def train_abs_multi(args):
""" Spawns 1 process per GPU """
init_logger()
nb_gpu = args.world_size
mp = torch.multiprocessing.get_context('spawn')
# 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 i in range(nb_gpu):
device_id = i
procs.append(
mp.Process(target=run,
args=(
args,
device_id,
error_queue,
),
daemon=True))
procs[i].start()
logger.info(" Starting process pid: %d " % procs[i].pid)
error_handler.add_child(procs[i].pid)
for p in procs:
p.join()
def test_abs(args, device_id, pt, step):
device = "cpu" if args.visible_gpus == '-1' else "cuda"
if (pt != ''):
test_from = pt
else:
test_from = args.test_from
logger.info('Loading checkpoint from %s' % test_from)
checkpoint = torch.load(test_from, map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
print(args)
model = AbsSummarizer(args, device, checkpoint)
model.eval()
test_iter = data_loader.Dataloader(args, load_dataset(args, 'test', shuffle=False),
args.test_batch_size, device,
shuffle=False, is_test=True)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True, cache_dir=args.temp_dir)
symbols = {'BOS': tokenizer.vocab['[unused0]'], 'EOS': tokenizer.vocab['[unused1]'],
'PAD': tokenizer.vocab['[PAD]'], 'EOQ': tokenizer.vocab['[unused2]']}
predictor = build_predictor(args, tokenizer, symbols, model, logger)
predictor.translate(test_iter, step)
def test_text_abs(args):
logger.info('Loading checkpoint from %s' % args.test_from)
device = "cpu" if args.visible_gpus == '-1' else "cuda"
checkpoint = torch.load(args.test_from,
map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
print(args)
model = AbsSummarizer(args, device, checkpoint)
model.eval()
test_iter = data_loader.load_text(args, args.text_src, args.text_tgt,
device)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased',
do_lower_case=True,
cache_dir=args.temp_dir)
symbols = {
'BOS': tokenizer.vocab['[unused0]'],
'EOS': tokenizer.vocab['[unused1]'],
'PAD': tokenizer.vocab['[PAD]'],
'EOQ': tokenizer.vocab['[unused2]']
}
predictor = build_predictor(args, tokenizer, symbols, model, logger)
predictor.translate(test_iter, -1)
def validate(args, device_id, pt, step):
device = "cpu" if args.visible_gpus == '-1' else "cuda"
if (pt != ''):
test_from = pt
else:
test_from = args.test_from
logger.info('Loading checkpoint from %s' % test_from)
checkpoint = torch.load(test_from, map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
print(args)
model = AbsSummarizer(args, device, checkpoint)
model.eval()
valid_iter = data_loader.Dataloader(args, load_dataset(args, 'valid', shuffle=False),
args.batch_size, device,
shuffle=False, is_test=False)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True, cache_dir=args.temp_dir)
symbols = {'BOS': tokenizer.vocab['[unused0]'], 'EOS': tokenizer.vocab['[unused1]'],
'PAD': tokenizer.vocab['[PAD]'], 'EOQ': tokenizer.vocab['[unused2]']}
valid_loss = abs_loss(model.generator, symbols, model.vocab_size, train=False, device=device)
trainer = build_trainer(args, device_id, model, None, valid_loss)
stats = trainer.validate(valid_iter, step)
return stats.xent()
def test_ext(args, device_id, pt, step):
device = "cpu" if args.visible_gpus == '-1' else "cuda"
if (pt != ''):
test_from = pt
else:
test_from = args.test_from
logger.info('Loading checkpoint from %s' % test_from)
checkpoint = torch.load(test_from,
map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
print(args)
model = ExtSummarizer(args, device, checkpoint)
model.eval()
test_iter = data_loader.Dataloader(args,
load_dataset(args,
'test',
shuffle=False),
args.test_batch_size,
device,
shuffle=False,
is_test=True)
trainer = build_trainer(args, device_id, model, None)
trainer.test(test_iter, step)
def validate_abs(args, device_id):
timestep = 0
if (args.test_all):
cp_files = sorted(
glob.glob(os.path.join(args.model_path, 'model_step_*.pt')))
cp_files.sort(key=os.path.getmtime)
xent_lst = []
for i, cp in enumerate(cp_files):
step = int(cp.split('.')[-2].split('_')[-1])
if (args.test_start_from != -1 and step < args.test_start_from):
xent_lst.append((1e6, cp))
continue
xent = validate(args, device_id, cp, step)
xent_lst.append((xent, cp))
max_step = xent_lst.index(min(xent_lst))
if (i - max_step > 10):
break
xent_lst = sorted(xent_lst, key=lambda x: x[0])[:5]
logger.info('PPL %s' % str(xent_lst))
for xent, cp in xent_lst:
step = int(cp.split('.')[-2].split('_')[-1])
test_abs(args, device_id, cp, step)
else:
while (True):
cp_files = sorted(
glob.glob(os.path.join(args.model_path, 'model_step_*.pt')))
cp_files.sort(key=os.path.getmtime)
if (cp_files):
cp = cp_files[-1]
time_of_cp = os.path.getmtime(cp)
if (not os.path.getsize(cp) > 0):
time.sleep(60)
continue
if (time_of_cp > timestep):
timestep = time_of_cp
step = int(cp.split('.')[-2].split('_')[-1])
validate(args, device_id, cp, step)
test_abs(args, device_id, cp, step)
cp_files = sorted(
glob.glob(os.path.join(args.model_path, 'model_step_*.pt')))
cp_files.sort(key=os.path.getmtime)
if (cp_files):
cp = cp_files[-1]
time_of_cp = os.path.getmtime(cp)
if (time_of_cp > timestep):
continue
else:
time.sleep(300)
def output(self, step, num_steps, learning_rate, start):
"""Write out statistics to stdout.
Args:
step (int): current step
n_batch (int): total batches
start (int): start time of step.
"""
t = self.elapsed_time()
logger.info(
("Step %2d/%5d; acc: %6.2f; ppl: %5.2f; xent: %4.2f; " +
"lr: %7.8f; %3.0f/%3.0f tok/s; %6.0f sec") %
(step, num_steps, self.accuracy(), self.ppl(), self.xent(),
learning_rate, self.n_src_words / (t + 1e-5), self.n_words /
(t + 1e-5), time.time() - start))
sys.stdout.flush()
def _save(self, step):
real_model = self.model
# real_generator = (self.generator.module
# if isinstance(self.generator, torch.nn.DataParallel)
# else self.generator)
model_state_dict = real_model.state_dict()
# generator_state_dict = real_generator.state_dict()
checkpoint = {
'model': model_state_dict,
# 'generator': generator_state_dict,
'opt': self.args,
'optim': self.optim,
}
checkpoint_path = os.path.join(self.args.model_path, 'model_step_%d.pt' % step)
logger.info("Saving checkpoint %s" % checkpoint_path)
# checkpoint_path = '%s_step_%d.pt' % (FLAGS.model_path, step)
if (not os.path.exists(checkpoint_path)):
torch.save(checkpoint, checkpoint_path)
return checkpoint, checkpoint_path
def output(self, step, num_steps, learning_rate, start):
"""Write out statistics to stdout.
Args:
step (int): current step
n_batch (int): total batches
start (int): start time of step.
"""
t = self.elapsed_time()
step_fmt = "%2d" % step
if num_steps > 0:
step_fmt = "%s/%5d" % (step_fmt, num_steps)
logger.info(
("Step %s; xent: %4.2f; " +
"lr: %7.7f; %3.0f docs/s; %6.0f sec")
% (step_fmt,
self.xent(),
learning_rate,
self.n_docs / (t + 1e-5),
time.time() - start))
sys.stdout.flush()
def build_trainer(args, device_id, model, optim):
"""
Simplify `Trainer` creation based on user `opt`s*
Args:
opt (:obj:`Namespace`): user options (usually from argument parsing)
model (:obj:`onmt.models.NMTModel`): the model to train
fields (dict): dict of fields
optim (:obj:`onmt.utils.Optimizer`): optimizer used during training
data_type (str): string describing the type of data
e.g. "text", "img", "audio"
model_saver(:obj:`onmt.models.ModelSaverBase`): the utility object
used to save the model
"""
grad_accum_count = args.accum_count
n_gpu = args.world_size
if device_id >= 0:
gpu_rank = int(args.gpu_ranks[device_id])
else:
gpu_rank = 0
n_gpu = 0
print('gpu_rank %d' % gpu_rank)
tensorboard_log_dir = args.model_path
writer = SummaryWriter(tensorboard_log_dir, comment="Unmt")
report_manager = ReportMgr(args.report_every, start_time=-1, tensorboard_writer=writer)
trainer = Trainer(args, model, optim, grad_accum_count, n_gpu, gpu_rank, report_manager)
# print(tr)
if (model):
n_params = _tally_parameters(model)
logger.info('* number of parameters: %d' % n_params)
return trainer
def test_text_ext(args):
logger.info('Loading checkpoint from %s' % args.test_from)
checkpoint = torch.load(args.test_from,
map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
#print(args)
device = "cpu" if args.visible_gpus == '-1' else "cuda"
device_id = 0 if device == "cuda" else -1
#print(device)
model = ExtSummarizer(args, device, checkpoint)
model.eval()
#print(model)
test_iter = data_loader.load_text(args, args.text_src, args.text_tgt,
device)
#print(data_loader)
trainer = build_trainer(args, device_id, model, None)
trainer.test(test_iter, -1)
def train_single_ext(args, device_id):
init_logger(args.log_file)
device = "cpu" if args.visible_gpus == '-1' else "cuda"
logger.info('Device ID %d' % device_id)
logger.info('Device %s' % device)
torch.manual_seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
if device_id >= 0:
torch.cuda.set_device(device_id)
torch.cuda.manual_seed(args.seed)
torch.manual_seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
if args.train_from != '':
logger.info('Loading checkpoint from %s' % args.train_from)
checkpoint = torch.load(args.train_from,
map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
else:
checkpoint = None
def train_iter_fct():
return data_loader.Dataloader(args,
load_dataset(args, 'train',
shuffle=True),
args.batch_size,
device,
shuffle=True,
is_test=False)
model = ExtSummarizer(args, device, checkpoint)
optim = model_builder.build_optim(args, model, checkpoint)
logger.info(model)
trainer = build_trainer(args, device_id, model, optim)
trainer.train(train_iter_fct, args.train_steps)
def train_abs_single(args, device_id):
init_logger(args.log_file)
logger.info(str(args))
device = "cpu" if args.visible_gpus == '-1' else "cuda"
logger.info('Device ID %d' % device_id)
logger.info('Device %s' % device)
torch.manual_seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
if device_id >= 0:
torch.cuda.set_device(device_id)
torch.cuda.manual_seed(args.seed)
if args.train_from != '':
logger.info('Loading checkpoint from %s' % args.train_from)
checkpoint = torch.load(args.train_from,
map_location=lambda storage, loc: storage)
opt = vars(checkpoint['opt'])
for k in opt.keys():
if (k in model_flags):
setattr(args, k, opt[k])
else:
checkpoint = None
if (args.load_from_extractive != ''):
logger.info('Loading bert from extractive model %s' %
args.load_from_extractive)
bert_from_extractive = torch.load(
args.load_from_extractive,
map_location=lambda storage, loc: storage)
bert_from_extractive = bert_from_extractive['model']
else:
bert_from_extractive = None
torch.manual_seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
def train_iter_fct():
return data_loader.Dataloader(args,
load_dataset(args, 'train',
shuffle=True),
args.batch_size,
device,
shuffle=True,
is_test=False)
model = AbsSummarizer(args, device, checkpoint, bert_from_extractive)
if (args.sep_optim):
optim_bert = model_builder.build_optim_bert(args, model, checkpoint)
optim_dec = model_builder.build_optim_dec(args, model, checkpoint)
optim = [optim_bert, optim_dec]
else:
optim = [model_builder.build_optim(args, model, checkpoint)]
logger.info(model)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased',
do_lower_case=True,
cache_dir=args.temp_dir)
symbols = {
'BOS': tokenizer.vocab['[unused0]'],
'EOS': tokenizer.vocab['[unused1]'],
'PAD': tokenizer.vocab['[PAD]'],
'EOQ': tokenizer.vocab['[unused2]']
}
train_loss = abs_loss(model.generator,
symbols,
model.vocab_size,
device,
train=True,
label_smoothing=args.label_smoothing)
trainer = build_trainer(args, device_id, model, optim, train_loss)
trainer.train(train_iter_fct, args.train_steps)
def log(self, *args, **kwargs):
logger.info(*args, **kwargs)
def test(self, test_iter, step, cal_lead=False, cal_oracle=False):
""" Validate model.
valid_iter: validate data iterator
Returns:
:obj:`nmt.Statistics`: validation loss statistics
"""
# Set model in validating mode.
def _get_ngrams(n, text):
ngram_set = set()
text_length = len(text)
max_index_ngram_start = text_length - n
for i in range(max_index_ngram_start + 1):
ngram_set.add(tuple(text[i:i + n]))
return ngram_set
def _block_tri(c, p):
tri_c = _get_ngrams(3, c.split())
for s in p:
tri_s = _get_ngrams(3, s.split())
if len(tri_c.intersection(tri_s)) > 0:
return True
return False
if (not cal_lead and not cal_oracle):
self.model.eval()
stats = Statistics()
can_path = '%s_step%d.candidate' % (self.args.result_path, step)
gold_path = '%s_step%d.gold' % (self.args.result_path, step)
with open(can_path, 'w') as save_pred:
with open(gold_path, 'w') as save_gold:
with torch.no_grad():
for batch in test_iter:
src = batch.src
segs = batch.segs
clss = batch.clss
mask = batch.mask_src
mask_cls = batch.mask_cls
gold = []
pred = []
if (cal_lead):
selected_ids = [list(range(batch.clss.size(1)))] * batch.batch_size
elif (cal_oracle):
labels = batch.src_sent_labels
selected_ids = [[j for j in range(batch.clss.size(1)) if labels[i][j] == 1] for i in
range(batch.batch_size)]
else:
sent_scores, mask = self.model(src, segs, clss, mask, mask_cls)
#print('sent scores - ')
#print(sent_scores)
sent_scores = sent_scores + mask.float()
sent_scores = sent_scores.cpu().data.numpy()
selected_ids = np.argsort(-sent_scores, 1)
if (hasattr(batch, 'src_sent_labels')):
labels = batch.src_sent_labels
loss = self.loss(sent_scores, labels.float())
loss = (loss * mask.float()).sum()
batch_stats = Statistics(float(loss.cpu().data.numpy()), len(labels))
stats.update(batch_stats)
for i, idx in enumerate(selected_ids):
_pred = []
if (len(batch.src_str[i]) == 0):
continue
for j in selected_ids[i][:len(batch.src_str[i])]:
if (j >= len(batch.src_str[i])):
continue
candidate = batch.src_str[i][j].strip()
if (self.args.block_trigram):
if (not _block_tri(candidate, _pred)):
_pred.append(candidate)
else:
_pred.append(candidate)
if ((not cal_oracle) and (not self.args.recall_eval) and len(_pred) == 5): #to change the number of sentences in the summary
break
_pred = '\n'.join(_pred)
if (self.args.recall_eval):
_pred = ' '.join(_pred.split()[:len(batch.tgt_str[i].split())])
pred.append(_pred)
gold.append(batch.tgt_str[i])
for i in range(len(gold)):
save_gold.write(gold[i].strip() + '\n')
for i in range(len(pred)):
save_pred.write(pred[i].strip() + '\n')
if (step != -1 and self.args.report_rouge):
rouges = test_rouge(self.args.temp_dir, can_path, gold_path)
logger.info('Rouges at step %d \n%s' % (step, rouge_results_to_str(rouges)))
self._report_step(0, step, valid_stats=stats)
return stats
def train(self, train_iter_fct, train_steps, valid_iter_fct=None, valid_steps=-1):
"""
The main training loops.
by iterating over training data (i.e. `train_iter_fct`)
and running validation (i.e. iterating over `valid_iter_fct`
Args:
train_iter_fct(function): a function that returns the train
iterator. e.g. something like
train_iter_fct = lambda: generator(*args, **kwargs)
valid_iter_fct(function): same as train_iter_fct, for valid data
train_steps(int):
valid_steps(int):
save_checkpoint_steps(int):
Return:
None
"""
logger.info('Start training...')
# step = self.optim._step + 1
step = self.optim._step + 1
true_batchs = []
accum = 0
normalization = 0
train_iter = train_iter_fct()
total_stats = Statistics()
report_stats = Statistics()
self._start_report_manager(start_time=total_stats.start_time)
while step <= train_steps:
reduce_counter = 0
for i, batch in enumerate(train_iter):
if self.n_gpu == 0 or (i % self.n_gpu == self.gpu_rank):
true_batchs.append(batch)
normalization += batch.batch_size
accum += 1
if accum == self.grad_accum_count:
reduce_counter += 1
if self.n_gpu > 1:
normalization = sum(distributed
.all_gather_list
(normalization))
self._gradient_accumulation(
true_batchs, normalization, total_stats,
report_stats)
report_stats = self._maybe_report_training(
step, train_steps,
self.optim.learning_rate,
report_stats)
true_batchs = []
accum = 0
normalization = 0
if (step % self.save_checkpoint_steps == 0 and self.gpu_rank == 0):
self._save(step)
step += 1
if step > train_steps:
break
train_iter = train_iter_fct()
return total_stats
