def train(opt):
ArgumentParser.validate_train_opts(opt)
ArgumentParser.update_model_opts(opt)
ArgumentParser.validate_model_opts(opt)
set_random_seed(opt.seed, False)
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
if 'vocab' in checkpoint:
logger.info('Loading vocab from checkpoint at %s.' %
opt.train_from)
vocab = checkpoint['vocab']
else:
vocab = torch.load(opt.data + '.vocab.pt')
else:
vocab = torch.load(opt.data + '.vocab.pt')
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
fields = vocab
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
nb_gpu = len(opt.gpu_ranks)
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
# Create a thread to listen for errors in the child processes.
error_queue = mp.SimpleQueue()
error_handler = ErrorHandler(error_queue)
# Train with multiprocessing.
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
procs.append(
mp.Process(target=run,
args=(opt, device_id, error_queue, q, semaphore),
daemon=True))
procs[device_id].start()
logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
producer = mp.Process(target=batch_producer,
args=(
train_iter,
queues,
semaphore,
opt,
),
daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs:
p.join()
producer.terminate()
elif nb_gpu == 1: # case 1 GPU only
single_main(opt, 0)
else: # case only CPU
single_main(opt, -1)
def train(opt):
ArgumentParser.validate_train_opts(opt)
set_random_seed(opt.seed, False)
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
vocab = torch.load(opt.data + '.vocab.pt')
if old_style_vocab(vocab):
fields = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
fields = vocab
patch_fields(opt, fields)
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
nb_gpu = len(opt.gpu_ranks)
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
procs.append(
mp.Process(target=run,
args=(opt, device_id, error_queue, q, semaphore),
daemon=True))
procs[device_id].start()
logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
producer = mp.Process(target=batch_producer,
args=(
train_iter,
queues,
semaphore,
opt,
),
daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs:
p.join()
producer.terminate()
elif nb_gpu == 1:
single_main(opt, 0)
else:
single_main(opt, -1)
def main(opt, device_id, batch_queue=None, semaphore=None):
# NOTE: It's important that ``opt`` has been validated and updated
# at this point.
configure_process(opt, device_id)
init_logger(opt.log_file)
assert len(opt.accum_count) == len(opt.accum_steps), \
'Number of accum_count values must match number of accum_steps'
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
checkpoint = None
model_opt = opt
vocab = torch.load(opt.data + '.vocab.pt')
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
fields = vocab
# Report src and tgt vocab sizes, including for features
for side in ['src', 'tgt']:
f = fields[side]
try:
f_iter = iter(f)
except TypeError:
f_iter = [(side, f)]
for sn, sf in f_iter:
if sf.use_vocab:
logger.info(' * %s vocab size = %d' % (sn, len(sf.vocab)))
# Build model.
model = build_model(model_opt, opt, fields, checkpoint)
n_params, enc, dec, nontrainable = _tally_parameters(model)
logger.info('encoder: %d' % enc)
logger.info('decoder: %d' % dec)
logger.info('non-trainable parameters (tgt_out_emb): %d' % nontrainable)
logger.info('* number of parameters: %d' % n_params)
_check_save_model_path(opt)
# Build optimizer.
optim = Optimizer.from_opt(model, opt, checkpoint=checkpoint)
# Build model saver
model_saver = build_model_saver(model_opt, opt, model, fields, optim)
trainer = build_trainer(opt,
device_id,
model,
fields,
optim,
model_saver=model_saver)
if batch_queue is None:
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
else:
assert semaphore is not None, \
"Using batch_queue requires semaphore as well"
def _train_iter():
while True:
batch = batch_queue.get()
semaphore.release()
yield batch
train_iter = _train_iter()
valid_iter = build_dataset_iter("valid", fields, opt, is_train=False)
if len(opt.gpu_ranks):
logger.info('Starting training on GPU: %s' % opt.gpu_ranks)
else:
logger.info('Starting training on CPU, could be very slow')
train_steps = opt.train_steps
if opt.single_pass and train_steps > 0:
logger.warning("Option single_pass is enabled, ignoring train_steps.")
train_steps = 0
trainer.train(train_iter,
train_steps,
save_checkpoint_steps=opt.save_checkpoint_steps,
valid_iter=valid_iter,
valid_steps=opt.valid_steps)
if trainer.report_manager.tensorboard_writer is not None:
trainer.report_manager.tensorboard_writer.close()
def train(opt):
ArgumentParser.validate_train_opts(opt)
ArgumentParser.update_model_opts(opt)
ArgumentParser.validate_model_opts(opt)
if opt.train_from != '':
raise Exception(
'train_from will be set automatically to the latest model, you should not set it manually'
)
# set gpu ranks automatically if not specified
if len(opt.gpu_ranks) == 0:
opt.gpu_ranks = list(range(opt.world_size))
# Set train_from to latest checkpoint if it exists
file_list = glob.glob(opt.save_model + '*.pt')
if len(os.listdir(os.path.dirname(
opt.save_model))) > 0 and len(file_list) == 0:
raise Exception(
'save_model directory is not empty but no pretrained models found')
if len(file_list) > 0:
ckpt_nos = list(
map(lambda x: int(x.split('_')[-1].split('.')[0]), file_list))
ckpt_no = max(ckpt_nos)
opt.train_from = opt.save_model + '_' + str(ckpt_no) + '.pt'
print(opt.train_from)
assert os.path.exists(opt.train_from)
set_random_seed(opt.seed, False)
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
vocab = torch.load(opt.data + '.vocab.pt')
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
fields = vocab
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
nb_gpu = len(opt.gpu_ranks)
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
# Create a thread to listen for errors in the child processes.
error_queue = mp.SimpleQueue()
error_handler = ErrorHandler(error_queue)
# Train with multiprocessing.
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
procs.append(
mp.Process(target=run,
args=(opt, device_id, error_queue, q, semaphore),
daemon=True))
procs[device_id].start()
logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
producer = mp.Process(target=batch_producer,
args=(
train_iter,
queues,
semaphore,
opt,
),
daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs:
p.join()
producer.terminate()
elif nb_gpu == 1: # case 1 GPU only
single_main(opt, 0)
else: # case only CPU
single_main(opt, -1)
def main(opt, device_id, batch_queue=None, semaphore=None):
# NOTE: It's important that ``opt`` has been validated and updated
# at this point.
configure_process(opt, device_id)
init_logger(opt.log_file)
assert len(opt.accum_count) == len(opt.accum_steps), \
'Number of accum_count values must match number of accum_steps'
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
checkpoint = None
model_opt = opt
vocab = torch.load(opt.data + '.vocab.pt')
if opt.teacher_model_path:
logger.info('Loading teacher model from {path}'.format(
path=opt.teacher_model_path))
teacher_model_ckpt = torch.load(
opt.teacher_model_path, map_location=lambda storage, loc: storage)
teacher_model_opt = ArgumentParser.ckpt_model_opts(
teacher_model_ckpt['opt'])
ArgumentParser.update_model_opts(teacher_model_opt)
ArgumentParser.validate_model_opts(teacher_model_opt)
logger.info('Loading vocab from checkpoint at {path}'.format(
path=opt.teacher_model_path))
teacher_vocab = teacher_model_ckpt['vocab']
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
fields = vocab
teacher_fields = teacher_vocab if opt.teacher_model_path else None
# patch for fields that may be missing in old data/model
# patch_fields(opt, fields)
# Report src and tgt vocab sizes, including for features
report_vocab_size(fields)
if teacher_fields is not None:
report_vocab_size(teacher_fields)
# Build model.
fields_opt = {"original": fields, "teacher": teacher_fields}
model = custom_builder.build_model(model_opt, opt, fields_opt, checkpoint)
# model = build_model(model_opt, opt, fields, checkpoint)
teacher_model = build_model(
teacher_model_opt, teacher_model_opt, teacher_fields,
teacher_model_ckpt) if opt.teacher_model_path else None
n_params, enc, dec = _tally_parameters(model)
logger.info('encoder: %d' % enc)
logger.info('decoder: %d' % dec)
logger.info('* number of parameters: %d' % n_params)
_check_save_model_path(opt)
if teacher_model is not None:
n_params, enc, dec = _tally_parameters(teacher_model)
logger.info('encoder: %d' % enc)
logger.info('decoder: %d' % dec)
logger.info('* number of parameters: %d' % n_params)
_check_save_model_path(teacher_model_opt)
# Build optimizer.
optim = Optimizer.from_opt(model, opt, checkpoint=checkpoint)
# Build model saver
# model_saver = build_model_saver(model_opt, opt, model, fields, optim)
model_saver = custom_model_saver.build_model_saver(model_opt, opt, model,
fields_opt, optim)
tgt_field = dict(teacher_fields)["tgt"].base_field if teacher_model is not None \
else dict(fields)["tgt"].base_field
sos_id = tgt_field.vocab.stoi[tgt_field.init_token]
if teacher_model is not None and opt.word_sampling:
sampler = Emulator(teacher_model,
teacher_fields,
device_id,
max_length=50,
random_sampling_topk=5)
else:
sampler = None
if teacher_model is not None:
trainer = build_trainer(opt,
device_id,
model,
teacher_fields,
optim,
model_saver,
teacher_model=teacher_model,
emulator=sampler)
else:
trainer = build_trainer(opt,
device_id,
model,
fields,
optim,
model_saver,
teacher_model=teacher_model,
emulator=sampler)
if batch_queue is None:
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
else:
assert semaphore is not None, \
"Using batch_queue requires semaphore as well"
def _train_iter():
while True:
batch = batch_queue.get()
semaphore.release()
yield batch
train_iter = _train_iter()
valid_iter = build_dataset_iter("valid", fields, opt, is_train=False)
if len(opt.gpu_ranks):
logger.info('Starting training on GPU: %s' % opt.gpu_ranks)
else:
logger.info('Starting training on CPU, could be very slow')
train_steps = opt.train_steps
if opt.single_pass and train_steps > 0:
logger.warning("Option single_pass is enabled, ignoring train_steps.")
train_steps = 0
trainer.train(train_iter,
train_steps,
sos_id=sos_id,
save_checkpoint_steps=opt.save_checkpoint_steps,
valid_iter=valid_iter,
valid_steps=opt.valid_steps)
if trainer.report_manager.tensorboard_writer is not None:
trainer.report_manager.tensorboard_writer.close()
def main(opt, device_id, batch_queue=None, semaphore=None):
# NOTE: It's important that ``opt`` has been validated and updated
# at this point.
configure_process(opt, device_id)
init_logger(opt.log_file)
# save training settings
if opt.log_file:
shutil.copy2(opt.config, opt.exp_dir)
logger.info(vars(opt))
assert len(opt.accum_count) == len(opt.accum_steps), \
'Number of accum_count values must match number of accum_steps'
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
checkpoint = None
model_opt = opt
# added by @memray for multiple datasets
if opt.vocab and opt.vocab != 'none':
vocab = torch.load(opt.vocab)
elif opt.encoder_type == 'pretrained':
vocab = None
else:
vocab = None
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(
vocab, opt.model_type, dynamic_dict=opt.copy_attn)
else:
fields = vocab
# @memray: a temporary workaround, as well as train.py line 43
if opt.model_type == "keyphrase":
if opt.tgt_type in ["one2one", "multiple"]:
if 'sep_indices' in fields:
del fields['sep_indices']
else:
if 'sep_indices' not in fields:
sep_indices = Field(
use_vocab=False, dtype=torch.long,
postprocessing=make_tgt, sequential=False)
fields["sep_indices"] = sep_indices
if 'src_ex_vocab' not in fields:
src_ex_vocab = RawField()
fields["src_ex_vocab"] = src_ex_vocab
tokenizer = None
if opt.pretrained_tokenizer:
tokenizer = load_pretrained_tokenizer(opt.pretrained_tokenizer, opt.cache_dir, opt.special_vocab_path)
setattr(opt, 'vocab_size', len(tokenizer))
if opt.data_type == 'news':
fields = reload_news_fields(fields, opt, tokenizer)
# Report src and tgt vocab sizes, including for features
for side in ['src', 'tgt']:
f = fields[side]
try:
f_iter = iter(f)
except TypeError:
f_iter = [(side, f)]
for sn, sf in f_iter:
if sf.use_vocab:
logger.info(' * %s vocab size = %d' % (sn, len(sf.vocab)))
# Build model.
model = build_model(model_opt, opt, fields, checkpoint)
n_params, enc, dec = _tally_parameters(model)
logger.info('encoder: %d' % enc)
logger.info('decoder: %d' % dec)
logger.info('* number of parameters: %d' % n_params)
# Build optimizer.
optim = Optimizer.from_opt(model, opt, checkpoint=checkpoint)
# Build model saver
model_saver = build_model_saver(model_opt, opt, model, fields, optim)
trainer = build_trainer(
opt, device_id, model, fields, optim, model_saver=model_saver)
if batch_queue is None:
if len(opt.data_ids) > 1:
# added by @memray, for loading multiple datasets
if opt.multi_dataset:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt, tokenizer=tokenizer)
else:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt, tokenizer=tokenizer)
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
else:
assert semaphore is not None, \
"Using batch_queue requires semaphore as well"
def _train_iter():
while True:
batch = batch_queue.get()
semaphore.release()
yield batch
train_iter = _train_iter()
if opt.valid:
valid_iter = build_dataset_iter(
"valid", fields, opt, is_train=False)
else:
valid_iter = None
if len(opt.gpu_ranks):
logger.info('Starting training on GPU: %s' % opt.gpu_ranks)
else:
logger.info('Starting training on CPU, could be very slow')
train_steps = opt.train_steps
if opt.single_pass and train_steps > 0:
logger.warning("Option single_pass is enabled, ignoring train_steps.")
train_steps = 0
trainer.train(
train_iter,
train_steps,
save_checkpoint_steps=opt.save_checkpoint_steps,
valid_iter=valid_iter,
valid_steps=opt.valid_steps)
if trainer.report_manager.tensorboard_writer is not None:
trainer.report_manager.tensorboard_writer.close()
def train(opt):
ArgumentParser.validate_train_opts(opt)
ArgumentParser.update_model_opts(opt)
ArgumentParser.validate_model_opts(opt)
set_random_seed(opt.seed, False)
# @Memray, check the dir existence beforehand to avoid path conflicting errors,
# and set save_model, tensorboard_log_dir, wandb_log_dir if not exist
train_single._check_save_model_path(opt)
if not os.path.exists(opt.tensorboard_log_dir):
os.makedirs(opt.tensorboard_log_dir)
# Scan previous checkpoint to resume training
latest_step = 0
latest_ckpt = None
for subdir, dirs, filenames in os.walk(opt.exp_dir):
for filename in sorted(filenames):
if not filename.endswith('.pt'):
continue
step = int(filename[filename.rfind('_') + 1:filename.rfind('.pt')])
if step > latest_step:
latest_ckpt = os.path.join(subdir, filename)
latest_step = step
# if not saved in the exp folder, check opt.save_model
if latest_ckpt is None and opt.save_model is not None:
save_model_dir = os.path.dirname(os.path.abspath(opt.save_model))
model_prefix = opt.save_model[opt.save_model.rfind(os.path.sep) + 1:]
for subdir, dirs, filenames in os.walk(save_model_dir):
for filename in sorted(filenames):
if not filename.endswith('.pt'):
continue
if not filename.startswith(model_prefix):
continue
step = int(filename[filename.rfind('_') +
1:filename.rfind('.pt')])
if step > latest_step:
latest_ckpt = os.path.join(subdir, filename)
latest_step = step
if latest_ckpt is not None:
logger.info("A previous checkpoint is found, train from it: %s" %
latest_ckpt)
setattr(opt, 'train_from', latest_ckpt)
setattr(opt, 'reset_optim', 'none')
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
elif opt.vocab and opt.vocab != 'none':
# added by @memray for multiple datasets
vocab = torch.load(opt.vocab)
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
vocab = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
elif opt.encoder_type == 'pretrained':
vocab = None
else:
vocab = None
fields = vocab
# @memray: a temporary workaround, as well as train_single.py line 78
if fields and opt.data_type == "keyphrase":
if opt.tgt_type in ["one2one", "multiple"]:
if 'sep_indices' in fields:
del fields['sep_indices']
else:
if 'sep_indices' not in fields:
sep_indices = Field(use_vocab=False,
dtype=torch.long,
postprocessing=make_tgt,
sequential=False)
fields["sep_indices"] = sep_indices
if 'src_ex_vocab' not in fields:
src_ex_vocab = RawField()
fields["src_ex_vocab"] = src_ex_vocab
# @memray reload fields for news dataset and pretrained models
tokenizer = None
if opt.pretrained_tokenizer is not None:
tokenizer = load_pretrained_tokenizer(opt.pretrained_tokenizer,
opt.cache_dir,
opt.special_vocab_path)
setattr(opt, 'vocab_size', len(tokenizer))
if opt.data_type == 'news':
fields = reload_news_fields(opt, tokenizer=tokenizer)
# elif opt.data_type == 'keyphrase':
# fields = reload_keyphrase_fields(opt, tokenizer=tokenizer)
if len(opt.data_ids) > 1:
# added by @memray, for loading multiple datasets
if opt.multi_dataset:
shard_base = "train"
train_iter = build_dataset_iter(shard_base,
fields,
opt,
multi=True)
else:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
nb_gpu = len(opt.gpu_ranks)
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
# Create a thread to listen for errors in the child processes.
error_queue = mp.SimpleQueue()
error_handler = ErrorHandler(error_queue)
# Train with multiprocessing.
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
procs.append(
mp.Process(target=run,
args=(opt, device_id, error_queue, q, semaphore),
daemon=True))
procs[device_id].start()
logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
producer = mp.Process(target=batch_producer,
args=(
train_iter,
queues,
semaphore,
opt,
),
daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs:
p.join()
producer.terminate()
elif nb_gpu == 1: # case 1 GPU only
single_main(opt, 0)
else: # case only CPU
single_main(opt, -1)
def main(opt,
device_id,
batch_queue=None,
semaphore=None,
train_iter=None,
passed_fields=None):
# NOTE: It's important that ``opt`` has been validated and updated
# at this point.
configure_process(opt, device_id)
init_logger(opt.log_file)
assert len(opt.accum_count) == len(opt.accum_steps), \
'Number of accum_count values must match number of accum_steps'
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
if opt.use_opt_from_trained:
model_opt = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
else:
model_opt = opt
ArgumentParser.validate_model_opts(model_opt)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
checkpoint = None
model_opt = opt
vocab = torch.load(opt.data + '.vocab.pt')
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
aux_fields = None
if passed_fields is not None:
fields = passed_fields['main']
aux_fields = passed_fields['crosslingual']
elif old_style_vocab(vocab):
fields = load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
fields = vocab
# Report src and tgt vocab sizes, including for features
for side in ['src', 'tgt']:
f = fields[side]
try:
f_iter = iter(f)
except TypeError:
f_iter = [(side, f)]
for sn, sf in f_iter:
if sf.use_vocab:
logger.info(' * %s vocab size = %d' % (sn, len(sf.vocab)))
# Build model.
model = build_model(model_opt,
opt,
fields,
checkpoint,
aux_fields=aux_fields)
n_params, enc, dec = _tally_parameters(model)
logger.info('encoder: %d' % enc)
logger.info('decoder: %d' % dec)
logger.info('* number of parameters: %d' % n_params)
_check_save_model_path(opt)
# Build optimizer.
if opt.almt_only:
almt = model.encoder.embeddings.almt_layers['mapping']
logger.info('Only training the alignment mapping.')
optim = Optimizer.from_opt(almt, opt, checkpoint=checkpoint)
else:
optim = Optimizer.from_opt(model, opt, checkpoint=checkpoint)
# Build model saver
model_saver = build_model_saver(model_opt,
opt,
model,
fields,
optim,
aux_fields=aux_fields)
trainer = build_trainer(opt,
device_id,
model,
fields,
optim,
model_saver=model_saver,
aux_fields=aux_fields)
if train_iter is not None:
pass # NOTE Use the passed one.
elif batch_queue is None:
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
else:
assert semaphore is not None, \
"Using batch_queue requires semaphore as well"
def _train_iter():
while True:
batch = batch_queue.get()
semaphore.release()
yield batch
train_iter = _train_iter()
cl_valid_iter = None
if opt.crosslingual:
valid_iter = build_dataset_iter("valid",
fields,
opt,
is_train=False,
task_cls=Eat2PlainMonoTask)
if opt.crosslingual_dev_data:
# NOTE I used 'train' to prepare this in `eat_prepare.sh`, so I use 'train' here as well.
cl_valid_iter = build_dataset_iter(
'train',
fields,
opt,
is_train=False,
data_attr='crosslingual_dev_data',
task_cls=Eat2PlainCrosslingualTask)
# NOTE This is for the second eat->plain task.
aux_valid_iter = build_dataset_iter('valid',
fields,
opt,
is_train=False,
data_attr='aux_train_data',
task_cls=Eat2PlainMonoTask)
valid_iters = [valid_iter, aux_valid_iter]
else:
valid_iters = [
build_dataset_iter("valid", fields, opt, is_train=False)
]
if len(opt.gpu_ranks):
logger.info('Starting training on GPU: %s' % opt.gpu_ranks)
else:
logger.info('Starting training on CPU, could be very slow')
train_steps = opt.train_steps
if opt.single_pass and train_steps > 0:
logger.warning("Option single_pass is enabled, ignoring train_steps.")
train_steps = 0
trainer.train(train_iter,
train_steps,
save_checkpoint_steps=opt.save_checkpoint_steps,
valid_iters=valid_iters,
valid_steps=opt.valid_steps,
cl_valid_iter=cl_valid_iter)
if opt.tensorboard:
trainer.report_manager.tensorboard_writer.close()
def train_single(self, output_model_dir: Path, opt, device_id, batch_queue=None, semaphore=None):
from roosterize.ml.onmt.CustomTrainer import CustomTrainer
from onmt.inputters.inputter import build_dataset_iter, load_old_vocab, old_style_vocab, build_dataset_iter_multiple
from onmt.model_builder import build_model
from onmt.train_single import configure_process, _tally_parameters, _check_save_model_path
from onmt.models import build_model_saver
from onmt.utils.optimizers import Optimizer
from onmt.utils.parse import ArgumentParser
configure_process(opt, device_id)
assert len(opt.accum_count) == len(opt.accum_steps), 'Number of accum_count values must match number of accum_steps'
# Load checkpoint if we resume from a previous training.
if opt.train_from:
self.logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from, map_location=lambda storage, loc: storage)
model_opt = ArgumentParser.ckpt_model_opts(checkpoint["opt"])
ArgumentParser.update_model_opts(model_opt)
ArgumentParser.validate_model_opts(model_opt)
self.logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
checkpoint = None
model_opt = opt
vocab = torch.load(opt.data + '.vocab.pt')
# end if
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(vocab, opt.model_type, dynamic_dict=opt.copy_attn)
else:
fields = vocab
# end if
# Report src and tgt vocab sizes, including for features
for side in ['src', 'tgt']:
f = fields[side]
try:
f_iter = iter(f)
except TypeError:
f_iter = [(side, f)]
# end try
for sn, sf in f_iter:
if sf.use_vocab: self.logger.info(' * %s vocab size = %d' % (sn, len(sf.vocab)))
# end for
# Build model
model = build_model(model_opt, opt, fields, checkpoint)
n_params, enc, dec = _tally_parameters(model)
self.logger.info('encoder: %d' % enc)
self.logger.info('decoder: %d' % dec)
self.logger.info('* number of parameters: %d' % n_params)
_check_save_model_path(opt)
# Build optimizer.
optim = Optimizer.from_opt(model, opt, checkpoint=checkpoint)
# Build model saver
model_saver = build_model_saver(model_opt, opt, model, fields, optim)
trainer = CustomTrainer.build_trainer(opt, device_id, model, fields, optim, model_saver=model_saver)
if batch_queue is None:
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
# end for
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
# end if
train_iter = build_dataset_iter(shard_base, fields, opt)
# end if
else:
assert semaphore is not None, "Using batch_queue requires semaphore as well"
def _train_iter():
while True:
batch = batch_queue.get()
semaphore.release()
yield batch
# end while
# end def
train_iter = _train_iter()
# end if
valid_iter = build_dataset_iter("valid", fields, opt, is_train=False)
if len(opt.gpu_ranks):
self.logger.info('Starting training on GPU: %s' % opt.gpu_ranks)
else:
self.logger.info('Starting training on CPU, could be very slow')
# end if
train_steps = opt.train_steps
if opt.single_pass and train_steps > 0:
self.logger.warning("Option single_pass is enabled, ignoring train_steps.")
train_steps = 0
# end if
trainer.train(
train_iter,
train_steps,
save_checkpoint_steps=opt.save_checkpoint_steps,
valid_iter=valid_iter,
valid_steps=opt.valid_steps)
time_begin = trainer.report_manager.start_time
time_end = time.time()
if opt.tensorboard: trainer.report_manager.tensorboard_writer.close()
# Dump train metrics
train_history = trainer.report_manager.get_joint_history()
train_metrics = {
"time_begin": time_begin,
"time_end": time_end,
"time": time_end - time_begin,
"train_history": train_history,
}
IOUtils.dump(output_model_dir/"train-metrics.json", train_metrics, IOUtils.Format.jsonNoSort)
# Get the best step, depending on the lowest val_xent (cross entropy)
best_loss = min([th["val_xent"] for th in train_history])
best_step = [th["step"] for th in train_history if th["val_xent"] == best_loss][-1] # Take the last if multiple
IOUtils.dump(output_model_dir/"best-step.json", best_step, IOUtils.Format.json)
return
def train_impl(self,
train_processed_data_dir: Path,
val_processed_data_dir: Path,
output_model_dir: Path,
) -> NoReturn:
from train import _get_parser as train_get_parser
from train import ErrorHandler, batch_producer
from onmt.inputters.inputter import old_style_vocab, load_old_vocab, build_dataset_iter, build_dataset_iter_multiple
import onmt.utils.distributed
from onmt.utils.parse import ArgumentParser
with IOUtils.cd(self.open_nmt_path):
parser = train_get_parser()
opt = parser.parse_args(
f" -data {output_model_dir}/processed-data"
f" -save_model {output_model_dir}/models/ckpt"
)
opt.gpu_ranks = [0]
opt.early_stopping = self.config.early_stopping_threshold
opt.report_every = 200
opt.valid_steps = 200
opt.save_checkpoint_steps = 200
opt.keep_checkpoint_max = self.config.ckpt_keep_max
opt.optim = "adam"
opt.learning_rate = self.config.learning_rate
opt.max_grad_norm = self.config.max_grad_norm
opt.batch_size = self.config.batch_size
opt.encoder_type = self.config.encoder
opt.decoder_type = self.config.decoder
opt.dropout = [self.config.dropout]
opt.src_word_vec_size = self.config.dim_embed
opt.tgt_word_vec_size = self.config.dim_embed
opt.layers = self.config.rnn_num_layers
opt.enc_rnn_size = self.config.dim_encoder_hidden
opt.dec_rnn_size = self.config.dim_decoder_hidden
if self.config.use_attn:
opt.global_attention = "general"
else:
opt.global_attention = "none"
# end if
if self.config.use_copy:
opt.copy_attn = True
opt.copy_attn_type = "general"
# end if
# train.main, one gpu case
ArgumentParser.validate_train_opts(opt)
ArgumentParser.update_model_opts(opt)
ArgumentParser.validate_model_opts(opt)
# Load checkpoint if we resume from a previous training.
if opt.train_from:
self.logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from, map_location=lambda storage, loc: storage)
self.logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
vocab = torch.load(opt.data + '.vocab.pt')
# end if
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(vocab, opt.model_type, dynamic_dict=opt.copy_attn)
else:
fields = vocab
# end if
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
# end for
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
# end if
train_iter = build_dataset_iter(shard_base, fields, opt)
# end if
nb_gpu = len(opt.gpu_ranks)
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
# Create a thread to listen for errors in the child processes.
error_queue = mp.SimpleQueue()
error_handler = ErrorHandler(error_queue)
# Train with multiprocessing.
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
def run(opt, device_id, error_queue, batch_queue, semaphore):
""" run process """
try:
gpu_rank = onmt.utils.distributed.multi_init(opt, device_id)
if gpu_rank != opt.gpu_ranks[device_id]:
raise AssertionError("An error occurred in Distributed initialization")
self.train_single(opt, device_id, batch_queue, semaphore)
except KeyboardInterrupt:
pass # killed by parent, do nothing
except Exception:
# propagate exception to parent process, keeping original traceback
import traceback
error_queue.put((opt.gpu_ranks[device_id], traceback.format_exc()))
# end try
# end def
procs.append(mp.Process(target=run, args=(opt, device_id, error_queue, q, semaphore), daemon=True))
procs[device_id].start()
self.logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
# end for
producer = mp.Process(target=batch_producer,args=(train_iter, queues, semaphore, opt,), daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs: p.join()
producer.terminate()
elif nb_gpu == 1: # case 1 GPU only
self.train_single(output_model_dir, opt, 0)
else: # case only CPU
self.train_single(output_model_dir, opt, -1)
# end if
# end with
return
def main(opt):
ArgumentParser.validate_train_opts(opt)
ArgumentParser.update_model_opts(opt)
ArgumentParser.validate_model_opts(opt)
# Load checkpoint if we resume from a previous training.
aux_vocab = None
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
if opt.crosslingual:
aux_vocab = checkpoint['aux_vocab']
elif opt.crosslingual:
assert opt.crosslingual in ['old', 'lm']
vocab = torch.load(opt.data + '.vocab.pt')
aux_vocab = torch.load(opt.aux_train_data + '.vocab.pt')
else:
vocab = torch.load(opt.data + '.vocab.pt')
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
def get_fields(vocab):
if old_style_vocab(vocab):
return load_old_vocab(vocab,
opt.model_type,
dynamic_dict=opt.copy_attn)
else:
return vocab
fields = get_fields(vocab)
aux_fields = None
if opt.crosslingual:
aux_fields = get_fields(aux_vocab)
if opt.crosslingual:
if opt.crosslingual == 'old':
aeq(len(opt.eat_formats), 3)
fields_info = [
('train', fields, 'data', Eat2PlainMonoTask, 'base',
opt.eat_formats[0]),
('train', aux_fields, 'aux_train_data', Eat2PlainAuxMonoTask,
'aux', opt.eat_formats[1]),
('train', aux_fields, 'aux_train_data',
Eat2PlainCrosslingualTask, 'crosslingual', opt.eat_format[2])
]
else:
aeq(len(opt.eat_formats), 4)
fields_info = [
('train', fields, 'data', Eat2PlainMonoTask, 'base',
opt.eat_formats[0]),
('train', fields, 'data', EatLMMonoTask, 'lm',
opt.eat_formats[1]),
('train', aux_fields, 'aux_train_data', Eat2PlainAuxMonoTask,
'aux', opt.eat_formats[2]),
('train', aux_fields, 'aux_train_data', EatLMCrosslingualTask,
'crosslingual', opt.eat_formats[3])
]
train_iter = build_crosslingual_dataset_iter(fields_info, opt)
elif len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
nb_gpu = len(opt.gpu_ranks)
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
# Create a thread to listen for errors in the child processes.
error_queue = mp.SimpleQueue()
error_handler = ErrorHandler(error_queue)
# Train with multiprocessing.
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
procs.append(
mp.Process(target=run,
args=(opt, device_id, error_queue, q, semaphore),
daemon=True))
procs[device_id].start()
logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
producer = mp.Process(target=batch_producer,
args=(
train_iter,
queues,
semaphore,
opt,
),
daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs:
p.join()
producer.terminate()
else:
device_id = 0 if nb_gpu == 1 else -1
# NOTE Only pass train_iter in my crosslingual mode.
train_iter = train_iter if opt.crosslingual else None
passed_fields = {
'main': fields,
'crosslingual': aux_fields
} if opt.crosslingual else None
single_main(opt,
device_id,
train_iter=train_iter,
passed_fields=passed_fields)
def main(opt):
ArgumentParser.validate_train_opts(opt)
ArgumentParser.update_model_opts(opt)
ArgumentParser.validate_model_opts(opt)
# Load checkpoint if we resume from a previous training.
if opt.train_from:
logger.info('Loading checkpoint from %s' % opt.train_from)
checkpoint = torch.load(opt.train_from,
map_location=lambda storage, loc: storage)
logger.info('Loading vocab from checkpoint at %s.' % opt.train_from)
vocab = checkpoint['vocab']
else:
vocab = torch.load(opt.data + '.vocab.pt')
# check for code where vocab is saved instead of fields
# (in the future this will be done in a smarter way)
if old_style_vocab(vocab):
fields = load_old_vocab(
vocab, opt.model_type, dynamic_dict=opt.copy_attn)
else:
fields = vocab
# @memray: a temporary workaround, as well as train_single.py line 78
if opt.model_type == "keyphrase":
if opt.tgt_type in ["one2one", "multiple"]:
del fields['sep_indices']
else:
if 'sep_indices' not in fields:
sep_indices = Field(
use_vocab=False, dtype=torch.long,
postprocessing=make_tgt, sequential=False)
fields["sep_indices"] = sep_indices
if 'src_ex_vocab' not in fields:
src_ex_vocab = RawField()
fields["src_ex_vocab"] = src_ex_vocab
if len(opt.data_ids) > 1:
train_shards = []
for train_id in opt.data_ids:
shard_base = "train_" + train_id
train_shards.append(shard_base)
train_iter = build_dataset_iter_multiple(train_shards, fields, opt)
else:
if opt.data_ids[0] is not None:
shard_base = "train_" + opt.data_ids[0]
else:
shard_base = "train"
train_iter = build_dataset_iter(shard_base, fields, opt)
nb_gpu = len(opt.gpu_ranks)
print(os.environ['PATH'])
if opt.world_size > 1:
queues = []
mp = torch.multiprocessing.get_context('spawn')
semaphore = mp.Semaphore(opt.world_size * opt.queue_size)
# Create a thread to listen for errors in the child processes.
error_queue = mp.SimpleQueue()
error_handler = ErrorHandler(error_queue)
# Train with multiprocessing.
procs = []
for device_id in range(nb_gpu):
q = mp.Queue(opt.queue_size)
queues += [q]
procs.append(mp.Process(target=run, args=(
opt, device_id, error_queue, q, semaphore), daemon=True))
procs[device_id].start()
logger.info(" Starting process pid: %d " % procs[device_id].pid)
error_handler.add_child(procs[device_id].pid)
producer = mp.Process(target=batch_producer,
args=(train_iter, queues, semaphore, opt,),
daemon=True)
producer.start()
error_handler.add_child(producer.pid)
for p in procs:
p.join()
producer.terminate()
elif nb_gpu == 1: # case 1 GPU only
single_main(opt, 0)
else: # case only CPU
single_main(opt, -1)
