary_freq == 0) or
(config.finish_after
and progress.uidx % config.finish_after == 0))
(factors, seqLen, batch_size) = x_in.shape
loss = updater.update(sess, x_in, x_mask_in, y_in, y_mask_in,
write_summary_for_this_batch)
total_loss += loss
n_sents += batch_size
n_words += int(numpy.sum(y_mask_in))
progress.uidx += 1
if config.disp_freq and progress.uidx % config.disp_freq == 0:
duration = time.time() - last_time
disp_time = datetime.now().strftime('[%Y-%m-%d %H:%M:%S]')
logging.info(
'{0} Epoch: {1} Update: {2} Loss/word: {3} Words/sec: {4} Sents/sec: {5}'
.format(disp_time, progress.eidx, progress.uidx,
total_loss / n_words, n_words / duration,
n_sents / duration))
last_time = time.time()
total_loss = 0.
n_sents = 0
n_words = 0
if config.sample_freq and progress.uidx % config.sample_freq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :
10], x_mask_in[:, :
10], y_in[:, :
10]
samples = model_set.sample(sess, x_small, x_mask_small)
assert len(samples) == len(x_small.T) == len(
y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
sample = util.seq2words(ss, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
logging.info('SAMPLE: {}'.format(sample))
if config.beam_freq and progress.uidx % config.beam_freq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :
10], x_mask_in[:, :
10], y_in[:, :
10]
samples = model_set.beam_search(
sess,
x_small,
x_mask_small,
config.beam_size,
normalization_alpha=config.normalization_alpha)
# samples is a list with shape batch x beam x len
assert len(samples) == len(x_small.T) == len(
y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
for i, (sample_seq, cost) in enumerate(ss):
sample = util.seq2words(sample_seq, num_to_target)
msg = 'SAMPLE {}: {} Cost/Len/Avg {}/{}/{}'.format(
i, sample, cost, len(sample), cost / len(sample))
logging.info(msg)
if config.valid_freq and progress.uidx % config.valid_freq == 0:
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
if (len(progress.history_errs) == 0
or valid_ce < min(progress.history_errs)):
progress.history_errs.append(valid_ce)
progress.bad_counter = 0
save_non_checkpoint(sess, saver, config.saveto)
progress_path = '{0}.progress.json'.format(config.saveto)
progress.save_to_json(progress_path)
else:
progress.history_errs.append(valid_ce)
progress.bad_counter += 1
if progress.bad_counter > config.patience:
logging.info('Early Stop!')
progress.estop = True
break
if config.valid_script is not None:
score = validate_with_script(sess, replicas[0], config)
need_to_save = (
score is not None
and (len(progress.valid_script_scores) == 0
or score > max(progress.valid_script_scores)))
if score is None:
score = 0.0 # ensure a valid value is written
progress.valid_script_scores.append(score)
if need_to_save:
save_path = config.saveto + ".best-valid-script"
save_non_checkpoint(sess, saver, save_path)
write_config_to_json_file(config, save_path)
progress_path = '{}.progress.json'.format(save_path)
progress.save_to_json(progress_path)
if config.save_freq and progress.uidx % config.save_freq == 0:
saver.save(sess,
save_path=config.saveto,
global_step=progress.uidx)
write_config_to_json_file(
config, "%s-%s" % (config.saveto, progress.uidx))
progress_path = '{0}-{1}.progress.json'.format(
config.saveto, progress.uidx)
progress.save_to_json(progress_path)
if config.finish_after and progress.uidx % config.finish_after == 0:
logging.info("Maximum number of updates reached")
saver.save(sess,
save_path=config.saveto,
global_step=progress.uidx)
write_config_to_json_file(
config, "%s-%s" % (config.saveto, progress.uidx))
progress.estop = True
progress_path = '{0}-{1}.progress.json'.format(
config.saveto, progress.uidx)
progress.save_to_json(progress_path)
break
if progress.estop:
break
def train(config, sess):
assert (config.prior_model != None and (tf.train.checkpoint_exists(os.path.abspath(config.prior_model))) or (config.map_decay_c==0.0)), \
"MAP training requires a prior model file: Use command-line option --prior_model"
# Construct the graph, with one model replica per GPU
num_gpus = len(util.get_available_gpus())
num_replicas = max(1, num_gpus)
logging.info('Building model...')
replicas = []
for i in range(num_replicas):
device_type = "GPU" if num_gpus > 0 else "CPU"
device_spec = tf.DeviceSpec(device_type=device_type, device_index=i)
with tf.device(device_spec):
with tf.variable_scope(tf.get_variable_scope(), reuse=(i>0)):
if config.model_type == "transformer":
model = TransformerModel(config)
else:
model = rnn_model.RNNModel(config)
replicas.append(model)
init = tf.zeros_initializer(dtype=tf.int32)
global_step = tf.get_variable('time', [], initializer=init, trainable=False)
if config.learning_schedule == "constant":
schedule = ConstantSchedule(config.learning_rate)
elif config.learning_schedule == "transformer":
schedule = TransformerSchedule(global_step=global_step,
dim=config.state_size,
warmup_steps=config.warmup_steps)
else:
logging.error('Learning schedule type is not valid: {}'.format(
config.learning_schedule))
sys.exit(1)
if config.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate=schedule.learning_rate,
beta1=config.adam_beta1,
beta2=config.adam_beta2,
epsilon=config.adam_epsilon)
else:
logging.error('No valid optimizer defined: {}'.format(config.optimizer))
sys.exit(1)
if config.summary_freq:
summary_dir = (config.summary_dir if config.summary_dir is not None
else os.path.abspath(os.path.dirname(config.saveto)))
writer = tf.summary.FileWriter(summary_dir, sess.graph)
else:
writer = None
updater = ModelUpdater(config, num_gpus, replicas, optimizer, global_step,
writer)
saver, progress = model_loader.init_or_restore_variables(
config, sess, train=True)
global_step.load(progress.uidx, sess)
# Use an InferenceModelSet to abstract over model types for sampling and
# beam search. Multi-GPU sampling and beam search are not currently
# supported, so we just use the first replica.
model_set = inference.InferenceModelSet([replicas[0]], [config])
#save model options
write_config_to_json_file(config, config.saveto)
text_iterator, valid_text_iterator = load_data(config)
_, _, num_to_source, num_to_target = util.load_dictionaries(config)
total_loss = 0.
n_sents, n_words = 0, 0
last_time = time.time()
logging.info("Initial uidx={}".format(progress.uidx))
for progress.eidx in range(progress.eidx, config.max_epochs):
logging.info('Starting epoch {0}'.format(progress.eidx))
for source_sents, target_sents in text_iterator:
if len(source_sents[0][0]) != config.factors:
logging.error('Mismatch between number of factors in settings ({0}), and number in training corpus ({1})\n'.format(config.factors, len(source_sents[0][0])))
sys.exit(1)
x_in, x_mask_in, y_in, y_mask_in = util.prepare_data(
source_sents, target_sents, config.factors, maxlen=None)
if x_in is None:
logging.info('Minibatch with zero sample under length {0}'.format(config.maxlen))
continue
write_summary_for_this_batch = config.summary_freq and ((progress.uidx % config.summary_freq == 0) or (config.finish_after and progress.uidx % config.finish_after == 0))
(factors, seqLen, batch_size) = x_in.shape
loss = updater.update(sess, x_in, x_mask_in, y_in, y_mask_in,
write_summary_for_this_batch)
total_loss += loss
n_sents += batch_size
n_words += int(numpy.sum(y_mask_in))
progress.uidx += 1
if config.disp_freq and progress.uidx % config.disp_freq == 0:
duration = time.time() - last_time
disp_time = datetime.now().strftime('[%Y-%m-%d %H:%M:%S]')
logging.info('{0} Epoch: {1} Update: {2} Loss/word: {3} Words/sec: {4} Sents/sec: {5}'.format(disp_time, progress.eidx, progress.uidx, total_loss/n_words, n_words/duration, n_sents/duration))
last_time = time.time()
total_loss = 0.
n_sents = 0
n_words = 0
if config.sample_freq and progress.uidx % config.sample_freq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :10], x_mask_in[:, :10], y_in[:, :10]
samples = model_set.sample(sess, x_small, x_mask_small)
assert len(samples) == len(x_small.T) == len(y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
sample = util.seq2words(ss, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
logging.info('SAMPLE: {}'.format(sample))
if config.beam_freq and progress.uidx % config.beam_freq == 0:
x_small, x_mask_small, y_small = x_in[:, :, :10], x_mask_in[:, :10], y_in[:,:10]
samples = model_set.beam_search(sess, x_small, x_mask_small,
config.beam_size,
normalization_alpha=config.normalization_alpha)
# samples is a list with shape batch x beam x len
assert len(samples) == len(x_small.T) == len(y_small.T), (len(samples), x_small.shape, y_small.shape)
for xx, yy, ss in zip(x_small.T, y_small.T, samples):
source = util.factoredseq2words(xx, num_to_source)
target = util.seq2words(yy, num_to_target)
logging.info('SOURCE: {}'.format(source))
logging.info('TARGET: {}'.format(target))
for i, (sample_seq, cost) in enumerate(ss):
sample = util.seq2words(sample_seq, num_to_target)
msg = 'SAMPLE {}: {} Cost/Len/Avg {}/{}/{}'.format(
i, sample, cost, len(sample), cost/len(sample))
logging.info(msg)
if config.valid_freq and progress.uidx % config.valid_freq == 0:
valid_ce = validate(sess, replicas[0], config,
valid_text_iterator)
if (len(progress.history_errs) == 0 or
valid_ce < min(progress.history_errs)):
progress.history_errs.append(valid_ce)
progress.bad_counter = 0
save_non_checkpoint(sess, saver, config.saveto)
progress_path = '{0}.progress.json'.format(config.saveto)
progress.save_to_json(progress_path)
else:
progress.history_errs.append(valid_ce)
progress.bad_counter += 1
if progress.bad_counter > config.patience:
logging.info('Early Stop!')
progress.estop = True
break
if config.valid_script is not None:
score = validate_with_script(sess, replicas[0], config)
need_to_save = (score is not None and
(len(progress.valid_script_scores) == 0 or
score > max(progress.valid_script_scores)))
if score is None:
score = 0.0 # ensure a valid value is written
progress.valid_script_scores.append(score)
if need_to_save:
progress.bad_counter = 0
save_path = config.saveto + ".best-valid-script"
save_non_checkpoint(sess, saver, save_path)
write_config_to_json_file(config, save_path)
progress_path = '{}.progress.json'.format(save_path)
progress.save_to_json(progress_path)
if config.save_freq and progress.uidx % config.save_freq == 0:
saver.save(sess, save_path=config.saveto, global_step=progress.uidx)
write_config_to_json_file(config, "%s-%s" % (config.saveto, progress.uidx))
progress_path = '{0}-{1}.progress.json'.format(config.saveto, progress.uidx)
progress.save_to_json(progress_path)
if config.finish_after and progress.uidx % config.finish_after == 0:
logging.info("Maximum number of updates reached")
saver.save(sess, save_path=config.saveto, global_step=progress.uidx)
write_config_to_json_file(config, "%s-%s" % (config.saveto, progress.uidx))
progress.estop=True
progress_path = '{0}-{1}.progress.json'.format(config.saveto, progress.uidx)
progress.save_to_json(progress_path)
break
if progress.estop:
break