def _compute_scores(self, src_filename, trg_filename):
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=src_filename,
vocabulary=self.vocab_src,
is_train_dataset=False,
),
TextLineDataset(data_path=trg_filename,
vocabulary=self.vocab_tgt,
is_train_dataset=False
)
)
# 其实好像还是会打乱
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=40,
use_bucket=True,
buffer_size=100000,
numbering=True,
shuffle=False
)
valid_iter = valid_iterator.build_generator()
score_result = dict()
self.model.eval()
with torch.no_grad():
for batch in valid_iter:
seq_numbers, seqs_x, seqs_y = batch
x, y = prepare_data(seqs_x, seqs_y, cuda=True)
y_inp = y[:, :-1].contiguous()
y_label = y[:, 1:].contiguous()
log_probs = self.model(x, y_inp) # [batch_size, seq_len, vocab_size]
batch_size, seq_len = y_label.shape
log_probs = log_probs.view(-1, self.vocab_tgt.max_n_words)
y_label = y_label.view(-1)
loss = F.nll_loss(log_probs, y_label, reduce=False, ignore_index=self.vocab_tgt.pad) # 越小越好
loss = loss.view(batch_size, seq_len)
loss = loss.sum(-1)
y_label = y_label.view(batch_size, seq_len)
valid_token = (y_label != self.vocab_tgt.pad).sum(-1)
norm_loss = loss.double().div(valid_token.double())
for seq_num, l, nl in zip(seq_numbers, loss, norm_loss):
score_result.update({seq_num: (l.item(), nl.item())})
# for i1, y_l in enumerate(y_label):
# score = 0
# for i2, y_index in enumerate(y_l):
# if y_index.item() == 0:
# break
# score += log_probs[i1][i2][y_index.item()].item()
# score_result.update({seq_numbers[i1]: score})
return score_result
def test_attack(config_path,
save_to,
model_name="attacker",
shuffle=True,
use_gpu=True):
"""
attack
:param config_path: attack attack configs
:param save_to: (string) saving directories
:param model_name: (string) for saving names
:param shuffle: (boolean) for batch scheme, shuffle data set
:param use_gpu: (boolean) on gpu or not
:return: attacked sequences
"""
# initiate
with open(config_path.strip()) as f:
configs = yaml.load(f)
attack_configs = configs["attack_configs"]
attacker_model_configs = configs["attacker_model_configs"]
attacker_optim_configs = configs["attacker_optimizer_configs"]
training_configs = configs["training_configs"]
victim_config_path = attack_configs["victim_configs"]
victim_model_path = attack_configs["victim_model"]
with open(victim_config_path.strip()) as v_f:
print("open victim configs...%s" % victim_config_path)
victim_configs = yaml.load(v_f)
data_configs = victim_configs["data_configs"]
# building inputs
vocab_src = Vocabulary(**data_configs["vocabularies"][0])
vocab_trg = Vocabulary(**data_configs["vocabularies"][1])
# parallel data binding
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0]),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_trg,
max_len=data_configs['max_len'][1]),
shuffle=shuffle)
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs["valid_data"][0],
vocabulary=vocab_src,
max_len=data_configs["max_len"][0]),
TextLineDataset(data_path=data_configs["valid_data"][1],
vocabulary=vocab_trg,
max_len=data_configs["max_len"][1]),
shuffle=shuffle)
train_batch_size = training_configs["batch_size"]
train_buffer_size = training_configs["buffer_size"]
training_iterator = DataIterator(
dataset=train_bitext_dataset,
batch_size=train_batch_size,
use_bucket=training_configs['use_bucket'],
buffer_size=train_buffer_size,
batching_func=training_configs['batching_key'])
# valid_iterator is bucketed by length to accelerate decoding (numbering to mark orders)
valid_iterator = DataIterator(
dataset=valid_bitext_dataset,
batch_size=training_configs["valid_batch_size"],
use_bucket=True,
buffer_size=50000,
numbering=True)
# initiate saver
model_collections = Collections()
checkpoint_saver = Saver(
save_prefix="{0}.ckpt".format(os.path.join(save_to, model_name)),
num_max_keeping=training_configs['num_kept_checkpoints'])
w2p, w2vocab = load_or_extract_near_vocab(
config_path=victim_config_path,
model_path=victim_model_path,
init_perturb_rate=attack_configs["init_perturb_rate"],
save_to=os.path.join(save_to, "near_vocab"),
save_to_full=os.path.join(save_to, "full_near_vocab"),
top_reserve=12,
emit_as_id=True)
# build attacker
# attacker = Attacker(n_words=vocab_src.max_n_words,
# **attacker_model_configs)
# if use_gpu:
# attacker = attacker.cuda()
# CURRENT_DEVICE = "cuda"
# else:
# CURRENT_DEVICE = "cpu"
# load embedding from trained NMT models
# load_embedding(attacker, model_path=victim_model_path, device=CURRENT_DEVICE)
# attacker.eval()
# for i in range(6):
train_iter = training_iterator.build_generator()
batch = train_iter.__next__()
print(batch[1][3])
def test_discriminator(config_path,
save_to,
model_name="Discriminator",
shuffle=True,
use_gpu=True):
with open(config_path.strip()) as f:
configs = yaml.load(f)
attack_configs = configs["attack_configs"]
discriminator_configs = configs["discriminator_configs"]
discriminator_model_configs = discriminator_configs[
"discriminator_model_configs"]
discriminator_optim_configs = discriminator_configs[
"discriminator_optimizer_configs"]
victim_config_path = attack_configs["victim_configs"]
victim_model_path = attack_configs["victim_model"]
with open(victim_config_path.strip()) as v_f:
print("open victim configs...%s" % victim_config_path)
victim_configs = yaml.load(v_f)
data_configs = victim_configs["data_configs"]
# building inputs
vocab_src = Vocabulary(**data_configs["vocabularies"][0])
vocab_trg = Vocabulary(**data_configs["vocabularies"][1])
# parallel data binding
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0]),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_trg,
max_len=data_configs['max_len'][1]),
shuffle=shuffle)
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs["valid_data"][0],
vocabulary=vocab_src,
max_len=data_configs["max_len"][0]),
TextLineDataset(data_path=data_configs["valid_data"][1],
vocabulary=vocab_trg,
max_len=data_configs["max_len"][1]),
shuffle=shuffle)
train_batch_size = attack_configs["batch_size"]
train_buffer_size = attack_configs["buffer_size"]
training_iterator = DataIterator(
dataset=train_bitext_dataset,
batch_size=train_batch_size,
use_bucket=attack_configs['use_bucket'],
buffer_size=train_buffer_size,
batching_func=attack_configs['batching_key'])
# valid_iterator is bucketed by length to accelerate decoding (numbering to mark orders)
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=attack_configs["batch_size"],
use_bucket=True,
buffer_size=50000,
numbering=True)
# initiate saver
model_collections = Collections()
checkpoint_saver = Saver(
save_prefix="{0}.ckpt".format(os.path.join(save_to, model_name)),
num_max_keeping=attack_configs['num_kept_checkpoints'])
# building model
model_D = TransDiscriminator(n_src_words=vocab_src.max_n_words,
n_trg_words=vocab_trg.max_n_words,
**discriminator_model_configs)
if use_gpu:
model_D = model_D.cuda()
CURRENT_DEVICE = "cuda"
else:
CURRENT_DEVICE = "cpu"
# load embedding from trained NMT models
load_embedding(model_D,
model_path=victim_model_path,
device=CURRENT_DEVICE)
# TODO reloading parameters
# classification need label smoothing to trigger Negative log likelihood loss
criterion = nn.CrossEntropyLoss()
# building optimizer
optim = Optimizer(
name=discriminator_optim_configs["optimizer"],
model=model_D,
lr=discriminator_optim_configs["learning_rate"],
grad_clip=discriminator_optim_configs["grad_clip"],
optim_args=discriminator_optim_configs["optimizer_params"])
# Build scheduler for optimizer if needed
if discriminator_optim_configs['schedule_method'] is not None:
if discriminator_optim_configs['schedule_method'] == "loss":
scheduler = ReduceOnPlateauScheduler(
optimizer=optim,
**discriminator_optim_configs["scheduler_configs"])
elif discriminator_optim_configs['schedule_method'] == "noam":
scheduler = NoamScheduler(
optimizer=optim,
**discriminator_optim_configs['scheduler_configs'])
elif discriminator_optim_configs["schedule_method"] == "rsqrt":
scheduler = RsqrtScheduler(
optimizer=optim,
**discriminator_optim_configs["scheduler_configs"])
else:
WARN(
"Unknown scheduler name {0}. Do not use lr_scheduling.".format(
discriminator_optim_configs['schedule_method']))
scheduler = None
else:
scheduler = None
# reload latest checkpoint
checkpoint_saver.load_latest(model=model_D,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections)
# prepare training
eidx = model_collections.get_collection("eidx", [0])[-1]
uidx = model_collections.get_collection("uidx", [0])[-1]
oom_count = model_collections.get_collection("oom_count", [0])[-1]
summary_writer = SummaryWriter(log_dir=save_to + "log")
w2p, w2vocab = load_or_extract_near_vocab(
config_path=victim_config_path,
model_path=victim_model_path,
init_perturb_rate=attack_configs["init_perturb_rate"],
save_to=os.path.join(save_to, "near_vocab"),
save_to_full=os.path.join(save_to, "full_near_vocab"),
top_reserve=12)
while True: # infinite loop for training epoch
training_iter = training_iterator.build_generator()
for batch in training_iter:
uidx += 1
if discriminator_optim_configs[
"schedule_method"] is not None and discriminator_optim_configs[
"schedule_method"] != "loss":
scheduler.step(global_step=uidx)
# training session
seqs_x, seqs_y = batch # returned tensor type of the data
optim.zero_grad()
try:
x, y, flags = prepare_D_data(w2p,
w2vocab,
victim_config_path,
seqs_x,
seqs_y,
use_gpu=use_gpu)
loss = compute_D_forward(model_D,
criterion=criterion,
seqs_x=x,
seqs_y=y,
gold_flags=flags)
optim.step()
print("loss:", loss)
except RuntimeError as e:
if "out of memory" in str(e):
print("WARNING: out of memory, skipping batch")
oom_count += 1
optim.zero_grad()
else:
raise e
# check for validation and save the model
if should_trigger_by_steps(
uidx,
eidx,
every_n_step=discriminator_configs["acc_valid_freq"]):
lrate = list(optim.get_lrate())[0]
summary_writer.add_scalar("lrate",
scalar_value=lrate,
global_step=uidx)
summary_writer.add_scalar("oom_count",
scalar_value=oom_count,
global_step=uidx)
if should_trigger_by_steps(
uidx, eidx, every_n_step=attack_configs["save_freq"]):
model_collections.add_to_collection("uidx", uidx)
model_collections.add_to_collection("eidx", eidx)
checkpoint_saver.save(global_step=uidx,
model=model_D,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections)
if should_trigger_by_steps(
uidx,
eidx,
every_n_step=discriminator_configs["acc_valid_freq"]):
# validate average loss over samples on validation set
n_sents = 0.
sum_loss = 0.0
valid_iter = valid_iterator.build_generator()
for batch in valid_iter:
_, seqs_x, seqs_y = batch
n_sents += len(seqs_x)
x, y, flags = prepare_D_data(w2p,
w2vocab,
victim_config_path,
seqs_x,
seqs_y,
use_gpu=use_gpu)
loss = compute_D_forward(model_D,
criterion,
x,
y,
gold_flags=flags,
eval=True)
if np.isnan(loss):
WARN("NaN detected!")
sum_loss += float(loss)
eval_loss = float(sum_loss / n_sents)
summary_writer.add_scalar("valid",
scalar_value=eval_loss,
global_step=uidx)
if should_trigger_by_steps(
uidx,
eidx,
every_n_step=discriminator_configs["acc_valid_freq"]):
# validate accuracy of the discriminator
acc = acc_validation(uidx=uidx,
discriminator_model=model_D,
valid_iterator=valid_iterator,
victim_configs=victim_config_path,
w2p=w2p,
w2vocab=w2vocab,
batch_size=attack_configs["batch_size"],
use_gpu=use_gpu)
summary_writer.add_scalar("accuracy",
scalar_value=acc,
global_step=uidx)
eidx += 1
pass
def train(FLAGS):
"""
FLAGS:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# write log of training to file.
write_log_to_file(
os.path.join(FLAGS.log_path,
"%s.log" % time.strftime("%Y%m%d-%H%M%S")))
GlobalNames.USE_GPU = FLAGS.use_gpu
if GlobalNames.USE_GPU:
CURRENT_DEVICE = "cpu"
else:
CURRENT_DEVICE = "cuda:0"
config_path = os.path.abspath(FLAGS.config_path)
with open(config_path.strip()) as f:
configs = yaml.load(f)
INFO(pretty_configs(configs))
# Add default configs
configs = default_configs(configs)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
optimizer_configs = configs['optimizer_configs']
training_configs = configs['training_configs']
GlobalNames.SEED = training_configs['seed']
set_seed(GlobalNames.SEED)
best_model_prefix = os.path.join(
FLAGS.saveto, FLAGS.model_name + GlobalNames.MY_BEST_MODEL_SUFFIX)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_tgt = Vocabulary(**data_configs["vocabularies"][0])
train_batch_size = training_configs["batch_size"] * max(
1, training_configs["update_cycle"])
train_buffer_size = training_configs["buffer_size"] * max(
1, training_configs["update_cycle"])
train_bitext_dataset = ZipDataset(TextLineDataset(
data_path=data_configs['train_data'][0],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][0],
),
shuffle=training_configs['shuffle'])
valid_bitext_dataset = ZipDataset(
TextLineDataset(
data_path=data_configs['valid_data'][0],
vocabulary=vocab_tgt,
))
training_iterator = DataIterator(
dataset=train_bitext_dataset,
batch_size=train_batch_size,
use_bucket=training_configs['use_bucket'],
buffer_size=train_buffer_size,
batching_func=training_configs['batching_key'])
valid_iterator = DataIterator(
dataset=valid_bitext_dataset,
batch_size=training_configs['valid_batch_size'],
use_bucket=True,
buffer_size=100000,
numbering=True)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
lrate = optimizer_configs['learning_rate']
is_early_stop = False
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
model_collections = Collections()
checkpoint_saver = Saver(
save_prefix="{0}.ckpt".format(
os.path.join(FLAGS.saveto, FLAGS.model_name)),
num_max_keeping=training_configs['num_kept_checkpoints'])
best_model_saver = Saver(
save_prefix=best_model_prefix,
num_max_keeping=training_configs['num_kept_best_model'])
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
lm_model = build_model(n_tgt_vocab=vocab_tgt.max_n_words, **model_configs)
INFO(lm_model)
params_total = sum([p.numel() for n, p in lm_model.named_parameters()])
params_with_embedding = sum([
p.numel() for n, p in lm_model.named_parameters()
if n.find('embedding') == -1
])
INFO('Total parameters: {}'.format(params_total))
INFO('Total parameters (excluding word embeddings): {}'.format(
params_with_embedding))
critic = NMTCriterion(label_smoothing=model_configs['label_smoothing'])
INFO(critic)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# 2. Move to GPU
if GlobalNames.USE_GPU:
lm_model = lm_model.cuda()
critic = critic.cuda()
# 3. Load pretrained model if needed
lm_model.init_parameters(FLAGS.pretrain_path, device=CURRENT_DEVICE)
# 4. Build optimizer
INFO('Building Optimizer...')
optim = Optimizer(name=optimizer_configs['optimizer'],
model=lm_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params'])
# 5. Build scheduler for optimizer if needed
if optimizer_configs['schedule_method'] is not None:
if optimizer_configs['schedule_method'] == "loss":
scheduler = ReduceOnPlateauScheduler(
optimizer=optim, **optimizer_configs["scheduler_configs"])
elif optimizer_configs['schedule_method'] == "noam":
scheduler = NoamScheduler(optimizer=optim,
**optimizer_configs['scheduler_configs'])
else:
WARN(
"Unknown scheduler name {0}. Do not use lr_scheduling.".format(
optimizer_configs['schedule_method']))
scheduler = None
else:
scheduler = None
# 6. build moving average
if training_configs['moving_average_method'] is not None:
ma = MovingAverage(
moving_average_method=training_configs['moving_average_method'],
named_params=lm_model.named_parameters(),
alpha=training_configs['moving_average_alpha'])
else:
ma = None
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# Reload from latest checkpoint
if FLAGS.reload:
checkpoint_saver.load_latest(model=lm_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma)
# ================================================================================== #
# Prepare training
eidx = model_collections.get_collection("eidx", [0])[-1]
uidx = model_collections.get_collection("uidx", [0])[-1]
bad_count = model_collections.get_collection("bad_count", [0])[-1]
oom_count = model_collections.get_collection("oom_count", [0])[-1]
summary_writer = SummaryWriter(log_dir=FLAGS.log_path)
cum_samples = 0
cum_words = 0
valid_loss = best_valid_loss = float('inf') # Max Float
saving_files = []
# Timer for computing speed
timer_for_speed = Timer()
timer_for_speed.tic()
INFO('Begin training...')
while True:
summary_writer.add_scalar("Epoch", (eidx + 1), uidx)
# Build iterator and progress bar
training_iter = training_iterator.build_generator()
training_progress_bar = tqdm(desc=' - (Epc {}, Upd {}) '.format(
eidx, uidx),
total=len(training_iterator),
unit="sents")
for batch in training_iter:
uidx += 1
if optimizer_configs[
"schedule_method"] is not None and optimizer_configs[
"schedule_method"] != "loss":
scheduler.step(global_step=uidx)
seqs_y = batch
n_samples_t = len(seqs_y)
n_words_t = sum(len(s) for s in seqs_y)
cum_samples += n_samples_t
cum_words += n_words_t
train_loss = 0.
optim.zero_grad()
try:
# Prepare data
for (seqs_y_t, ) in split_shard(
seqs_y, split_size=training_configs['update_cycle']):
y = prepare_data(seqs_y_t, cuda=GlobalNames.USE_GPU)
loss = compute_forward(
model=lm_model,
critic=critic,
# seqs_x=x,
seqs_y=y,
eval=False,
normalization=n_samples_t,
norm_by_words=training_configs["norm_by_words"])
train_loss += loss / y.size(
1) if not training_configs["norm_by_words"] else loss
optim.step()
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
oom_count += 1
optim.zero_grad()
else:
raise e
if ma is not None and eidx >= training_configs[
'moving_average_start_epoch']:
ma.step()
training_progress_bar.update(n_samples_t)
training_progress_bar.set_description(
' - (Epc {}, Upd {}) '.format(eidx, uidx))
training_progress_bar.set_postfix_str(
'TrainLoss: {:.2f}, ValidLoss(best): {:.2f} ({:.2f})'.format(
train_loss, valid_loss, best_valid_loss))
summary_writer.add_scalar("train_loss",
scalar_value=train_loss,
global_step=uidx)
# ================================================================================== #
# Display some information
if should_trigger_by_steps(
uidx, eidx, every_n_step=training_configs['disp_freq']):
# words per second and sents per second
words_per_sec = cum_words / (timer.toc(return_seconds=True))
sents_per_sec = cum_samples / (timer.toc(return_seconds=True))
lrate = list(optim.get_lrate())[0]
summary_writer.add_scalar("Speed(words/sec)",
scalar_value=words_per_sec,
global_step=uidx)
summary_writer.add_scalar("Speed(sents/sen)",
scalar_value=sents_per_sec,
global_step=uidx)
summary_writer.add_scalar("lrate",
scalar_value=lrate,
global_step=uidx)
summary_writer.add_scalar("oom_count",
scalar_value=oom_count,
global_step=uidx)
# Reset timer
timer.tic()
cum_words = 0
cum_samples = 0
# ================================================================================== #
# Saving checkpoints
if should_trigger_by_steps(
uidx,
eidx,
every_n_step=training_configs['save_freq'],
debug=FLAGS.debug):
model_collections.add_to_collection("uidx", uidx)
model_collections.add_to_collection("eidx", eidx)
model_collections.add_to_collection("bad_count", bad_count)
if not is_early_stop:
checkpoint_saver.save(global_step=uidx,
model=lm_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma)
# ================================================================================== #
# Loss Validation & Learning rate annealing
if should_trigger_by_steps(
global_step=uidx,
n_epoch=eidx,
every_n_step=training_configs['loss_valid_freq'],
debug=FLAGS.debug):
if ma is not None:
origin_state_dict = deepcopy(lm_model.state_dict())
lm_model.load_state_dict(ma.export_ma_params(),
strict=False)
valid_loss = loss_validation(
model=lm_model,
critic=critic,
valid_iterator=valid_iterator,
norm_by_words=training_configs["norm_by_words"])
model_collections.add_to_collection("history_losses",
valid_loss)
min_history_loss = np.array(
model_collections.get_collection("history_losses")).min()
summary_writer.add_scalar("loss", valid_loss, global_step=uidx)
summary_writer.add_scalar("best_loss",
min_history_loss,
global_step=uidx)
if ma is not None:
lm_model.load_state_dict(origin_state_dict)
del origin_state_dict
if optimizer_configs["schedule_method"] == "loss":
scheduler.step(metric=best_valid_loss)
# If model get new best valid loss
if valid_loss < best_valid_loss:
bad_count = 0
if is_early_stop is False:
# 1. save the best model
torch.save(lm_model.state_dict(),
best_model_prefix + ".final")
# 2. record all several best models
best_model_saver.save(global_step=uidx, model=lm_model)
else:
bad_count += 1
# At least one epoch should be traversed
if bad_count >= training_configs[
'early_stop_patience'] and eidx > 0:
is_early_stop = True
WARN("Early Stop!")
best_valid_loss = min_history_loss
summary_writer.add_scalar("bad_count", bad_count, uidx)
INFO("{0} Loss: {1:.2f} lrate: {2:6f} patience: {3}".format(
uidx, valid_loss, lrate, bad_count))
training_progress_bar.close()
eidx += 1
if eidx > training_configs["max_epochs"]:
break
def train(config_path, model_path, model_type, src_filename, trg_filename):
"""
flags:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# ================================================================================== #
# Initialization for training on different devices
# - CPU/GPU
# - Single/Distributed
Constants.USE_GPU = True
print(config_path)
print(model_path)
print(model_type)
world_size = 1
rank = 0
local_rank = 0
if Constants.USE_GPU:
torch.cuda.set_device(local_rank)
Constants.CURRENT_DEVICE = "cuda:{0}".format(local_rank)
else:
Constants.CURRENT_DEVICE = "cpu"
# ================================================================================== #
# Parsing configuration files
# - Load default settings
# - Load pre-defined settings
# - Load user-defined settings
configs = prepare_configs(config_path)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
training_configs = configs['training_configs']
INFO(pretty_configs(configs))
Constants.SEED = training_configs['seed']
set_seed(Constants.SEED)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary.build_from_file(**data_configs['vocabularies'][0])
vocab_tgt = Vocabulary.build_from_file(**data_configs['vocabularies'][1])
Constants.EOS = vocab_src.eos
Constants.PAD = vocab_src.pad
Constants.BOS = vocab_src.bos
valid_bitext_dataset = ZipDataset(
TextLineDataset(
data_path=src_filename,
vocabulary=vocab_src,
max_len=100,
is_train_dataset=False,
),
TextLineDataset(
data_path=trg_filename,
vocabulary=vocab_tgt,
is_train_dataset=False,
max_len=100,
))
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=20,
use_bucket=training_configs['use_bucket'],
buffer_size=training_configs['buffer_size'],
numbering=True,
world_size=world_size,
rank=rank)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words,
padding_idx=vocab_src.pad,
vocab_src=vocab_src,
**model_configs)
INFO(nmt_model)
# 2. Move to GPU
if Constants.USE_GPU:
nmt_model = nmt_model.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model,
model_path,
device=Constants.CURRENT_DEVICE)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# ================================================================================== #
# Prepare training
sent_per_sec_meter = TimeMeter()
tok_per_sec_meter = TimeMeter()
grad_denom = 0
train_loss = 0.0
cum_n_words = 0
valid_loss = best_valid_loss = float('inf')
sent_per_sec_meter.start()
tok_per_sec_meter.start()
INFO('Begin training...')
eidx = 0
uidx = 0
score_result = dict()
# Build iterator and progress bar
training_iter = valid_iterator.build_generator()
training_progress_bar = tqdm(desc=' - (Epc {}, Upd {}) '.format(
eidx, uidx),
total=len(valid_iterator),
unit="sents")
for batch in training_iter:
seqs_numbers, seqs_x, seqs_y = batch
batch_size = len(seqs_x)
cum_n_words += sum(len(s) for s in seqs_y)
try:
# Prepare data
x, y = prepare_data(seqs_x, seqs_y, cuda=Constants.USE_GPU)
y_inp = y[:, :-1].contiguous()
y_label = y[:, 1:].contiguous() # [batch_size, seq_len]
log_probs = nmt_model(
x, y_inp, log_probs=True) # [batch_size, seq_len, vocab_size]
_, seq_len = y_label.shape
log_probs = log_probs.view(-1, vocab_tgt.max_n_words)
y_label = y_label.view(-1)
loss = F.nll_loss(log_probs,
y_label,
reduce=False,
ignore_index=vocab_tgt.pad)
loss = loss.view(batch_size, seq_len)
loss = loss.sum(-1)
y_label = y_label.view(batch_size, seq_len)
valid_token = (y_label != vocab_tgt.pad).sum(-1)
loss = loss.double().div(valid_token.double())
for seq_num, l in zip(seqs_numbers, loss):
assert seq_num not in score_result
score_result.update({seq_num: l.item()})
uidx += 1
grad_denom += batch_size
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
else:
raise e
if training_progress_bar is not None:
training_progress_bar.update(batch_size)
training_progress_bar.set_description(
' - (Epc {}, Upd {}) '.format(eidx, uidx))
postfix_str = 'TrainLoss: {:.2f}, ValidLoss(best): {:.2f} ({:.2f}), '.format(
train_loss, valid_loss, best_valid_loss)
training_progress_bar.set_postfix_str(postfix_str)
training_progress_bar.close()
return score_result
def train2(flags):
"""
flags:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# ================================================================================== #
# Initialization for training on different devices
# - CPU/GPU
# - Single/Distributed
Constants.USE_GPU = flags.use_gpu
world_size = 1
rank = 0
local_rank = 0
if Constants.USE_GPU:
torch.cuda.set_device(local_rank)
Constants.CURRENT_DEVICE = "cuda:{0}".format(local_rank)
else:
Constants.CURRENT_DEVICE = "cpu"
# ================================================================================== #
# Parsing configuration files
# - Load default settings
# - Load pre-defined settings
# - Load user-defined settings
configs = prepare_configs(flags.config_path, flags.predefined_config)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
training_configs = configs['training_configs']
bt_configs = configs['bt_configs'] if 'bt_configs' in configs else None
if bt_configs is not None:
print("btconfigs ", bt_configs)
if 'bt_attribute_data' not in bt_configs:
Constants.USE_BT = False
bt_configs = None
else:
Constants.USE_BT = True
Constants.USE_BTTAG = bt_configs['use_bttag']
Constants.USE_CONFIDENCE = bt_configs['use_confidence']
INFO(pretty_configs(configs))
Constants.SEED = training_configs['seed']
set_seed(Constants.SEED)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary.build_from_file(**data_configs['vocabularies'][0])
vocab_tgt = Vocabulary.build_from_file(**data_configs['vocabularies'][1])
Constants.EOS = vocab_src.eos
Constants.PAD = vocab_src.pad
Constants.BOS = vocab_src.bos
# bt tag dataset
if Constants.USE_BT:
if Constants.USE_BTTAG:
Constants.BTTAG = vocab_src.bttag
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0],
is_train_dataset=True
),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][1],
is_train_dataset=True
),
AttributeDataset(data_path=bt_configs['bt_attribute_data'], is_train_dataset=True)
)
else:
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0],
is_train_dataset=True
),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][1],
is_train_dataset=True
)
)
training_iterator = DataIterator(dataset=train_bitext_dataset,
batch_size=training_configs["batch_size"],
use_bucket=training_configs['use_bucket'],
buffer_size=training_configs['buffer_size'],
batching_func=training_configs['batching_key'],
world_size=world_size, numbering=True,
rank=rank)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words, padding_idx=vocab_src.pad, vocab_src=vocab_src,
vocab_tgt=vocab_tgt,
**model_configs)
INFO(nmt_model)
# 2. Move to GPU
if Constants.USE_GPU:
nmt_model = nmt_model.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model, flags.pretrain_path, exclude_prefix=flags.pretrain_exclude_prefix,
device=Constants.CURRENT_DEVICE)
nmt_model = nmt_model.encoder
INFO('Done. Elapsed time {0}'.format(timer.toc()))
INFO('Begin training...')
# 计算train集合每个句子的表示：mean pool
training_iter = training_iterator.build_generator()
nmt_model.eval()
all_seq_numbers = []
encoder_filename = "/home/wangdq/encoder.mean.output"
seq_numbers_filename = '/home/wangdq/seq_numbers.output'
processd = 0
with open(encoder_filename, 'w') as f_encoder, open(seq_numbers_filename, 'w') as f_seq_numbers:
for batch in training_iter:
bt_attrib = None
# bt attrib data
if Constants.USE_BT:
seq_numbers, seqs_x, seqs_y, bt_attrib = batch # seq_numerbs从0开始编号
else:
seq_numbers, seqs_x, seqs_y = batch
x = prepare_data(seqs_x, seqs_y=None, cuda=Constants.USE_GPU, bt_attrib=bt_attrib)
try:
with torch.no_grad():
encoder_hidden, mask = nmt_model(x)
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
else:
raise e
valid_hidden = (mask == False).float().cuda()
sum_encoder_hidden = (encoder_hidden * valid_hidden.unsqueeze(-1)).sum(dim=1)
valid_tokens = (mask == False).sum(-1)
mean_encoder_hidden = sum_encoder_hidden.float().div(valid_tokens.unsqueeze(1))
all_seq_numbers.extend(seq_numbers)
# if all_mean_encoder_hidden is None:
# all_mean_encoder_hidden = mean_encoder_hidden.cpu()
# else:
# all_mean_encoder_hidden = torch.cat((all_mean_encoder_hidden, mean_encoder_hidden.cpu()), dim=0)
mean_encoder_list = mean_encoder_hidden.cpu().numpy().tolist()
content = [[str(i) for i in mean] for mean in mean_encoder_list]
content = [' '.join(mean) + '\n' for mean in content]
f_encoder.writelines(content)
processd += len(seq_numbers)
print(processd)
content = [str(i) for i in all_seq_numbers]
content = ' '.join(content)
f_seq_numbers.writelines(content)
def test_data(flags):
Constants.USE_GPU = flags.use_gpu
world_size = 1
rank = 0
local_rank = 0
if Constants.USE_GPU:
torch.cuda.set_device(local_rank)
Constants.CURRENT_DEVICE = "cuda:{0}".format(local_rank)
else:
Constants.CURRENT_DEVICE = "cpu"
# ================================================================================== #
# Parsing configuration files
# - Load default settings
# - Load pre-defined settings
# - Load user-defined settings
configs = prepare_configs(flags.config_path, flags.predefined_config)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
training_configs = configs['training_configs']
bt_configs = configs['bt_configs'] if 'bt_configs' in configs else None
if bt_configs is not None:
print("btconfigs ", bt_configs)
if 'bt_attribute_data' not in bt_configs:
Constants.USE_BT = False
bt_configs = None
else:
Constants.USE_BT = True
Constants.USE_BTTAG = bt_configs['use_bttag']
Constants.USE_CONFIDENCE = bt_configs['use_confidence']
INFO(pretty_configs(configs))
Constants.SEED = training_configs['seed']
set_seed(Constants.SEED)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary.build_from_file(**data_configs['vocabularies'][0])
vocab_tgt = Vocabulary.build_from_file(**data_configs['vocabularies'][1])
Constants.EOS = vocab_src.eos
Constants.PAD = vocab_src.pad
Constants.BOS = vocab_src.bos
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['valid_data'][0],
vocabulary=vocab_src,
is_train_dataset=False,
),
TextLineDataset(data_path=data_configs['valid_data'][1],
vocabulary=vocab_tgt,
is_train_dataset=False
)
)
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=training_configs['valid_batch_size'],
use_bucket=True, buffer_size=100000, numbering=True,
world_size=world_size, rank=rank, shuffle=False)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words, padding_idx=vocab_src.pad, vocab_src=vocab_src,
vocab_tgt=vocab_tgt,
**model_configs)
INFO(nmt_model)
# 2. Move to GPU
if Constants.USE_GPU:
nmt_model = nmt_model.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model, flags.pretrain_path, exclude_prefix=flags.pretrain_exclude_prefix,
device=Constants.CURRENT_DEVICE)
nmt_model = nmt_model.encoder
INFO('Done. Elapsed time {0}'.format(timer.toc()))
INFO('Begin training...')
# 计算train集合每个句子的表示：mean pool
nmt_model.eval()
# 计算test集合每个句子的表示： mean pool
valid_iter = valid_iterator.build_generator()
all_seq_numbers = []
all_mean_encoder_hidden = None
for batch in valid_iter:
bt_attrib = None
seq_numbers, seqs_x, seqs_y = batch
all_seq_numbers.extend(seq_numbers)
x = prepare_data(seqs_x, seqs_y=None, cuda=Constants.USE_GPU, bt_attrib=bt_attrib)
try:
with torch.no_grad():
encoder_hidden, mask = nmt_model(x)
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
else:
raise e
valid_hidden = (mask == False).float().cuda()
sum_encoder_hidden = (encoder_hidden * valid_hidden.unsqueeze(-1)).sum(dim=1)
valid_tokens = (mask == False).sum(-1)
mean_encoder_hidden = sum_encoder_hidden.float().div(valid_tokens.unsqueeze(1))
if all_mean_encoder_hidden is None:
all_mean_encoder_hidden = mean_encoder_hidden
else:
all_mean_encoder_hidden = torch.cat((all_mean_encoder_hidden, mean_encoder_hidden), dim=0)
return all_mean_encoder_hidden, all_seq_numbers
def _compute_scores(self, src_filename, trg_filename):
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=src_filename,
vocabulary=self.vocab_src,
is_train_dataset=False,
max_len=100),
TextLineDataset(data_path=trg_filename,
vocabulary=self.vocab_tgt,
is_train_dataset=False,
max_len=100))
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=20,
use_bucket=True,
buffer_size=1000,
numbering=True,
shuffle=False)
valid_iter = valid_iterator.build_generator()
score_result = dict()
self.model.eval()
eidx = 0
uidx = 0
training_progress_bar = tqdm(desc=' - (Epc {}, Upd {}) '.format(
eidx, uidx),
total=len(valid_iterator),
unit="sents")
with torch.no_grad():
for batch in valid_iter:
seq_numbers, seqs_x, seqs_y = batch
x, y = prepare_data(seqs_x, seqs_y, cuda=True)
y_inp = y[:, :-1].contiguous()
y_label = y[:, 1:].contiguous() # [batch_size, seq_len]
log_probs = self.model(
x, y_inp,
log_probs=True) # [batch_size, seq_len, vocab_size]
batch_size, seq_len = y_label.shape
log_probs = log_probs.view(-1, self.vocab_tgt.max_n_words)
y_label = y_label.view(-1)
loss = F.nll_loss(log_probs,
y_label,
reduce=False,
ignore_index=self.vocab_tgt.pad)
loss = loss.view(batch_size, seq_len)
loss = loss.sum(-1)
y_label = y_label.view(batch_size, seq_len)
valid_token = (y_label != self.vocab_tgt.pad).sum(-1)
loss = loss.double().div(valid_token.double())
for seq_num, l in zip(seq_numbers, loss):
assert seq_num not in score_result
score_result.update({seq_num: l.item()})
training_progress_bar.update(batch_size)
training_progress_bar.set_description(
' - (Epc {}, Upd {}) '.format(eidx, uidx))
# for i1, y_l in enumerate(y_label):
# score = 0
# for i2, y_index in enumerate(y_l):
# if y_index.item() == 0:
# break
# score += log_probs[i1][i2][y_index.item()].item()
# valid_token = (y_label != self.vocab_tgt.pad).long().sum().item()
# score = -1 * score / valid_token
# score_result.update({seq_numbers[i1]: score})
return score_result
def run():
# default actor threads as 1
os.environ["OMP_NUM_THREADS"] = "1"
mp = _mp.get_context('spawn')
args = parser.parse_args()
if not os.path.exists(args.save_to):
os.mkdir(args.save_to)
with open(args.config_path, "r") as f, \
open(os.path.join(args.save_to, "current_attack_configs.yaml"), "w") as current_configs:
configs = yaml.load(f)
yaml.dump(configs, current_configs)
attack_configs = configs["attack_configs"]
attacker_configs = configs["attacker_configs"]
attacker_model_configs = attacker_configs["attacker_model_configs"]
attacker_optimizer_configs = attacker_configs["attacker_optimizer_configs"]
discriminator_configs = configs["discriminator_configs"]
# training_configs = configs["training_configs"]
# initial best saver for global model
global_saver = Saver(
save_prefix="{0}.final".format(os.path.join(args.save_to, "ACmodel")),
num_max_keeping=attack_configs["num_kept_checkpoints"])
# the Global variable of USE_GPU is mainly used for environments
GlobalNames.SEED = attack_configs["seed"]
GlobalNames.USE_GPU = args.use_gpu
torch.manual_seed(GlobalNames.SEED)
# build vocabulary and data iterator for env
with open(attack_configs["victim_configs"], "r") as victim_f:
victim_configs = yaml.load(victim_f)
data_configs = victim_configs["data_configs"]
src_vocab = Vocabulary(**data_configs["vocabularies"][0])
trg_vocab = Vocabulary(**data_configs["vocabularies"][1])
data_set = ZipDataset(
TextLineDataset(
data_path=data_configs["train_data"][0],
vocabulary=src_vocab,
),
TextLineDataset(
data_path=data_configs["train_data"][1],
vocabulary=trg_vocab,
),
shuffle=attack_configs["shuffle"]
) # we build the parallel data sets and iterate inside a thread
# global model variables (trg network to save the results)
global_attacker = attacker.Attacker(src_vocab.max_n_words,
**attacker_model_configs)
global_attacker = global_attacker.cpu()
global_attacker.share_memory()
if args.share_optim:
# initiate optimizer and set to share mode
optimizer = Optimizer(
name=attacker_optimizer_configs["optimizer"],
model=global_attacker,
lr=attacker_optimizer_configs["learning_rate"],
grad_clip=attacker_optimizer_configs["grad_clip"],
optim_args=attacker_optimizer_configs["optimizer_params"])
optimizer.optim.share_memory()
# Build scheduler for optimizer if needed
if attacker_optimizer_configs['schedule_method'] is not None:
if attacker_optimizer_configs['schedule_method'] == "loss":
scheduler = ReduceOnPlateauScheduler(
optimizer=optimizer,
**attacker_optimizer_configs["scheduler_configs"])
elif attacker_optimizer_configs['schedule_method'] == "noam":
scheduler = NoamScheduler(
optimizer=optimizer,
**attacker_optimizer_configs['scheduler_configs'])
elif attacker_optimizer_configs["schedule_method"] == "rsqrt":
scheduler = RsqrtScheduler(
optimizer=optimizer,
**attacker_optimizer_configs["scheduler_configs"])
else:
WARN("Unknown scheduler name {0}. Do not use lr_scheduling.".
format(attacker_optimizer_configs['schedule_method']))
scheduler = None
else:
scheduler = None
else:
optimizer = None
scheduler = None
# load from checkpoint for global model
global_saver.load_latest(model=global_attacker,
optim=optimizer,
lr_scheduler=scheduler)
if args.use_gpu:
# collect available devices and distribute env on the available gpu
device = "cuda"
devices = []
for i in range(torch.cuda.device_count()):
devices += ["cuda:%d" % i]
print("available gpus:", devices)
else:
device = "cpu"
devices = [device]
process = []
counter = mp.Value("i", 0)
lock = mp.Lock() # for multiple attackers update
INFO("extract near candidates")
_, _ = load_or_extract_near_vocab(
config_path=attack_configs["victim_configs"],
model_path=attack_configs["victim_model"],
init_perturb_rate=attack_configs["init_perturb_rate"],
save_to=os.path.join(args.save_to, "near_vocab"),
save_to_full=os.path.join(args.save_to, "full_near_vocab"),
top_reserve=12,
emit_as_id=True)
# train(0, device, args, counter, lock,
# attack_configs, discriminator_configs,
# src_vocab, trg_vocab, data_set,
# global_attacker, attacker_configs,
# optimizer, scheduler,
# global_saver)
# valid(args.n, device, args,
# attack_configs, discriminator_configs,
# src_vocab, trg_vocab, data_set,
# global_attacker, attacker_configs, counter)
# run multiple training process of local attacker to update global one
for rank in range(args.n):
print("initialize training thread on cuda:%d" % (rank + 1))
p = mp.Process(target=train,
args=(rank, "cuda:%d" % (rank + 1), args, counter, lock,
attack_configs, discriminator_configs, src_vocab,
trg_vocab, data_set, global_attacker,
attacker_configs, optimizer, scheduler,
global_saver))
p.start()
process.append(p)
# run the dev thread for initiation
print("initialize dev thread on cuda:0")
p = mp.Process(target=valid,
args=(0, "cuda:0", args, attack_configs,
discriminator_configs, src_vocab, trg_vocab, data_set,
global_attacker, attacker_configs, counter))
p.start()
process.append(p)
for p in process:
p.join()
def run():
# default actor threads as 1
os.environ["OMP_NUM_THREADS"] = "1"
mp = _mp.get_context('spawn')
args = parser.parse_args()
if not os.path.exists(args.save_to):
os.mkdir(args.save_to)
# load reinforce configs
with open(args.config_path, "r") as f, \
open(os.path.join(args.save_to, "current_reinforce.yaml"), "w") as current_configs:
INFO("load reinforce configures")
configs = yaml.load(f, Loader=yaml.FullLoader)
yaml.dump(configs, current_configs)
reinforce_configs = configs["reinforce_configs"]
agent_configs = configs["agent_configs"]
rephraser_model_configs = agent_configs["rephraser_model_configs"]
rephraser_optimizer_configs = agent_configs["rephraser_optimizer_configs"]
annunciator_configs = configs["annunciator_configs"]
# the Global variable of USE_GPU is mainly used for environments
GlobalNames.SEED = reinforce_configs["seed"]
GlobalNames.USE_GPU = args.use_gpu
torch.manual_seed(GlobalNames.SEED)
# build vocabulary and data iterator for env
with open(reinforce_configs["victim_configs"], "r") as victim_f:
victim_configs = yaml.load(victim_f, Loader=yaml.FullLoader)
data_configs = victim_configs["data_configs"]
src_vocab = Vocabulary(**data_configs["vocabularies"][0])
trg_vocab = Vocabulary(**data_configs["vocabularies"][1])
data_set = ZipDataset(
TextLineDataset(data_path=data_configs["train_data"][0],
vocabulary=src_vocab, max_len=data_configs["max_len"][0]),
TextLineDataset(data_path=data_configs["train_data"][1],
vocabulary=trg_vocab, max_len=data_configs["max_len"][1]),
shuffle=reinforce_configs["shuffle"]
) # we build the parallel data sets and iterate inside a thread
# collect range of action space:
_, _, limit_dist = load_or_extract_near_vocab(
config_path=reinforce_configs["victim_configs"],
model_path=reinforce_configs["victim_model"],
init_perturb_rate=reinforce_configs["init_perturb_rate"],
save_to=os.path.join(args.save_to, "near_vocab"),
save_to_full=os.path.join(args.save_to, "full_near_vocab"),
top_reserve=12,
emit_as_id=True, use_max_dist=True)
# build global SACAgent for the final policy (on cpu)
global_agent = rephraser.SACAgent(
device="cpu",
d_word_vec=victim_configs["model_configs"]["d_word_vec"],
d_model=rephraser_model_configs["d_model"],
limit_dist=limit_dist,
dropout=rephraser_model_configs["dropout"],
learnable_temperature=rephraser_model_configs["learnable_temperature"],
init_temperature=rephraser_model_configs["init_temperature"],
rephraser_optimizer_configs=rephraser_optimizer_configs,
save_to=args.save_to,
num_kept_checkpoints=reinforce_configs["num_kept_checkpoints"]
)
# load global ckp (only for the AC parameters) if needed
global_step = global_agent.load_model()
print("global_step:", global_step)
if global_step != 0:
INFO("restarting at step %d"%global_step)
else: # save the initial model
global_agent.save_model(global_step)
global_summary_writer = SummaryWriter(
log_dir=os.path.join(args.save_to, "global_summary"))
global_replay_buffer = SharedReplayBuffer(
max_sen_len=data_configs["max_len"][0],
state_dim=victim_configs["model_configs"]["d_word_vec"],
action_dim=victim_configs["model_configs"]["d_word_vec"],
capacity=reinforce_configs["replay_buffer_capacity"])
# test_for_throughput(global_replay_buffer)
# make global objects shared memory
global_agent.share_memory()
global_replay_buffer.share_memory()
# collect available devices and distribute env on the available gpu
if args.use_gpu:
device = "cuda"
devices = []
for i in range(torch.cuda.device_count()):
devices += ["cuda:%d" % i]
print("available gpus:", devices)
else:
device = "cpu"
devices = [device]
# initialize global parameters for the current training trial
global_step_lock = mp.Lock()
global_step_counter = mp.Value("i", global_step) # "i is the type code for c_int"
# train_thread(0, device, args,
# reinforce_configs, annunciator_configs,
# src_vocab, trg_vocab, data_set,
# global_agent, global_replay_buffer,
# global_step_counter, global_step_lock,
# agent_configs)
# valid_thread(device, args,
# reinforce_configs, annunciator_configs,
# src_vocab, trg_vocab, data_set,
# global_agent, global_replay_buffer,
# global_step_counter, global_step_lock,
# agent_configs)
# build multi thread for learning and validation
process = []
for rank in range(args.n):
print("initialize training thread on cuda:%d" % (rank+1))
p=mp.Process(
target=train_thread,
args=(rank, "cuda:%d"%(rank+1), args,
reinforce_configs, annunciator_configs,
src_vocab, trg_vocab, data_set,
global_agent, global_replay_buffer,
global_step_counter, global_step_lock,
agent_configs)
)
p.start()
process.append(p)
# run the dev thread for initiation
print("initialize dev thread on cuda:0")
p = mp.Process(
target=valid_thread,
args=("cuda:0", args,
reinforce_configs, annunciator_configs,
src_vocab, trg_vocab, data_set,
global_agent, global_replay_buffer,
global_step_counter, global_step_lock,
agent_configs)
)
p.start()
process.append(p)
for p in process:
p.join()
def tune(flags):
"""
flags:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# ================================================================================== #
# Initialization for training on different devices
# - CPU/GPU
# - Single/Distributed
Constants.USE_GPU = flags.use_gpu
if flags.multi_gpu:
dist.distributed_init(flags.shared_dir)
world_size = dist.get_world_size()
rank = dist.get_rank()
local_rank = dist.get_local_rank()
else:
world_size = 1
rank = 0
local_rank = 0
if Constants.USE_GPU:
torch.cuda.set_device(local_rank)
Constants.CURRENT_DEVICE = "cuda:{0}".format(local_rank)
else:
Constants.CURRENT_DEVICE = "cpu"
# If not root_rank, close logging
# else write log of training to file.
if rank == 0:
write_log_to_file(
os.path.join(flags.log_path,
"%s.log" % time.strftime("%Y%m%d-%H%M%S")))
else:
close_logging()
# ================================================================================== #
# Parsing configuration files
# - Load default settings
# - Load pre-defined settings
# - Load user-defined settings
configs = prepare_configs(flags.config_path, flags.predefined_config)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
optimizer_configs = configs['optimizer_configs']
training_configs = configs['training_configs']
INFO(pretty_configs(configs))
Constants.SEED = training_configs['seed']
set_seed(Constants.SEED)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary.build_from_file(**data_configs['vocabularies'][0])
vocab_tgt = Vocabulary.build_from_file(**data_configs['vocabularies'][1])
Constants.EOS = vocab_src.eos
Constants.PAD = vocab_src.pad
Constants.BOS = vocab_src.bos
# bt tag dataset
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0],
is_train_dataset=True),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][1],
is_train_dataset=True))
training_iterator = DataIterator(
dataset=train_bitext_dataset,
batch_size=training_configs["batch_size"],
use_bucket=training_configs['use_bucket'],
buffer_size=training_configs['buffer_size'],
batching_func=training_configs['batching_key'],
world_size=world_size,
rank=rank)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
lrate = optimizer_configs['learning_rate']
model_collections = Collections()
checkpoint_saver = Saver(
save_prefix="{0}.ckpt".format(
os.path.join(flags.saveto, flags.model_name)),
num_max_keeping=training_configs['num_kept_checkpoints'])
best_model_prefix = os.path.join(
flags.saveto, flags.model_name + Constants.MY_BEST_MODEL_SUFFIX)
best_model_saver = Saver(
save_prefix=best_model_prefix,
num_max_keeping=training_configs['num_kept_best_model'])
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words,
padding_idx=vocab_src.pad,
vocab_src=vocab_src,
vocab_tgt=vocab_tgt,
**model_configs)
INFO(nmt_model)
critic = NMTCriterion(label_smoothing=model_configs['label_smoothing'],
padding_idx=vocab_tgt.pad)
INFO(critic)
# 2. Move to GPU
if Constants.USE_GPU:
nmt_model = nmt_model.cuda()
critic = critic.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model,
flags.pretrain_path,
exclude_prefix=flags.pretrain_exclude_prefix,
device=Constants.CURRENT_DEVICE)
# froze_parameters
froze_params(nmt_model, flags.froze_config)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# 4. Build optimizer
INFO('Building Optimizer...')
if not flags.multi_gpu:
optim = Optimizer(name=optimizer_configs['optimizer'],
model=nmt_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params'],
update_cycle=training_configs['update_cycle'])
else:
optim = dist.DistributedOptimizer(
name=optimizer_configs['optimizer'],
model=nmt_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params'],
device_id=local_rank)
# 5. Build scheduler for optimizer if needed
scheduler = build_scheduler(
schedule_method=optimizer_configs['schedule_method'],
optimizer=optim,
scheduler_configs=optimizer_configs['scheduler_configs'])
# 6. build moving average
if training_configs['moving_average_method'] is not None:
ma = MovingAverage(
moving_average_method=training_configs['moving_average_method'],
named_params=nmt_model.named_parameters(),
alpha=training_configs['moving_average_alpha'])
else:
ma = None
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# Reload from latest checkpoint
if flags.reload:
checkpoint_saver.load_latest(model=nmt_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma,
device=Constants.CURRENT_DEVICE)
# broadcast parameters and optimizer states
if world_size > 1:
INFO("Broadcasting model parameters...")
dist.broadcast_parameters(params=nmt_model.state_dict())
INFO("Broadcasting optimizer states...")
dist.broadcast_optimizer_state(optimizer=optim.optim)
INFO('Done.')
# ================================================================================== #
# Prepare training
eidx = model_collections.get_collection("eidx", [0])[-1]
uidx = model_collections.get_collection("uidx", [1])[-1]
bad_count = model_collections.get_collection("bad_count", [0])[-1]
oom_count = model_collections.get_collection("oom_count", [0])[-1]
is_early_stop = model_collections.get_collection("is_early_stop", [
False,
])[-1]
train_loss_meter = AverageMeter()
sent_per_sec_meter = TimeMeter()
tok_per_sec_meter = TimeMeter()
update_cycle = training_configs['update_cycle']
grad_denom = 0
train_loss = 0.0
cum_n_words = 0
valid_loss = best_valid_loss = float('inf')
if rank == 0:
summary_writer = SummaryWriter(log_dir=flags.log_path)
else:
summary_writer = None
sent_per_sec_meter.start()
tok_per_sec_meter.start()
INFO('Begin training...')
while True:
if summary_writer is not None:
summary_writer.add_scalar("Epoch", (eidx + 1), uidx)
# Build iterator and progress bar
training_iter = training_iterator.build_generator()
if rank == 0:
training_progress_bar = tqdm(desc=' - (Epc {}, Upd {}) '.format(
eidx, uidx),
total=len(training_iterator),
unit="sents")
else:
training_progress_bar = None
# INFO(Constants.USE_BT)
for batch in training_iter:
# bt attrib data
seqs_x, seqs_y = batch
batch_size = len(seqs_x)
cum_n_words += sum(len(s) for s in seqs_y)
try:
# Prepare data
x, y = prepare_data(seqs_x, seqs_y, cuda=Constants.USE_GPU)
loss = compute_forward(
model=nmt_model,
critic=critic,
seqs_x=x,
seqs_y=y,
eval=False,
normalization=1.0,
norm_by_words=training_configs["norm_by_words"])
update_cycle -= 1
grad_denom += batch_size
train_loss += loss
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
oom_count += 1
else:
raise e
# When update_cycle becomes 0, it means end of one batch. Several things will be done:
# - update parameters
# - reset update_cycle and grad_denom, update uidx
# - learning rate scheduling
# - update moving average
if update_cycle == 0:
# 0. reduce variables
if world_size > 1:
grad_denom = dist.all_reduce_py(grad_denom)
train_loss = dist.all_reduce_py(train_loss)
cum_n_words = dist.all_reduce_py(cum_n_words)
# 1. update parameters
optim.step(denom=grad_denom)
optim.zero_grad()
if training_progress_bar is not None:
training_progress_bar.update(grad_denom)
training_progress_bar.set_description(
' - (Epc {}, Upd {}) '.format(eidx, uidx))
postfix_str = 'TrainLoss: {:.2f}, ValidLoss(best): {:.2f} ({:.2f}), '.format(
train_loss, valid_loss, best_valid_loss)
training_progress_bar.set_postfix_str(postfix_str)
# 2. learning rate scheduling
if scheduler is not None and optimizer_configs[
"schedule_method"] != "loss":
scheduler.step(global_step=uidx)
# 3. update moving average
if ma is not None and eidx >= training_configs[
'moving_average_start_epoch']:
ma.step()
# 4. update meters
train_loss_meter.update(train_loss, grad_denom)
sent_per_sec_meter.update(grad_denom)
tok_per_sec_meter.update(cum_n_words)
# 5. reset accumulated variables, update uidx
update_cycle = training_configs['update_cycle']
grad_denom = 0
uidx += 1
cum_n_words = 0.0
train_loss = 0.0
else:
continue
# ================================================================================== #
# Display some information
if should_trigger_by_steps(
uidx, eidx, every_n_step=training_configs['disp_freq']):
lrate = list(optim.get_lrate())[0]
if summary_writer is not None:
summary_writer.add_scalar(
"Speed(sents/sec)",
scalar_value=sent_per_sec_meter.ave,
global_step=uidx)
summary_writer.add_scalar(
"Speed(words/sec)",
scalar_value=tok_per_sec_meter.ave,
global_step=uidx)
summary_writer.add_scalar(
"train_loss",
scalar_value=train_loss_meter.ave,
global_step=uidx)
summary_writer.add_scalar("lrate",
scalar_value=lrate,
global_step=uidx)
summary_writer.add_scalar("oom_count",
scalar_value=oom_count,
global_step=uidx)
# Reset Meters
sent_per_sec_meter.reset()
tok_per_sec_meter.reset()
train_loss_meter.reset()
# ================================================================================== #
# Saving checkpoints
# if should_trigger_by_steps(uidx, eidx, every_n_step=training_configs['save_freq'], debug=flags.debug):
# model_collections.add_to_collection("uidx", uidx)
# model_collections.add_to_collection("eidx", eidx)
# model_collections.add_to_collection("bad_count", bad_count)
#
# if not is_early_stop:
# if rank == 0:
# checkpoint_saver.save(global_step=uidx,
# model=nmt_model,
# optim=optim,
# lr_scheduler=scheduler,
# collections=model_collections,
# ma=ma)
torch.save(nmt_model.state_dict(), best_model_prefix + ".final")
if training_progress_bar is not None:
training_progress_bar.close()
eidx += 1
if eidx > training_configs["max_epochs"]:
break
def train(FLAGS):
"""
FLAGS:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# ================================================================================== #
# Initialization for training on different devices
# - CPU/GPU
# - Single/Distributed
GlobalNames.USE_GPU = FLAGS.use_gpu
if FLAGS.multi_gpu:
if hvd is None or distributed is None:
ERROR("Distributed training is disable. Please check the installation of Horovod.")
hvd.init()
world_size = hvd.size()
rank = hvd.rank()
local_rank = hvd.local_rank()
else:
world_size = 1
rank = 0
local_rank = 0
if GlobalNames.USE_GPU:
torch.cuda.set_device(local_rank)
CURRENT_DEVICE = "cuda:{0}".format(local_rank)
else:
CURRENT_DEVICE = "cpu"
# If not root_rank, close logging
if rank != 0:
close_logging()
# write log of training to file.
if rank == 0:
write_log_to_file(os.path.join(FLAGS.log_path, "%s.log" % time.strftime("%Y%m%d-%H%M%S")))
# ================================================================================== #
# Parsing configuration files
config_path = os.path.abspath(FLAGS.config_path)
with open(config_path.strip()) as f:
configs = yaml.load(f)
INFO(pretty_configs(configs))
# Add default configs
configs = default_baseline_configs(configs)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
optimizer_configs = configs['optimizer_configs']
training_configs = configs['training_configs']
GlobalNames.SEED = training_configs['seed']
set_seed(GlobalNames.SEED)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary(**data_configs["vocabularies"][0])
vocab_tgt = Vocabulary(**data_configs["vocabularies"][1])
actual_buffer_size = training_configs["buffer_size"] * max(1, training_configs["update_cycle"])
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0],
),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][1],
)
)
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['valid_data'][0],
vocabulary=vocab_src,
),
TextLineDataset(data_path=data_configs['valid_data'][1],
vocabulary=vocab_tgt,
)
)
training_iterator = DataIterator(dataset=train_bitext_dataset,
batch_size=training_configs["batch_size"],
use_bucket=training_configs['use_bucket'],
buffer_size=actual_buffer_size,
batching_func=training_configs['batching_key'],
world_size=world_size,
rank=rank)
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=training_configs['valid_batch_size'],
use_bucket=True, buffer_size=100000, numbering=True,
world_size=world_size, rank=rank)
bleu_scorer = SacreBLEUScorer(reference_path=data_configs["bleu_valid_reference"],
num_refs=data_configs["num_refs"],
lang_pair=data_configs["lang_pair"],
sacrebleu_args=training_configs["bleu_valid_configs"]['sacrebleu_args'],
postprocess=training_configs["bleu_valid_configs"]['postprocess']
)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
lrate = optimizer_configs['learning_rate']
is_early_stop = False
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
model_collections = Collections()
best_model_prefix = os.path.join(FLAGS.saveto, FLAGS.model_name + GlobalNames.MY_BEST_MODEL_SUFFIX)
checkpoint_saver = Saver(save_prefix="{0}.ckpt".format(os.path.join(FLAGS.saveto, FLAGS.model_name)),
num_max_keeping=training_configs['num_kept_checkpoints']
)
best_model_saver = Saver(save_prefix=best_model_prefix, num_max_keeping=training_configs['num_kept_best_model'])
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words, **model_configs)
INFO(nmt_model)
critic = NMTCriterion(label_smoothing=model_configs['label_smoothing'])
INFO(critic)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# 2. Move to GPU
if GlobalNames.USE_GPU:
nmt_model = nmt_model.cuda()
critic = critic.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model, FLAGS.pretrain_path, exclude_prefix=None, device=CURRENT_DEVICE)
# 4. Build optimizer
INFO('Building Optimizer...')
optim = Optimizer(name=optimizer_configs['optimizer'],
model=nmt_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params'],
distributed=True if world_size > 1 else False,
update_cycle=training_configs['update_cycle']
)
# 5. Build scheduler for optimizer if needed
if optimizer_configs['schedule_method'] is not None:
if optimizer_configs['schedule_method'] == "loss":
scheduler = ReduceOnPlateauScheduler(optimizer=optim,
**optimizer_configs["scheduler_configs"]
)
elif optimizer_configs['schedule_method'] == "noam":
scheduler = NoamScheduler(optimizer=optim, **optimizer_configs['scheduler_configs'])
else:
WARN("Unknown scheduler name {0}. Do not use lr_scheduling.".format(optimizer_configs['schedule_method']))
scheduler = None
else:
scheduler = None
# 6. build moving average
if training_configs['moving_average_method'] is not None:
ma = MovingAverage(moving_average_method=training_configs['moving_average_method'],
named_params=nmt_model.named_parameters(),
alpha=training_configs['moving_average_alpha'])
else:
ma = None
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# Reload from latest checkpoint
if FLAGS.reload:
checkpoint_saver.load_latest(model=nmt_model, optim=optim, lr_scheduler=scheduler,
collections=model_collections, ma=ma)
# broadcast parameters and optimizer states
if world_size > 1:
hvd.broadcast_parameters(params=nmt_model.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(optimizer=optim.optim, root_rank=0)
# ================================================================================== #
# Prepare training
eidx = model_collections.get_collection("eidx", [0])[-1]
uidx = model_collections.get_collection("uidx", [1])[-1]
bad_count = model_collections.get_collection("bad_count", [0])[-1]
oom_count = model_collections.get_collection("oom_count", [0])[-1]
cum_n_samples = 0
cum_n_words = 0
best_valid_loss = 1.0 * 1e10 # Max Float
update_cycle = training_configs['update_cycle']
grad_denom = 0
if rank == 0:
summary_writer = SummaryWriter(log_dir=FLAGS.log_path)
else:
summary_writer = None
# Timer for computing speed
timer_for_speed = Timer()
timer_for_speed.tic()
INFO('Begin training...')
while True:
if summary_writer is not None:
summary_writer.add_scalar("Epoch", (eidx + 1), uidx)
# Build iterator and progress bar
training_iter = training_iterator.build_generator()
if rank == 0:
training_progress_bar = tqdm(desc=' - (Epoch %d) ' % eidx,
total=len(training_iterator),
unit="sents"
)
else:
training_progress_bar = None
for batch in training_iter:
seqs_x, seqs_y = batch
batch_size = len(seqs_x)
cum_n_samples += batch_size
cum_n_words += sum(len(s) for s in seqs_y)
try:
# Prepare data
x, y = prepare_data(seqs_x, seqs_y, cuda=GlobalNames.USE_GPU)
loss = compute_forward(model=nmt_model,
critic=critic,
seqs_x=x,
seqs_y=y,
eval=False,
normalization=1.0,
norm_by_words=training_configs["norm_by_words"])
update_cycle -= 1
grad_denom += batch_size
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
oom_count += 1
else:
raise e
# When update_cycle becomes 0, it means end of one batch. Several things will be done:
# - update parameters
# - reset update_cycle and grad_denom
# - update uidx
# - update moving average
if update_cycle == 0:
if world_size > 1:
grad_denom = distributed.all_reduce(grad_denom)
optim.step(denom=grad_denom)
optim.zero_grad()
if training_progress_bar is not None:
training_progress_bar.update(grad_denom)
update_cycle = training_configs['update_cycle']
grad_denom = 0
uidx += 1
if scheduler is None:
pass
elif optimizer_configs["schedule_method"] == "loss":
scheduler.step(metric=best_valid_loss)
else:
scheduler.step(global_step=uidx)
if ma is not None and eidx >= training_configs['moving_average_start_epoch']:
ma.step()
else:
continue
# ================================================================================== #
# Display some information
if should_trigger_by_steps(uidx, eidx, every_n_step=training_configs['disp_freq']):
if world_size > 1:
cum_n_words = sum(distributed.all_gather(cum_n_words))
cum_n_samples = sum(distributed.all_gather(cum_n_samples))
# words per second and sents per second
words_per_sec = cum_n_words / (timer.toc(return_seconds=True))
sents_per_sec = cum_n_samples / (timer.toc(return_seconds=True))
lrate = list(optim.get_lrate())[0]
if summary_writer is not None:
summary_writer.add_scalar("Speed(words/sec)", scalar_value=words_per_sec, global_step=uidx)
summary_writer.add_scalar("Speed(sents/sen)", scalar_value=sents_per_sec, global_step=uidx)
summary_writer.add_scalar("lrate", scalar_value=lrate, global_step=uidx)
summary_writer.add_scalar("oom_count", scalar_value=oom_count, global_step=uidx)
# Reset timer
timer.tic()
cum_n_words = 0
cum_n_samples = 0
# ================================================================================== #
# Loss Validation & Learning rate annealing
if should_trigger_by_steps(global_step=uidx, n_epoch=eidx, every_n_step=training_configs['loss_valid_freq'],
debug=FLAGS.debug):
valid_loss = loss_validation(model=nmt_model,
critic=critic,
valid_iterator=valid_iterator,
rank=rank,
world_size=world_size
)
model_collections.add_to_collection("history_losses", valid_loss)
min_history_loss = np.array(model_collections.get_collection("history_losses")).min()
best_valid_loss = min_history_loss
if summary_writer is not None:
summary_writer.add_scalar("loss", valid_loss, global_step=uidx)
summary_writer.add_scalar("best_loss", min_history_loss, global_step=uidx)
# ================================================================================== #
# BLEU Validation & Early Stop
if should_trigger_by_steps(global_step=uidx, n_epoch=eidx,
every_n_step=training_configs['bleu_valid_freq'],
min_step=training_configs['bleu_valid_warmup'],
debug=FLAGS.debug):
valid_bleu = bleu_validation(uidx=uidx,
valid_iterator=valid_iterator,
batch_size=training_configs["bleu_valid_batch_size"],
model=nmt_model,
bleu_scorer=bleu_scorer,
vocab_tgt=vocab_tgt,
valid_dir=FLAGS.valid_path,
max_steps=training_configs["bleu_valid_configs"]["max_steps"],
beam_size=training_configs["bleu_valid_configs"]["beam_size"],
alpha=training_configs["bleu_valid_configs"]["alpha"],
world_size=world_size,
rank=rank,
)
model_collections.add_to_collection(key="history_bleus", value=valid_bleu)
best_valid_bleu = float(np.array(model_collections.get_collection("history_bleus")).max())
if summary_writer is not None:
summary_writer.add_scalar("bleu", valid_bleu, uidx)
summary_writer.add_scalar("best_bleu", best_valid_bleu, uidx)
# If model get new best valid bleu score
if valid_bleu >= best_valid_bleu:
bad_count = 0
if is_early_stop is False:
if rank == 0:
# 1. save the best model
torch.save(nmt_model.state_dict(), best_model_prefix + ".final")
# 2. record all several best models
best_model_saver.save(global_step=uidx, model=nmt_model, ma=ma)
else:
bad_count += 1
# At least one epoch should be traversed
if bad_count >= training_configs['early_stop_patience'] and eidx > 0:
is_early_stop = True
WARN("Early Stop!")
if summary_writer is not None:
summary_writer.add_scalar("bad_count", bad_count, uidx)
INFO("{0} Loss: {1:.2f} BLEU: {2:.2f} lrate: {3:6f} patience: {4}".format(
uidx, valid_loss, valid_bleu, lrate, bad_count
))
# ================================================================================== #
# Saving checkpoints
if should_trigger_by_steps(uidx, eidx, every_n_step=training_configs['save_freq'], debug=FLAGS.debug):
model_collections.add_to_collection("uidx", uidx)
model_collections.add_to_collection("eidx", eidx)
model_collections.add_to_collection("bad_count", bad_count)
if not is_early_stop:
if rank == 0:
checkpoint_saver.save(global_step=uidx,
model=nmt_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma)
if training_progress_bar is not None:
training_progress_bar.close()
eidx += 1
if eidx > training_configs["max_epochs"]:
break
def train(FLAGS):
"""
FLAGS:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# write log of training to file.
write_log_to_file(os.path.join(FLAGS.log_path, "%s.log" % time.strftime("%Y%m%d-%H%M%S")))
GlobalNames.USE_GPU = FLAGS.use_gpu
if GlobalNames.USE_GPU:
CURRENT_DEVICE = "cpu"
else:
CURRENT_DEVICE = "cuda:0"
config_path = os.path.abspath(FLAGS.config_path)
with open(config_path.strip()) as f:
configs = yaml.load(f)
INFO(pretty_configs(configs))
# Add default configs
configs = default_configs(configs)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
optimizer_configs = configs['optimizer_configs']
training_configs = configs['training_configs']
GlobalNames.SEED = training_configs['seed']
set_seed(GlobalNames.SEED)
best_model_prefix = os.path.join(FLAGS.saveto, FLAGS.model_name + GlobalNames.MY_BEST_MODEL_SUFFIX)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary(**data_configs["vocabularies"][0])
vocab_tgt = Vocabulary(**data_configs["vocabularies"][1])
train_batch_size = training_configs["batch_size"] * max(1, training_configs["update_cycle"])
train_buffer_size = training_configs["buffer_size"] * max(1, training_configs["update_cycle"])
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0],
),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][1],
),
shuffle=training_configs['shuffle']
)
valid_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['valid_data'][0],
vocabulary=vocab_src,
),
TextLineDataset(data_path=data_configs['valid_data'][1],
vocabulary=vocab_tgt,
)
)
training_iterator = DataIterator(dataset=train_bitext_dataset,
batch_size=train_batch_size,
use_bucket=training_configs['use_bucket'],
buffer_size=train_buffer_size,
batching_func=training_configs['batching_key'])
valid_iterator = DataIterator(dataset=valid_bitext_dataset,
batch_size=training_configs['valid_batch_size'],
use_bucket=True, buffer_size=100000, numbering=True)
bleu_scorer = SacreBLEUScorer(reference_path=data_configs["bleu_valid_reference"],
num_refs=data_configs["num_refs"],
lang_pair=data_configs["lang_pair"],
sacrebleu_args=training_configs["bleu_valid_configs"]['sacrebleu_args'],
postprocess=training_configs["bleu_valid_configs"]['postprocess']
)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
lrate = optimizer_configs['learning_rate']
is_early_stop = False
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
model_collections = Collections()
checkpoint_saver = Saver(save_prefix="{0}.ckpt".format(os.path.join(FLAGS.saveto, FLAGS.model_name)),
num_max_keeping=training_configs['num_kept_checkpoints']
)
best_model_saver = Saver(save_prefix=best_model_prefix, num_max_keeping=training_configs['num_kept_best_model'])
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words, **model_configs)
INFO(nmt_model)
critic = NMTCriterion(label_smoothing=model_configs['label_smoothing'])
INFO(critic)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# 2. Move to GPU
if GlobalNames.USE_GPU:
nmt_model = nmt_model.cuda()
critic = critic.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model, FLAGS.pretrain_path, exclude_prefix=None, device=CURRENT_DEVICE)
# 4. Build optimizer
INFO('Building Optimizer...')
optim = Optimizer(name=optimizer_configs['optimizer'],
model=nmt_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params']
)
# 5. Build scheduler for optimizer if needed
if optimizer_configs['schedule_method'] is not None:
if optimizer_configs['schedule_method'] == "loss":
scheduler = ReduceOnPlateauScheduler(optimizer=optim,
**optimizer_configs["scheduler_configs"]
)
elif optimizer_configs['schedule_method'] == "noam":
scheduler = NoamScheduler(optimizer=optim, **optimizer_configs['scheduler_configs'])
else:
WARN("Unknown scheduler name {0}. Do not use lr_scheduling.".format(optimizer_configs['schedule_method']))
scheduler = None
else:
scheduler = None
# 6. build EMA
if training_configs['ema_decay'] > 0.0:
ema = ExponentialMovingAverage(named_params=nmt_model.named_parameters(), decay=training_configs['ema_decay'])
else:
ema = None
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# Reload from latest checkpoint
if FLAGS.reload:
checkpoint_saver.load_latest(model=nmt_model, optim=optim, lr_scheduler=scheduler,
collections=model_collections)
# ================================================================================== #
# Prepare training
eidx = model_collections.get_collection("eidx", [0])[-1]
uidx = model_collections.get_collection("uidx", [0])[-1]
bad_count = model_collections.get_collection("bad_count", [0])[-1]
summary_writer = SummaryWriter(log_dir=FLAGS.log_path)
cum_samples = 0
cum_words = 0
best_valid_loss = 1.0 * 1e10 # Max Float
saving_files = []
# Timer for computing speed
timer_for_speed = Timer()
timer_for_speed.tic()
INFO('Begin training...')
while True:
summary_writer.add_scalar("Epoch", (eidx + 1), uidx)
# Build iterator and progress bar
training_iter = training_iterator.build_generator()
training_progress_bar = tqdm(desc=' - (Epoch %d) ' % eidx,
total=len(training_iterator),
unit="sents"
)
for batch in training_iter:
uidx += 1
if scheduler is None:
pass
elif optimizer_configs["schedule_method"] == "loss":
scheduler.step(metric=best_valid_loss)
else:
scheduler.step(global_step=uidx)
seqs_x, seqs_y = batch
n_samples_t = len(seqs_x)
n_words_t = sum(len(s) for s in seqs_y)
cum_samples += n_samples_t
cum_words += n_words_t
training_progress_bar.update(n_samples_t)
optim.zero_grad()
# Prepare data
for seqs_x_t, seqs_y_t in split_shard(seqs_x, seqs_y, split_size=training_configs['update_cycle']):
x, y = prepare_data(seqs_x_t, seqs_y_t, cuda=GlobalNames.USE_GPU)
loss = compute_forward(model=nmt_model,
critic=critic,
seqs_x=x,
seqs_y=y,
eval=False,
normalization=n_samples_t,
norm_by_words=training_configs["norm_by_words"])
optim.step()
if ema is not None:
ema.step()
# ================================================================================== #
# Display some information
if should_trigger_by_steps(uidx, eidx, every_n_step=training_configs['disp_freq']):
# words per second and sents per second
words_per_sec = cum_words / (timer.toc(return_seconds=True))
sents_per_sec = cum_samples / (timer.toc(return_seconds=True))
lrate = list(optim.get_lrate())[0]
summary_writer.add_scalar("Speed(words/sec)", scalar_value=words_per_sec, global_step=uidx)
summary_writer.add_scalar("Speed(sents/sen)", scalar_value=sents_per_sec, global_step=uidx)
summary_writer.add_scalar("lrate", scalar_value=lrate, global_step=uidx)
# Reset timer
timer.tic()
cum_words = 0
cum_samples = 0
# ================================================================================== #
# Saving checkpoints
if should_trigger_by_steps(uidx, eidx, every_n_step=training_configs['save_freq'], debug=FLAGS.debug):
model_collections.add_to_collection("uidx", uidx)
model_collections.add_to_collection("eidx", eidx)
model_collections.add_to_collection("bad_count", bad_count)
if not is_early_stop:
checkpoint_saver.save(global_step=uidx, model=nmt_model, optim=optim, lr_scheduler=scheduler,
collections=model_collections, ema=ema)
# ================================================================================== #
# Loss Validation & Learning rate annealing
if should_trigger_by_steps(global_step=uidx, n_epoch=eidx, every_n_step=training_configs['loss_valid_freq'],
debug=FLAGS.debug):
if ema is not None:
origin_state_dict = deepcopy(nmt_model.state_dict())
nmt_model.load_state_dict(ema.state_dict(), strict=False)
valid_loss = loss_validation(model=nmt_model,
critic=critic,
valid_iterator=valid_iterator,
)
model_collections.add_to_collection("history_losses", valid_loss)
min_history_loss = np.array(model_collections.get_collection("history_losses")).min()
summary_writer.add_scalar("loss", valid_loss, global_step=uidx)
summary_writer.add_scalar("best_loss", min_history_loss, global_step=uidx)
best_valid_loss = min_history_loss
if ema is not None:
nmt_model.load_state_dict(origin_state_dict)
del origin_state_dict
# ================================================================================== #
# BLEU Validation & Early Stop
if should_trigger_by_steps(global_step=uidx, n_epoch=eidx,
every_n_step=training_configs['bleu_valid_freq'],
min_step=training_configs['bleu_valid_warmup'],
debug=FLAGS.debug):
if ema is not None:
origin_state_dict = deepcopy(nmt_model.state_dict())
nmt_model.load_state_dict(ema.state_dict(), strict=False)
valid_bleu = bleu_validation(uidx=uidx,
valid_iterator=valid_iterator,
batch_size=training_configs["bleu_valid_batch_size"],
model=nmt_model,
bleu_scorer=bleu_scorer,
vocab_tgt=vocab_tgt,
valid_dir=FLAGS.valid_path,
max_steps=training_configs["bleu_valid_configs"]["max_steps"],
beam_size=training_configs["bleu_valid_configs"]["beam_size"],
alpha=training_configs["bleu_valid_configs"]["alpha"]
)
model_collections.add_to_collection(key="history_bleus", value=valid_bleu)
best_valid_bleu = float(np.array(model_collections.get_collection("history_bleus")).max())
summary_writer.add_scalar("bleu", valid_bleu, uidx)
summary_writer.add_scalar("best_bleu", best_valid_bleu, uidx)
# If model get new best valid bleu score
if valid_bleu >= best_valid_bleu:
bad_count = 0
if is_early_stop is False:
# 1. save the best model
torch.save(nmt_model.state_dict(), best_model_prefix + ".final")
# 2. record all several best models
best_model_saver.save(global_step=uidx, model=nmt_model)
else:
bad_count += 1
# At least one epoch should be traversed
if bad_count >= training_configs['early_stop_patience'] and eidx > 0:
is_early_stop = True
WARN("Early Stop!")
summary_writer.add_scalar("bad_count", bad_count, uidx)
if ema is not None:
nmt_model.load_state_dict(origin_state_dict)
del origin_state_dict
INFO("{0} Loss: {1:.2f} BLEU: {2:.2f} lrate: {3:6f} patience: {4}".format(
uidx, valid_loss, valid_bleu, lrate, bad_count
))
training_progress_bar.close()
eidx += 1
if eidx > training_configs["max_epochs"]:
break
def train(flags):
"""
flags:
saveto: str
reload: store_true
config_path: str
pretrain_path: str, default=""
model_name: str
log_path: str
"""
# ================================================================================== #
# Initialization for training on different devices
# - CPU/GPU
# - Single/Distributed
Constants.USE_GPU = flags.use_gpu
if flags.multi_gpu:
dist.distributed_init(flags.shared_dir)
world_size = dist.get_world_size()
rank = dist.get_rank()
local_rank = dist.get_local_rank()
else:
world_size = 1
rank = 0
local_rank = 0
if Constants.USE_GPU:
torch.cuda.set_device(local_rank)
Constants.CURRENT_DEVICE = "cuda:{0}".format(local_rank)
else:
Constants.CURRENT_DEVICE = "cpu"
# If not root_rank, close logging
# else write log of training to file.
if rank == 0:
write_log_to_file(
os.path.join(flags.log_path,
"%s.log" % time.strftime("%Y%m%d-%H%M%S")))
else:
close_logging()
# ================================================================================== #
# Parsing configuration files
# - Load default settings
# - Load pre-defined settings
# - Load user-defined settings
configs = prepare_configs(flags.config_path, flags.predefined_config)
data_configs = configs['data_configs']
model_configs = configs['model_configs']
optimizer_configs = configs['optimizer_configs']
training_configs = configs['training_configs']
INFO(pretty_configs(configs))
# use odc
if training_configs['use_odc'] is True:
ave_best_k = check_odc_config(training_configs)
else:
ave_best_k = 0
Constants.SEED = training_configs['seed']
set_seed(Constants.SEED)
timer = Timer()
# ================================================================================== #
# Load Data
INFO('Loading data...')
timer.tic()
# Generate target dictionary
vocab_src = Vocabulary.build_from_file(**data_configs['vocabularies'][0])
vocab_tgt = Vocabulary.build_from_file(**data_configs['vocabularies'][1])
Constants.EOS = vocab_src.eos
Constants.PAD = vocab_src.pad
Constants.BOS = vocab_src.bos
train_bitext_dataset = ZipDataset(
TextLineDataset(data_path=data_configs['train_data'][0],
vocabulary=vocab_src,
max_len=data_configs['max_len'][0],
is_train_dataset=True),
TextLineDataset(data_path=data_configs['train_data'][1],
vocabulary=vocab_tgt,
max_len=data_configs['max_len'][1],
is_train_dataset=True))
valid_bitext_dataset = ZipDataset(
TextLineDataset(
data_path=data_configs['valid_data'][0],
vocabulary=vocab_src,
is_train_dataset=False,
),
TextLineDataset(data_path=data_configs['valid_data'][1],
vocabulary=vocab_tgt,
is_train_dataset=False))
training_iterator = DataIterator(
dataset=train_bitext_dataset,
batch_size=training_configs["batch_size"],
use_bucket=training_configs['use_bucket'],
buffer_size=training_configs['buffer_size'],
batching_func=training_configs['batching_key'],
world_size=world_size,
rank=rank)
valid_iterator = DataIterator(
dataset=valid_bitext_dataset,
batch_size=training_configs['valid_batch_size'],
use_bucket=True,
buffer_size=100000,
numbering=True,
world_size=world_size,
rank=rank)
bleu_scorer = SacreBLEUScorer(
reference_path=data_configs["bleu_valid_reference"],
num_refs=data_configs["num_refs"],
lang_pair=data_configs["lang_pair"],
sacrebleu_args=training_configs["bleu_valid_configs"]
['sacrebleu_args'],
postprocess=training_configs["bleu_valid_configs"]['postprocess'])
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# ================================ Begin ======================================== #
# Build Model & Optimizer
# We would do steps below on after another
# 1. build models & criterion
# 2. move models & criterion to gpu if needed
# 3. load pre-trained model if needed
# 4. build optimizer
# 5. build learning rate scheduler if needed
# 6. load checkpoints if needed
# 0. Initial
lrate = optimizer_configs['learning_rate']
model_collections = Collections()
checkpoint_saver = Saver(
save_prefix="{0}.ckpt".format(
os.path.join(flags.saveto, flags.model_name)),
num_max_keeping=training_configs['num_kept_checkpoints'])
best_model_prefix = os.path.join(
flags.saveto, flags.model_name + Constants.MY_BEST_MODEL_SUFFIX)
best_k_saver = BestKSaver(
save_prefix="{0}.best_k_ckpt".format(
os.path.join(flags.saveto, flags.model_name)),
num_max_keeping=training_configs['num_kept_best_k_checkpoints'])
# 1. Build Model & Criterion
INFO('Building model...')
timer.tic()
nmt_model = build_model(n_src_vocab=vocab_src.max_n_words,
n_tgt_vocab=vocab_tgt.max_n_words,
padding_idx=vocab_src.pad,
vocab_src=vocab_src,
**model_configs)
INFO(nmt_model)
# build teacher model
teacher_model, teacher_model_path = get_teacher_model(
training_configs, model_configs, vocab_src, vocab_tgt, flags)
# build critic
critic = CombinationCriterion(model_configs['loss_configs'],
padding_idx=vocab_tgt.pad,
teacher=teacher_model)
# INFO(critic)
critic.INFO()
# 2. Move to GPU
if Constants.USE_GPU:
nmt_model = nmt_model.cuda()
critic = critic.cuda()
# 3. Load pretrained model if needed
load_pretrained_model(nmt_model,
flags.pretrain_path,
exclude_prefix=None,
device=Constants.CURRENT_DEVICE)
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# 4. Build optimizer
INFO('Building Optimizer...')
if not flags.multi_gpu:
optim = Optimizer(name=optimizer_configs['optimizer'],
model=nmt_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params'],
update_cycle=training_configs['update_cycle'])
else:
optim = dist.DistributedOptimizer(
name=optimizer_configs['optimizer'],
model=nmt_model,
lr=lrate,
grad_clip=optimizer_configs['grad_clip'],
optim_args=optimizer_configs['optimizer_params'],
device_id=local_rank)
# 5. Build scheduler for optimizer if needed
scheduler = build_scheduler(
schedule_method=optimizer_configs['schedule_method'],
optimizer=optim,
scheduler_configs=optimizer_configs['scheduler_configs'])
# 6. build moving average
ma = build_ma(training_configs, nmt_model.named_parameters())
INFO('Done. Elapsed time {0}'.format(timer.toc()))
# Reload from latest checkpoint
if flags.reload:
checkpoint_saver.load_latest(model=nmt_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma,
device=Constants.CURRENT_DEVICE)
# broadcast parameters and optimizer states
if world_size > 1:
INFO("Broadcasting model parameters...")
dist.broadcast_parameters(params=nmt_model.state_dict())
INFO("Broadcasting optimizer states...")
dist.broadcast_optimizer_state(optimizer=optim.optim)
INFO('Done.')
# ================================================================================== #
# Prepare training
eidx = model_collections.get_collection("eidx", [0])[-1]
uidx = model_collections.get_collection("uidx", [1])[-1]
bad_count = model_collections.get_collection("bad_count", [0])[-1]
oom_count = model_collections.get_collection("oom_count", [0])[-1]
is_early_stop = model_collections.get_collection("is_early_stop", [
False,
])[-1]
teacher_patience = model_collections.get_collection(
"teacher_patience", [training_configs['teacher_patience']])[-1]
train_loss_meter = AverageMeter()
train_loss_dict_meter = AverageMeterDict(critic.get_critic_name())
sent_per_sec_meter = TimeMeter()
tok_per_sec_meter = TimeMeter()
update_cycle = training_configs['update_cycle']
grad_denom = 0
train_loss = 0.0
cum_n_words = 0
train_loss_dict = dict()
valid_loss = best_valid_loss = float('inf')
if rank == 0:
summary_writer = SummaryWriter(log_dir=flags.log_path)
else:
summary_writer = None
sent_per_sec_meter.start()
tok_per_sec_meter.start()
INFO('Begin training...')
while True:
if summary_writer is not None:
summary_writer.add_scalar("Epoch", (eidx + 1), uidx)
# Build iterator and progress bar
training_iter = training_iterator.build_generator()
if rank == 0:
training_progress_bar = tqdm(desc=' - (Epc {}, Upd {}) '.format(
eidx, uidx),
total=len(training_iterator),
unit="sents")
else:
training_progress_bar = None
for batch in training_iter:
seqs_x, seqs_y = batch
batch_size = len(seqs_x)
cum_n_words += sum(len(s) for s in seqs_y)
try:
# Prepare data
x, y = prepare_data(seqs_x, seqs_y, cuda=Constants.USE_GPU)
loss, loss_dict = compute_forward(
model=nmt_model,
critic=critic,
seqs_x=x,
seqs_y=y,
eval=False,
normalization=1.0,
norm_by_words=training_configs["norm_by_words"])
update_cycle -= 1
grad_denom += batch_size
train_loss += loss
train_loss_dict = add_dict_value(train_loss_dict, loss_dict)
except RuntimeError as e:
if 'out of memory' in str(e):
print('| WARNING: ran out of memory, skipping batch')
oom_count += 1
else:
raise e
# When update_cycle becomes 0, it means end of one batch. Several things will be done:
# - update parameters
# - reset update_cycle and grad_denom, update uidx
# - learning rate scheduling
# - update moving average
if update_cycle == 0:
# 0. reduce variables
if world_size > 1:
grad_denom = dist.all_reduce_py(grad_denom)
train_loss = dist.all_reduce_py(train_loss)
train_loss_dict = dist.all_reduce_py(train_loss_dict)
cum_n_words = dist.all_reduce_py(cum_n_words)
# 1. update parameters
optim.step(denom=grad_denom)
optim.zero_grad()
if training_progress_bar is not None:
training_progress_bar.update(grad_denom)
training_progress_bar.set_description(
' - (Epc {}, Upd {}) '.format(eidx, uidx))
postfix_str = 'TrainLoss: {:.2f}, ValidLoss(best): {:.2f} ({:.2f}), '.format(
train_loss, valid_loss, best_valid_loss)
for critic_name, loss_value in train_loss_dict.items():
postfix_str += (critic_name +
': {:.2f}, ').format(loss_value)
training_progress_bar.set_postfix_str(postfix_str)
# 2. learning rate scheduling
if scheduler is not None and optimizer_configs[
"schedule_method"] != "loss":
scheduler.step(global_step=uidx)
# 3. update moving average
if ma is not None and eidx >= training_configs[
'moving_average_start_epoch']:
ma.step()
# 4. update meters
train_loss_meter.update(train_loss, grad_denom)
train_loss_dict_meter.update(train_loss_dict, grad_denom)
sent_per_sec_meter.update(grad_denom)
tok_per_sec_meter.update(cum_n_words)
# 5. reset accumulated variables, update uidx
update_cycle = training_configs['update_cycle']
grad_denom = 0
uidx += 1
cum_n_words = 0.0
train_loss = 0.0
train_loss_dict = dict()
else:
continue
# ================================================================================== #
# Display some information
if should_trigger_by_steps(
uidx, eidx, every_n_step=training_configs['disp_freq']):
lrate = list(optim.get_lrate())[0]
if summary_writer is not None:
summary_writer.add_scalar(
"Speed(sents/sec)",
scalar_value=sent_per_sec_meter.ave,
global_step=uidx)
summary_writer.add_scalar(
"Speed(words/sec)",
scalar_value=tok_per_sec_meter.ave,
global_step=uidx)
summary_writer.add_scalar(
"train_loss",
scalar_value=train_loss_meter.ave,
global_step=uidx)
# add loss for every critic
if flags.display_loss_detail:
combination_loss = train_loss_dict_meter.value
for key, value in combination_loss.items():
summary_writer.add_scalar(key,
scalar_value=value,
global_step=uidx)
summary_writer.add_scalar("lrate",
scalar_value=lrate,
global_step=uidx)
summary_writer.add_scalar("oom_count",
scalar_value=oom_count,
global_step=uidx)
# Reset Meters
sent_per_sec_meter.reset()
tok_per_sec_meter.reset()
train_loss_meter.reset()
train_loss_dict_meter.reset()
# ================================================================================== #
# Loss Validation & Learning rate annealing
if should_trigger_by_steps(
global_step=uidx,
n_epoch=eidx,
every_n_step=training_configs['loss_valid_freq'],
debug=flags.debug):
with cache_parameters(nmt_model):
valid_loss, valid_loss_dict = loss_evaluation(
model=nmt_model,
critic=critic,
valid_iterator=valid_iterator,
rank=rank,
world_size=world_size)
if scheduler is not None and optimizer_configs[
"schedule_method"] == "loss":
scheduler.step(metric=valid_loss)
model_collections.add_to_collection("history_losses",
valid_loss)
min_history_loss = np.array(
model_collections.get_collection("history_losses")).min()
best_valid_loss = min_history_loss
if summary_writer is not None:
summary_writer.add_scalar("loss",
valid_loss,
global_step=uidx)
summary_writer.add_scalar("best_loss",
min_history_loss,
global_step=uidx)
# ================================================================================== #
# BLEU Validation & Early Stop
if should_trigger_by_steps(
global_step=uidx,
n_epoch=eidx,
every_n_step=training_configs['bleu_valid_freq'],
min_step=training_configs['bleu_valid_warmup'],
debug=flags.debug):
with cache_parameters(nmt_model):
valid_bleu = bleu_evaluation(
uidx=uidx,
valid_iterator=valid_iterator,
batch_size=training_configs["bleu_valid_batch_size"],
model=nmt_model,
bleu_scorer=bleu_scorer,
vocab_src=vocab_src,
vocab_tgt=vocab_tgt,
valid_dir=flags.valid_path,
max_steps=training_configs["bleu_valid_configs"]
["max_steps"],
beam_size=training_configs["bleu_valid_configs"]
["beam_size"],
alpha=training_configs["bleu_valid_configs"]["alpha"],
world_size=world_size,
rank=rank,
)
model_collections.add_to_collection(key="history_bleus",
value=valid_bleu)
best_valid_bleu = float(
np.array(model_collections.get_collection(
"history_bleus")).max())
if summary_writer is not None:
summary_writer.add_scalar("bleu", valid_bleu, uidx)
summary_writer.add_scalar("best_bleu", best_valid_bleu,
uidx)
# If model get new best valid bleu score
if valid_bleu >= best_valid_bleu:
bad_count = 0
if is_early_stop is False:
if rank == 0:
# 1. save the best model
torch.save(nmt_model.state_dict(),
best_model_prefix + ".final")
else:
bad_count += 1
# At least one epoch should be traversed
if bad_count >= training_configs[
'early_stop_patience'] and eidx > 0:
is_early_stop = True
WARN("Early Stop!")
exit(0)
if rank == 0:
best_k_saver.save(global_step=uidx,
metric=valid_bleu,
model=nmt_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma)
# ODC
if training_configs['use_odc'] is True:
if valid_bleu >= best_valid_bleu:
pass
# choose method to generate teachers from checkpoints
# - best
# - ave_k_best
# - ma
if training_configs['teacher_choice'] == 'ma':
teacher_params = ma.export_ma_params()
elif training_configs['teacher_choice'] == 'best':
teacher_params = nmt_model.state_dict()
elif "ave_best" in training_configs['teacher_choice']:
if best_k_saver.num_saved >= ave_best_k:
teacher_params = average_checkpoints(
best_k_saver.get_all_ckpt_path()
[-ave_best_k:])
else:
teacher_params = nmt_model.state_dict()
else:
raise ValueError(
"can not support teacher choice %s" %
training_configs['teacher_choice'])
torch.save(teacher_params, teacher_model_path)
del teacher_params
teacher_patience = 0
critic.set_use_KD(False)
else:
teacher_patience += 1
if teacher_patience >= training_configs[
'teacher_refresh_warmup']:
teacher_params = torch.load(
teacher_model_path,
map_location=Constants.CURRENT_DEVICE)
teacher_model.load_state_dict(teacher_params,
strict=False)
del teacher_params
critic.reset_teacher(teacher_model)
critic.set_use_KD(True)
if summary_writer is not None:
summary_writer.add_scalar("bad_count", bad_count, uidx)
info_str = "{0} Loss: {1:.2f} BLEU: {2:.2f} lrate: {3:6f} patience: {4} ".format(
uidx, valid_loss, valid_bleu, lrate, bad_count)
for key, value in valid_loss_dict.items():
info_str += (key + ': {0:.2f} '.format(value))
INFO(info_str)
# ================================================================================== #
# Saving checkpoints
if should_trigger_by_steps(
uidx,
eidx,
every_n_step=training_configs['save_freq'],
debug=flags.debug):
model_collections.add_to_collection("uidx", uidx)
model_collections.add_to_collection("eidx", eidx)
model_collections.add_to_collection("bad_count", bad_count)
model_collections.add_to_collection("teacher_patience",
teacher_patience)
if not is_early_stop:
if rank == 0:
checkpoint_saver.save(global_step=uidx,
model=nmt_model,
optim=optim,
lr_scheduler=scheduler,
collections=model_collections,
ma=ma)
if training_progress_bar is not None:
training_progress_bar.close()
eidx += 1
if eidx > training_configs["max_epochs"]:
break