def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# make logger
model_dir = make_model_dir(cfg["training"]["model_dir"],
overwrite=cfg["training"].get("overwrite", False))
_ = make_logger(model_dir, mode="train") # version string returned
# TODO: save version number in model checkpoints
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# load the data
train_data, dev_data, test_data, src_vocab, trg_vocab = load_data(
data_cfg=cfg["data"])
# build an encoder-decoder model
model = build_model(cfg["model"], src_vocab=src_vocab, trg_vocab=trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg)
log_data_info(train_data=train_data, valid_data=dev_data,
test_data=test_data, src_vocab=src_vocab, trg_vocab=trg_vocab)
logger.info(str(model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(model_dir, trainer.stats.best_ckpt_iter)
output_name = "{:08d}.hyps".format(trainer.stats.best_ckpt_iter)
output_path = os.path.join(model_dir, output_name)
datasets_to_test = {"dev": dev_data, "test": test_data,
"src_vocab": src_vocab, "trg_vocab": trg_vocab}
test(cfg_file, ckpt=ckpt, output_path=output_path,
datasets=datasets_to_test)
def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
train_cfg = cfg["training"]
data_cfg = cfg["data"]
# set the random seed
set_seed(seed=train_cfg.get("random_seed", 42))
# load the data
data = load_data(data_cfg)
train_data = data["train_data"]
dev_data = data["dev_data"]
test_data = data["test_data"]
vocabs = data["vocabs"]
# build an encoder-decoder model
model = build_model(cfg["model"], vocabs=vocabs)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, join(trainer.model_dir, "config.yaml"))
# log all entries of config
log_cfg(cfg, trainer.logger)
log_data_info(
train_data=train_data,
valid_data=dev_data,
test_data=test_data,
vocabs=vocabs,
logging_function=trainer.logger.info)
trainer.logger.info(str(model))
# store the vocabs
model_dir = train_cfg["model_dir"]
for field_name, vocab in vocabs.items():
vocab_file = join(model_dir, field_name + "_vocab.txt")
vocab.to_file(vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
# predict with the best model on validation (and test, if available)
ckpt = join(trainer.model_dir, str(trainer.best_ckpt_iteration) + ".ckpt")
output_name = "{:08d}.hyps".format(trainer.best_ckpt_iteration)
output_path = join(trainer.model_dir, output_name)
test(cfg_file, ckpt=ckpt, output_path=output_path, logger=trainer.logger)
def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
print(f'Loading data...')
# load the data
train_data, dev_data, _, trg_vocab = load_data(data_cfg=cfg["data"],
get_test=False)
print(f'Building model...')
# build an encoder-decoder model
model = build_model(cfg["model"], trg_vocab=trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg, trainer.logger)
# log_data_info(train_data=train_data, valid_data=dev_data,
# test_data=test_data, trg_vocab=trg_vocab,
# logging_function=trainer.logger.info)
trainer.logger.info(str(model))
# store the vocabs
# trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
# trg_vocab.to_file(trg_vocab_file)
print(f'Initiating Training...')
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(trainer.model_dir, trainer.best_ckpt_iteration)
output_name = "{:08d}.hyps".format(trainer.best_ckpt_iteration)
output_path = os.path.join(trainer.model_dir, output_name)
return test(cfg_file,
ckpt=ckpt,
output_path=output_path,
logger=trainer.logger,
trg_vocab=trg_vocab)
def train_norm(model, cfg_file: str, skip_test: bool = False) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
:param skip_test: whether a test should be run or not after training
"""
cfg = load_config(cfg_file)
# make logger
model_dir = make_model_dir(cfg["training"]["model_dir"],
overwrite=cfg["training"].get(
"overwrite", False))
_ = make_logger(model_dir, mode="train") # version string returned
# TODO: save version number in model checkpoints
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# load the data
train_data, dev_data, test_data, src_vocab, trg_vocab = load_data(
data_cfg=cfg["data"],
src_lang=cfg["data"].get("src"),
trg_lang=cfg["data"].get("trg"))
# build an encoder-decoder model
#model = build_model(cfg["model"], src_vocab=src_vocab, trg_vocab=trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg)
log_data_info(train_data=train_data,
valid_data=dev_data,
test_data=test_data,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
logger.info(str(model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
shards_dir = os.path.dirname(cfg["data"]["shard_path"])
if not os.path.exists(shards_dir):
os.makedirs(shards_dir)
if cfg["data"].get("shard_data", False):
assert cfg["data"].get(
"n_shards", 0) > 0, "n_shards needs to exist and be at least 1"
shard_data(path=cfg["data"]["train"],
src_lang=cfg["data"]["src"],
tgt_lang=cfg["data"]["trg"],
n_shards=cfg["data"]["n_shards"],
shard_path=cfg["data"]["shard_path"])
# load the data
load_train_whole = True if cfg["data"].get("n_shards", 0) < 1 else False
train_data, dev_data, test_data, src_vocab, trg_vocab, src_field, trg_field = load_data(
data_cfg=cfg["data"], load_train=load_train_whole)
if not load_train_whole:
sharded_iterator = ShardedEpochDatasetIterator(
n_shards=cfg["data"]["n_shards"],
percent_to_sample=cfg["data"].get("percent_to_sample_from_shard",
1.0),
data_path=cfg["data"]["train"],
shard_path=cfg["data"]["shard_path"],
extensions=(cfg["data"]["src"], cfg["data"]["trg"]),
fields=(src_field, trg_field),
n_epochs=cfg["training"]["epochs"],
filter_pred=lambda x: len(vars(x)[
'src']) <= cfg["data"]["max_sent_length"] and len(
vars(x)['trg']) <= cfg["data"]["max_sent_length"])
else:
sharded_iterator = None
# build an encoder-decoder model
model = build_model(cfg["model"], src_vocab=src_vocab, trg_vocab=trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg, trainer.logger)
if load_train_whole:
log_data_info(train_data=train_data,
valid_data=dev_data,
test_data=test_data,
src_vocab=src_vocab,
trg_vocab=trg_vocab,
logging_function=trainer.logger.info)
trainer.logger.info(str(model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data,
valid_data=dev_data,
sharded_iterator=sharded_iterator)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(trainer.model_dir, trainer.best_ckpt_iteration)
output_name = "{:08d}.hyps".format(trainer.best_ckpt_iteration)
output_path = os.path.join(trainer.model_dir, output_name)
test(cfg_file, ckpt=ckpt, output_path=output_path, logger=trainer.logger)
def Q_learning(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file) # config is a dict
# make logger
model_dir = make_model_dir(cfg["training"]["model_dir"],
overwrite=cfg["training"].get(
"overwrite", False))
_ = make_logger(model_dir, mode="train") # version string returned
# TODO: save version number in model checkpoints
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# load the data
print("loadding data here")
train_data, dev_data, test_data, src_vocab, trg_vocab = load_data(
data_cfg=cfg["data"])
# The training data is filtered to include sentences up to `max_sent_length`
# on source and target side.
# training config:
train_config = cfg["training"]
shuffle = train_config.get("shuffle", True)
batch_size = train_config["batch_size"]
mini_BATCH_SIZE = train_config["mini_batch_size"]
batch_type = train_config.get("batch_type", "sentence")
outer_epochs = train_config.get("outer_epochs", 10)
inner_epochs = train_config.get("inner_epochs", 10)
TARGET_UPDATE = train_config.get("target_update", 10)
Gamma = train_config.get("Gamma", 0.999)
use_cuda = train_config["use_cuda"] and torch.cuda.is_available()
# validation part config
# validation
validation_freq = train_config.get("validation_freq", 1000)
ckpt_queue = queue.Queue(maxsize=train_config.get("keep_last_ckpts", 5))
eval_batch_size = train_config.get("eval_batch_size", batch_size)
level = cfg["data"]["level"]
eval_metric = train_config.get("eval_metric", "bleu")
n_gpu = torch.cuda.device_count() if use_cuda else 0
eval_batch_type = train_config.get("eval_batch_type", batch_type)
# eval options
test_config = cfg["testing"]
bpe_type = test_config.get("bpe_type", "subword-nmt")
sacrebleu = {"remove_whitespace": True, "tokenize": "13a"}
max_output_length = train_config.get("max_output_length", None)
minimize_metric = True
# initialize training statistics
stats = TrainStatistics(
steps=0,
stop=False,
total_tokens=0,
best_ckpt_iter=0,
best_ckpt_score=np.inf if minimize_metric else -np.inf,
minimize_metric=minimize_metric)
early_stopping_metric = train_config.get("early_stopping_metric",
"eval_metric")
if early_stopping_metric in ["ppl", "loss"]:
stats.minimize_metric = True
stats.best_ckpt_score = np.inf
elif early_stopping_metric == "eval_metric":
if eval_metric in [
"bleu", "chrf", "token_accuracy", "sequence_accuracy"
]:
stats.minimize_metric = False
stats.best_ckpt_score = -np.inf
# eval metric that has to get minimized (not yet implemented)
else:
stats.minimize_metric = True
# data loader(modified from train_and_validate function
# Returns a torchtext iterator for a torchtext dataset.
# param dataset: torchtext dataset containing src and optionally trg
train_iter = make_data_iter(train_data,
batch_size=batch_size,
batch_type=batch_type,
train=True,
shuffle=shuffle)
# initialize the Replay Memory D with capacity N
memory = ReplayMemory(10000)
steps_done = 0
# initialize two DQN networks
policy_net = build_model(cfg["model"],
src_vocab=src_vocab,
trg_vocab=trg_vocab) # Q_network
target_net = build_model(cfg["model"],
src_vocab=src_vocab,
trg_vocab=trg_vocab) # Q_hat_network
#logger.info(policy_net.src_vocab.stoi)
#print("###############trg vocab: ", len(target_net.trg_vocab.stoi))
#print("trg embed: ", target_net.trg_embed.vocab_size)
if use_cuda:
policy_net.cuda()
target_net.cuda()
target_net.load_state_dict(policy_net.state_dict())
# Initialize target net Q_hat with weights equal to policy_net
target_net.eval() # target_net not update the parameters, test mode
# Optimizer
optimizer = build_optimizer(config=cfg["training"],
parameters=policy_net.parameters())
# Loss function
mse_loss = torch.nn.MSELoss()
pad_index = policy_net.pad_index
# print('!!!'*10, pad_index)
cross_entropy_loss = XentLoss(pad_index=pad_index)
policy_net.loss_function = cross_entropy_loss
# learning rate scheduling
scheduler, scheduler_step_at = build_scheduler(
config=train_config,
scheduler_mode="min" if minimize_metric else "max",
optimizer=optimizer,
hidden_size=cfg["model"]["encoder"]["hidden_size"])
# model parameters
if "load_model" in train_config.keys():
load_model_path = train_config["load_model"]
reset_best_ckpt = train_config.get("reset_best_ckpt", False)
reset_scheduler = train_config.get("reset_scheduler", False)
reset_optimizer = train_config.get("reset_optimizer", False)
reset_iter_state = train_config.get("reset_iter_state", False)
print('settings', reset_best_ckpt, reset_iter_state, reset_optimizer,
reset_scheduler)
logger.info("Loading model from %s", load_model_path)
model_checkpoint = load_checkpoint(path=load_model_path,
use_cuda=use_cuda)
# restore model and optimizer parameters
policy_net.load_state_dict(model_checkpoint["model_state"])
if not reset_optimizer:
optimizer.load_state_dict(model_checkpoint["optimizer_state"])
else:
logger.info("Reset optimizer.")
if not reset_scheduler:
if model_checkpoint["scheduler_state"] is not None and \
scheduler is not None:
scheduler.load_state_dict(model_checkpoint["scheduler_state"])
else:
logger.info("Reset scheduler.")
if not reset_best_ckpt:
stats.best_ckpt_score = model_checkpoint["best_ckpt_score"]
stats.best_ckpt_iter = model_checkpoint["best_ckpt_iteration"]
print('stats.best_ckpt_score', stats.best_ckpt_score)
print('stats.best_ckpt_iter', stats.best_ckpt_iter)
else:
logger.info("Reset tracking of the best checkpoint.")
if (not reset_iter_state and model_checkpoint.get(
'train_iter_state', None) is not None):
train_iter_state = model_checkpoint["train_iter_state"]
# move parameters to cuda
target_net.load_state_dict(policy_net.state_dict())
# Initialize target net Q_hat with weights equal to policy_net
target_net.eval()
if use_cuda:
policy_net.cuda()
target_net.cuda()
for i_episode in range(outer_epochs):
# Outer loop
# get batch
for i, batch in enumerate(iter(train_iter)): # joeynmt training.py 377
# create a Batch object from torchtext batch
# ( use class Batch from batch.py)
# return the sentences same length (with padding) in one batch
batch = Batch(batch, policy_net.pad_index, use_cuda=use_cuda)
# we want to get batch.src and batch.trg
# the shape of batch.src: (batch_size * length of the sentence)
# source here is represented by the word index not word embedding.
encoder_output_batch, _, _, _ = policy_net(
return_type="encode",
src=batch.src,
src_length=batch.src_length,
src_mask=batch.src_mask,
)
trans_output_batch, _ = transformer_greedy(
src_mask=batch.src_mask,
max_output_length=max_output_length,
model=policy_net,
encoder_output=encoder_output_batch,
steps_done=steps_done,
use_cuda=use_cuda)
#print('steps_done',steps_done)
steps_done += 1
#print('trans_output_batch.shape is:', trans_output_batch.shape)
# batch_size * max_translation_sentence_length
#print('batch.src', batch.src)
#print('batch.trg', batch.trg)
print('batch.trg.shape is:', batch.trg.shape)
print('trans_output_batch', trans_output_batch)
reward_batch = [
] # Get the reward_batch (Get the bleu score of the sentences in a batch)
for i in range(int(batch.src.shape[0])):
all_outputs = [(trans_output_batch[i])[1:]]
all_ref = [batch.trg[i]]
sentence_score = calculate_bleu(model=policy_net,
level=level,
raw_hypo=all_outputs,
raw_ref=all_ref)
reward_batch.append(sentence_score)
print('reward batch is', reward_batch)
reward_batch = torch.tensor(reward_batch, dtype=torch.float)
# reward_batch = bleu(hypotheses, references, tokenize="13a")
# print('reward_batch.shape', reward_batch.shape)
# make prefix and push tuples into memory
push_sample_to_memory(model=policy_net,
level=level,
eos_index=policy_net.eos_index,
memory=memory,
src_batch=batch.src,
trg_batch=batch.trg,
trans_output_batch=trans_output_batch,
reward_batch=reward_batch,
max_output_length=max_output_length)
print(memory.capacity, len(memory.memory))
if len(memory.memory) == memory.capacity:
# inner loop
for t in range(inner_epochs):
# Sample mini-batch from the memory
transitions = memory.sample(mini_BATCH_SIZE)
# transition = [Transition(source=array([]), prefix=array([]), next_word= int, reward= int),
# Transition(source=array([]), prefix=array([]), next_word= int, reward= int,...]
# Each Transition is what we push into memory for one sentence: memory.push(source, prefix, next_word, reward_batch[i])
mini_batch = Transition(*zip(*transitions))
# merge the same class in transition together
# mini_batch = Transition(source=(array([]), array([]),...), prefix=(array([],...),
# next_word=array([...]), reward=array([...]))
# mini_batch.reward is tuple: length is mini_BATCH_SIZE.
#print('mini_batch', mini_batch)
#concatenate together into a tensor.
words = []
for word in mini_batch.next_word:
new_word = word.unsqueeze(0)
words.append(new_word)
mini_next_word = torch.cat(
words) # shape (mini_BATCH_SIZE,)
mini_reward = torch.tensor(
mini_batch.reward) # shape (mini_BATCH_SIZE,)
#print('mini_batch.finish', mini_batch.finish)
mini_is_eos = torch.Tensor(mini_batch.finish)
#print(mini_is_eos)
mini_src_length = [
len(item) for item in mini_batch.source_sentence
]
mini_src_length = torch.Tensor(mini_src_length)
mini_src = pad_sequence(mini_batch.source_sentence,
batch_first=True,
padding_value=float(pad_index))
# shape (mini_BATCH_SIZE, max_length_src)
length_prefix = [len(item) for item in mini_batch.prefix]
mini_prefix_length = torch.Tensor(length_prefix)
prefix_list = []
for prefix_ in mini_batch.prefix:
prefix_ = torch.from_numpy(prefix_)
prefix_list.append(prefix_)
mini_prefix = pad_sequence(prefix_list,
batch_first=True,
padding_value=pad_index)
# shape (mini_BATCH_SIZE, max_length_prefix)
mini_src_mask = (mini_src != pad_index).unsqueeze(1)
mini_trg_mask = (mini_prefix != pad_index).unsqueeze(1)
#print('mini_src', mini_src)
#print('mini_src_length', mini_src_length)
#print('mini_src_mask', mini_src_mask)
#print('mini_prefix', mini_prefix)
#print('mini_trg_mask', mini_trg_mask)
#print('mini_reward', mini_reward)
# max_length_src = torch.max(mini_src_length) #max([len(item) for item in mini_batch.source_sentence])
if use_cuda:
mini_src = mini_src.cuda()
mini_prefix = mini_prefix.cuda()
mini_src_mask = mini_src_mask.cuda()
mini_src_length = mini_src_length.cuda()
mini_trg_mask = mini_trg_mask.cuda()
mini_next_word = mini_next_word.cuda()
# print(next(policy_net.parameters()).is_cuda)
# print(mini_trg_mask.get_device())
# calculate the Q_value
logits_Q, _, _, _ = policy_net._encode_decode(
src=mini_src,
trg_input=mini_prefix,
src_mask=mini_src_mask,
src_length=mini_src_length,
trg_mask=
mini_trg_mask # trg_mask = (self.trg_input != pad_index).unsqueeze(1)
)
#print('mini_prefix_length', mini_prefix_length)
#print('logits_Q.shape', logits_Q.shape) # torch.Size([64, 99, 31716])
#print('logits_Q', logits_Q)
# length_prefix = max([len(item) for item in mini_batch.prefix])
# logits_Q shape: batch_size * length of the sentence * total number of words in corpus.
logits_Q = logits_Q[range(mini_BATCH_SIZE),
mini_prefix_length.long() - 1, :]
#print('logits_Q_.shape', logits_Q.shape) #shape(mini_batch_size, num_words)
# logits shape: mini_batch_size * total number of words in corpus
Q_value = logits_Q[range(mini_BATCH_SIZE), mini_next_word]
#print('mini_next_word', mini_next_word)
#print("Q_value", Q_value)
mini_prefix_add = torch.cat(
[mini_prefix, mini_next_word.unsqueeze(1)], dim=1)
#print('mini_prefix_add', mini_prefix_add)
mini_trg_mask_add = (mini_prefix_add !=
pad_index).unsqueeze(1)
#print('mini_trg_mask_add', mini_trg_mask_add)
if use_cuda:
mini_prefix_add = mini_prefix_add.cuda()
mini_trg_mask_add = mini_trg_mask_add.cuda()
logits_Q_hat, _, _, _ = target_net._encode_decode(
src=mini_src,
trg_input=mini_prefix_add,
src_mask=mini_src_mask,
src_length=mini_src_length,
trg_mask=mini_trg_mask_add)
#print('mini_prefix_add.shape', mini_prefix_add.shape)
#print('logits_Q_hat.shape', logits_Q_hat.shape)
#print('mini_prefix_length.long()', mini_prefix_length.long())
logits_Q_hat = logits_Q_hat[range(mini_BATCH_SIZE),
mini_prefix_length.long(), :]
Q_hat_value, _ = torch.max(logits_Q_hat, dim=1)
#print('Q_hat_value', Q_hat_value)
if use_cuda:
Q_hat_value = Q_hat_value.cuda()
mini_reward = mini_reward.cuda()
mini_is_eos = mini_is_eos.cuda()
yj = mini_reward.float() + Gamma * Q_hat_value
#print('yj', yj)
index = mini_is_eos.long()
#print('mini_is_eos', mini_is_eos)
yj[index] = mini_reward[index]
#print('yj', yj)
#print('Q_value1', Q_value)
yj.detach()
# Optimize the model
policy_net.zero_grad()
# Compute loss
loss = mse_loss(yj, Q_value)
print('loss', loss)
logger.info("step = {}, loss = {}".format(
stats.steps, loss.item()))
loss.backward()
#for param in policy_net.parameters():
# param.grad.data.clamp_(-1, 1)
optimizer.step()
stats.steps += 1
#print('step', stats.steps)
if stats.steps % TARGET_UPDATE == 0:
#print('update the parameters in target_net.')
target_net.load_state_dict(policy_net.state_dict())
if stats.steps % validation_freq == 0: # Validation
print('Start validation')
valid_score, valid_loss, valid_ppl, valid_sources, \
valid_sources_raw, valid_references, valid_hypotheses, \
valid_hypotheses_raw, valid_attention_scores = \
validate_on_data(
model=policy_net,
data=dev_data,
batch_size=eval_batch_size,
use_cuda=use_cuda,
level=level,
eval_metric=eval_metric,
n_gpu=n_gpu,
compute_loss=True,
beam_size=1,
beam_alpha=-1,
batch_type=eval_batch_type,
postprocess=True,
bpe_type=bpe_type,
sacrebleu=sacrebleu,
max_output_length=max_output_length
)
print(
'validation_loss: {}, validation_score: {}'.format(
valid_loss, valid_score))
logger.info(valid_loss)
print('average loss: total_loss/n_tokens:', valid_ppl)
if early_stopping_metric == "loss":
ckpt_score = valid_loss
elif early_stopping_metric in ["ppl", "perplexity"]:
ckpt_score = valid_ppl
else:
ckpt_score = valid_score
if stats.is_best(ckpt_score):
stats.best_ckpt_score = ckpt_score
stats.best_ckpt_iter = stats.steps
logger.info(
'Hooray! New best validation result [%s]!',
early_stopping_metric)
if ckpt_queue.maxsize > 0:
logger.info("Saving new checkpoint.")
# def _save_checkpoint(self) -> None:
"""
Save the model's current parameters and the training state to a
checkpoint.
The training state contains the total number of training steps,
the total number of training tokens,
the best checkpoint score and iteration so far,
and optimizer and scheduler states.
"""
model_path = "{}/{}.ckpt".format(
model_dir, stats.steps)
model_state_dict = policy_net.module.state_dict() \
if isinstance(policy_net, torch.nn.DataParallel) \
else policy_net.state_dict()
state = {
"steps": stats.steps,
"total_tokens": stats.total_tokens,
"best_ckpt_score": stats.best_ckpt_score,
"best_ckpt_iteration":
stats.best_ckpt_iter,
"model_state": model_state_dict,
"optimizer_state": optimizer.state_dict(),
# "scheduler_state": scheduler.state_dict() if
# self.scheduler is not None else None,
# 'amp_state': amp.state_dict() if self.fp16 else None
}
torch.save(state, model_path)
if ckpt_queue.full():
to_delete = ckpt_queue.get(
) # delete oldest ckpt
try:
os.remove(to_delete)
except FileNotFoundError:
logger.warning(
"Wanted to delete old checkpoint %s but "
"file does not exist.", to_delete)
ckpt_queue.put(model_path)
best_path = "{}/best.ckpt".format(model_dir)
try:
# create/modify symbolic link for best checkpoint
symlink_update(
"{}.ckpt".format(stats.steps),
best_path)
except OSError:
# overwrite best.ckpt
torch.save(state, best_path)
def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# load the data
train_data, dev_data, test_data, src_vocab, trg_vocab = load_data(
data_cfg=cfg["data"])
# build an encoder-decoder model
model = build_model(cfg["model"], src_vocab=src_vocab, trg_vocab=trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg, trainer.logger)
log_data_info(train_data=train_data,
valid_data=dev_data,
test_data=test_data,
src_vocab=src_vocab,
trg_vocab=trg_vocab,
logging_function=trainer.logger.info)
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
# test the model with the best checkpoint
if test_data is not None:
# load checkpoint
if trainer.best_ckpt_iteration > 0:
checkpoint_path = "{}/{}.ckpt".format(trainer.model_dir,
trainer.best_ckpt_iteration)
else:
## For save_checkpoint by save_freq
checkpoint_path = get_latest_checkpoint(trainer.model_dir)
try:
trainer.init_from_checkpoint(checkpoint_path)
except AssertionError:
trainer.logger.warning(
"Checkpoint %s does not exist. "
"Skipping testing.", checkpoint_path)
if trainer.best_ckpt_iteration == 0 \
and trainer.best_ckpt_score in [np.inf, -np.inf]:
trainer.logger.warning(
"It seems like no checkpoint was written, "
"since no improvement was obtained over the initial model."
)
return
# generate hypotheses for test data
if "testing" in cfg.keys():
beam_size = cfg["testing"].get("beam_size", 0)
beam_alpha = cfg["testing"].get("alpha", -1)
return_logp = cfg["testing"].get("return_logp", False)
else:
beam_size = 0
beam_alpha = -1
return_logp = False
# pylint: disable=unused-variable
score, loss, ppl, sources, sources_raw, references, hypotheses, \
hypotheses_raw, attention_scores, log_probs = validate_on_data(
data=test_data, batch_size=trainer.batch_size,
eval_metric=trainer.eval_metric, level=trainer.level,
max_output_length=trainer.max_output_length,
model=model, use_cuda=trainer.use_cuda, loss_function=None,
beam_size=beam_size, beam_alpha=beam_alpha,
return_logp=return_logp)
if "trg" in test_data.fields:
decoding_description = "Greedy decoding" if beam_size == 0 else \
"Beam search decoding with beam size = {} and alpha = {}"\
.format(beam_size, beam_alpha)
trainer.logger.info("Test data result: %f %s [%s]", score,
trainer.eval_metric, decoding_description)
else:
trainer.logger.info(
"No references given for %s.%s -> no evaluation.",
cfg["data"]["test"], cfg["data"]["src"])
output_path_set = "{}/{}.{}".format(trainer.model_dir, "test",
cfg["data"]["trg"])
with open(output_path_set, mode="w", encoding="utf-8") as f:
for h in hypotheses:
f.write("{}\n".format(h))
trainer.logger.info("Test translations saved to: %s", output_path_set)
if return_logp:
output_path_set_logp = output_path_set + ".logp"
with open(output_path_set_logp, mode="w", encoding="utf-8") as f:
for l in log_probs:
f.write("{}\n".format(l))
trainer.logger.info("Test log probs saved to: %s",
output_path_set_logp)
def train_transfer(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# set the random seed
set_seed(seed=cfg["pretraining"].get("random_seed", 42))
# load the data
pre_train_data, pre_dev_data, pre_test_data, pre_src_vocab, pre_trg_vocab = load_data(
data_cfg=cfg["pretrained_data"])
# build an encoder-decoder model
pretrained_model = build_model(cfg["model"],
src_vocab=pre_src_vocab,
trg_vocab=pre_trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=pretrained_model,
config=cfg,
training_key="pretraining",
name_log="pre_train")
# store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg, trainer.logger)
log_data_info(train_data=pre_train_data,
valid_data=pre_dev_data,
test_data=pre_test_data,
src_vocab=pre_src_vocab,
trg_vocab=pre_trg_vocab,
logging_function=trainer.logger.info)
trainer.logger.info(str(pretrained_model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["pretraining"]["model_dir"])
pre_src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["pretraining"]["model_dir"])
pre_trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.train_and_validate(train_data=pre_train_data,
valid_data=pre_dev_data)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(trainer.model_dir, trainer.best_ckpt_iteration)
output_name = "{:08d}.hyps".format(trainer.best_ckpt_iteration)
output_path = os.path.join(trainer.model_dir, output_name)
test(cfg_file,
ckpt=ckpt,
output_path=output_path,
logger=trainer.logger,
key_training="pretraining",
key_data="pretrained_data")
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# load the data
train_data, dev_data, test_data, src_vocab, trg_vocab = load_data(
data_cfg=cfg["data"])
# build an encoder-decoder model
model = build_pretrained_model(cfg["model"],
pretrained_model=pretrained_model,
pretrained_src_vocab=pre_src_vocab,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg, training_key="training")
# store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg, trainer.logger)
log_data_info(train_data=train_data,
valid_data=dev_data,
test_data=test_data,
src_vocab=src_vocab,
trg_vocab=trg_vocab,
logging_function=trainer.logger.info)
trainer.logger.info(str(model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(trainer.model_dir, trainer.best_ckpt_iteration)
output_name = "{:08d}.hyps".format(trainer.best_ckpt_iteration)
output_path = os.path.join(trainer.model_dir, output_name)
test(cfg_file,
ckpt=ckpt,
output_path=output_path,
logger=trainer.logger,
key_training="training",
key_data="data")
def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
kb_task = bool(cfg["data"].get("kb_task", False))
# load the data
train_data, dev_data, test_data,\
src_vocab, trg_vocab,\
train_kb, dev_kb, test_kb,\
train_kb_lookup, dev_kb_lookup, test_kb_lookup,\
train_kb_lengths, dev_kb_lengths, test_kb_lengths,\
train_kb_truvals, dev_kb_truvals, test_kb_truvals,\
trv_vocab, canonizer,\
dev_data_canon, test_data_canon\
= load_data(data_cfg=cfg["data"])
# build an encoder-decoder model
model = build_model(cfg["model"],
src_vocab=src_vocab,
trg_vocab=trg_vocab,
trv_vocab=trv_vocab,
canonizer=canonizer)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, trainer.model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg, trainer.logger)
log_data_info(train_data=train_data,
valid_data=dev_data,
test_data=test_data,
src_vocab=src_vocab,
trg_vocab=trg_vocab,
logging_function=trainer.logger.info)
trainer.logger.info(str(model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
if kb_task:
trv_vocab_file = "{}/trv_vocab.txt".format(
cfg["training"]["model_dir"])
trv_vocab.to_file(trv_vocab_file)
# train the model
trainer.train_and_validate(train_data=train_data, valid_data=dev_data, kb_task=kb_task,\
train_kb=train_kb, train_kb_lkp=train_kb_lookup, train_kb_lens=train_kb_lengths, train_kb_truvals=train_kb_truvals,\
valid_kb=dev_kb, valid_kb_lkp=dev_kb_lookup, valid_kb_lens=dev_kb_lengths, valid_kb_truvals=dev_kb_truvals,\
valid_data_canon=dev_data_canon)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(trainer.model_dir, trainer.best_ckpt_iteration)
output_name = "{:08d}.hyps".format(trainer.best_ckpt_iteration)
output_path = os.path.join(trainer.model_dir, output_name)
test(cfg_file, ckpt=ckpt, output_path=output_path, logger=trainer.logger)
def train(cfg_file: str) -> None:
"""
Main training function. After training, also test on test data if given.
:param cfg_file: path to configuration yaml file
"""
cfg = load_config(cfg_file)
# make logger
model_dir = make_model_dir(cfg["training"]["model_dir"],
overwrite=cfg["training"].get(
"overwrite", False))
_ = make_logger(model_dir, mode="train") # version string returned
# TODO: save version number in model checkpoints
# set the random seed
set_seed(seed=cfg["training"].get("random_seed", 42))
# load the data
train_tasks_list = []
valid_tasks_list = []
src_tasks = cfg["data"].get("src")
trg_tasks = cfg["data"].get("trg")
for x in range(len(src_tasks)):
src_lang = src_tasks[x]
trg_lang = trg_tasks[x]
train_data, dev_data, _, _, _ = load_data(data_cfg=cfg["data"],
src_lang=src_lang,
trg_lang=trg_lang)
train_tasks_list.append(train_data)
valid_tasks_list.append(dev_data)
#build vocabulary
logger.info("Building vocabulary...")
src_max_size = cfg["data"].get("src_voc_limit", sys.maxsize)
src_min_freq = cfg["data"].get("src_voc_min_freq", 1)
trg_max_size = cfg["data"].get("trg_voc_limit", sys.maxsize)
trg_min_freq = cfg["data"].get("trg_voc_min_freq", 1)
src_vocab_file = cfg["data"].get("src_vocab", None)
trg_vocab_file = cfg["data"].get("trg_vocab", None)
src_vocab = build_vocab(field="src",
min_freq=src_min_freq,
max_size=src_max_size,
dataset=train_tasks_list[0],
vocab_file=src_vocab_file)
trg_vocab = build_vocab(field="trg",
min_freq=trg_min_freq,
max_size=trg_max_size,
dataset=train_tasks_list[0],
vocab_file=trg_vocab_file)
# build an encoder-decoder model
model = build_model(cfg["model"], src_vocab=src_vocab, trg_vocab=src_vocab)
# for training management, e.g. early stopping and model selection
trainer = TrainManager(model=model, config=cfg)
# store copy of original training config in model dir
shutil.copy2(cfg_file, model_dir + "/config.yaml")
# log all entries of config
log_cfg(cfg)
# log_data_info(train_data=train_data,
# valid_data=dev_data,
# test_data=test_data,
# src_vocab=src_vocab,
# trg_vocab=trg_vocab)
logger.info(str(model))
# store the vocabs
src_vocab_file = "{}/src_vocab.txt".format(cfg["training"]["model_dir"])
src_vocab.to_file(src_vocab_file)
trg_vocab_file = "{}/trg_vocab.txt".format(cfg["training"]["model_dir"])
trg_vocab.to_file(trg_vocab_file)
# train the model
trainer.maml_train_and_validate(train_tasks=train_tasks_list,
valid_tasks=valid_tasks_list)
# predict with the best model on validation and test
# (if test data is available)
ckpt = "{}/{}.ckpt".format(model_dir, trainer.stats.best_ckpt_iter)
output_name = "{:08d}.hyps".format(trainer.stats.best_ckpt_iter)
output_path = os.path.join(model_dir, output_name)
