def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--bert_model",
default='bert-base-uncased',
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
'--task',
type=str,
default=None,
required=True,
help="Task code in {hotpot_open, hotpot_distractor, squad, nq}")
# Other parameters
parser.add_argument(
"--max_seq_length",
default=378,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=1,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=5,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam. (def: 5e-5)")
parser.add_argument("--num_train_epochs",
default=5.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument('--local_rank', default=-1, type=int)
# RNN graph retriever-specific parameters
parser.add_argument("--example_limit", default=None, type=int)
parser.add_argument("--max_para_num", default=10, type=int)
parser.add_argument(
"--neg_chunk",
default=8,
type=int,
help="The chunk size of negative examples during training (to "
"reduce GPU memory consumption with negative sampling)")
parser.add_argument(
"--eval_chunk",
default=100000,
type=int,
help=
"The chunk size of evaluation examples (to reduce RAM consumption during evaluation)"
)
parser.add_argument(
"--split_chunk",
default=300,
type=int,
help=
"The chunk size of BERT encoding during inference (to reduce GPU memory consumption)"
)
parser.add_argument('--train_file_path',
type=str,
default=None,
help="File path to the training data")
parser.add_argument('--dev_file_path',
type=str,
default=None,
help="File path to the eval data")
parser.add_argument('--beam', type=int, default=1, help="Beam size")
parser.add_argument('--min_select_num',
type=int,
default=1,
help="Minimum number of selected paragraphs")
parser.add_argument('--max_select_num',
type=int,
default=3,
help="Maximum number of selected paragraphs")
parser.add_argument(
"--use_redundant",
action='store_true',
help="Whether to use simulated seqs (only for training)")
parser.add_argument(
"--use_multiple_redundant",
action='store_true',
help="Whether to use multiple simulated seqs (only for training)")
parser.add_argument(
'--max_redundant_num',
type=int,
default=100000,
help=
"Whether to limit the number of the initial TF-IDF pool (only for open-domain eval)"
)
parser.add_argument(
"--no_links",
action='store_true',
help=
"Whether to omit any links (or in other words, only use TF-IDF-based paragraphs)"
)
parser.add_argument("--pruning_by_links",
action='store_true',
help="Whether to do pruning by links (and top 1)")
parser.add_argument(
"--expand_links",
action='store_true',
help=
"Whether to expand links with paragraphs in the same article (for NQ)")
parser.add_argument(
'--tfidf_limit',
type=int,
default=None,
help=
"Whether to limit the number of the initial TF-IDF pool (only for open-domain eval)"
)
parser.add_argument("--pred_file",
default=None,
type=str,
help="File name to write paragraph selection results")
parser.add_argument("--tagme",
action='store_true',
help="Whether to use tagme at inference")
parser.add_argument(
'--topk',
type=int,
default=2,
help="Whether to use how many paragraphs from the previous steps")
parser.add_argument(
"--model_suffix",
default=None,
type=str,
help="Suffix to load a model file ('pytorch_model_' + suffix +'.bin')")
parser.add_argument("--db_save_path",
default=None,
type=str,
help="File path to DB")
parser.add_argument("--fp16", default=False, action='store_true')
parser.add_argument("--fp16_opt_level", default="O1", type=str)
parser.add_argument("--do_label",
default=False,
action='store_true',
help="For pre-processing features only.")
parser.add_argument("--oss_cache_dir", default=None, type=str)
parser.add_argument("--cache_dir", default=None, type=str)
parser.add_argument("--dist",
default=False,
action='store_true',
help='use distributed training.')
parser.add_argument("--save_steps", default=5000, type=int)
parser.add_argument("--resume", default=None, type=int)
parser.add_argument("--oss_pretrain", default=None, type=str)
parser.add_argument("--model_version", default='v1', type=str)
parser.add_argument("--disable_rnn_layer_norm",
default=False,
action='store_true')
args = parser.parse_args()
if args.dist:
dist.init_process_group(backend='nccl')
print(f"local rank: {args.local_rank}")
print(f"global rank: {dist.get_rank()}")
print(f"world size: {dist.get_world_size()}")
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of synchronizing nodes/GPUs
dist.init_process_group(backend='nccl')
if args.dist:
global_rank = dist.get_rank()
world_size = dist.get_world_size()
if world_size > 1:
args.local_rank = global_rank
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if args.train_file_path is not None:
do_train = True
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".
format(args.output_dir))
if args.local_rank in [-1, 0]:
os.makedirs(args.output_dir, exist_ok=True)
elif args.dev_file_path is not None:
do_train = False
else:
raise ValueError(
'One of train_file_path: {} or dev_file_path: {} must be non-None'.
format(args.train_file_path, args.dev_file_path))
processor = DataProcessor()
# Configurations of the graph retriever
graph_retriever_config = GraphRetrieverConfig(
example_limit=args.example_limit,
task=args.task,
max_seq_length=args.max_seq_length,
max_select_num=args.max_select_num,
max_para_num=args.max_para_num,
tfidf_limit=args.tfidf_limit,
train_file_path=args.train_file_path,
use_redundant=args.use_redundant,
use_multiple_redundant=args.use_multiple_redundant,
max_redundant_num=args.max_redundant_num,
dev_file_path=args.dev_file_path,
beam=args.beam,
min_select_num=args.min_select_num,
no_links=args.no_links,
pruning_by_links=args.pruning_by_links,
expand_links=args.expand_links,
eval_chunk=args.eval_chunk,
tagme=args.tagme,
topk=args.topk,
db_save_path=args.db_save_path,
disable_rnn_layer_norm=args.disable_rnn_layer_norm)
logger.info(graph_retriever_config)
logger.info(args)
tokenizer = AutoTokenizer.from_pretrained(args.bert_model)
if args.model_version == 'roberta':
from modeling_graph_retriever_roberta import RobertaForGraphRetriever
elif args.model_version == 'v3':
from modeling_graph_retriever_roberta import RobertaForGraphRetrieverIterV3 as RobertaForGraphRetriever
else:
raise RuntimeError()
##############################
# Training #
##############################
if do_train:
_model_state_dict = None
if args.oss_pretrain is not None:
_model_state_dict = torch.load(load_pretrain_from_oss(
args.oss_pretrain),
map_location='cpu')
logger.info(f"Loaded pretrained model from {args.oss_pretrain}")
if args.resume is not None:
_model_state_dict = torch.load(load_buffer_from_oss(
os.path.join(args.oss_cache_dir,
f"pytorch_model_{args.resume}.bin")),
map_location='cpu')
model = RobertaForGraphRetriever.from_pretrained(
args.bert_model,
graph_retriever_config=graph_retriever_config,
state_dict=_model_state_dict)
model.to(device)
global_step = 0
POSITIVE = 1.0
NEGATIVE = 0.0
_cache_file_name = f"cache_roberta_train_{args.max_seq_length}_{args.max_para_num}"
_examples_cache_file_name = f"examples_{_cache_file_name}"
_features_cache_file_name = f"features_{_cache_file_name}"
# Load training examples
logger.info(f"Loading training examples and features.")
try:
if args.cache_dir is not None and os.path.exists(
os.path.join(args.cache_dir, _features_cache_file_name)):
logger.info(
f"Loading pre-processed features from {os.path.join(args.cache_dir, _features_cache_file_name)}"
)
train_features = torch.load(
os.path.join(args.cache_dir, _features_cache_file_name))
else:
# train_examples = torch.load(load_buffer_from_oss(os.path.join(oss_features_cache_dir,
# _examples_cache_file_name)))
train_features = torch.load(
load_buffer_from_oss(
os.path.join(oss_features_cache_dir,
_features_cache_file_name)))
logger.info(
f"Pre-processed features are loaded from oss: "
f"{os.path.join(oss_features_cache_dir, _features_cache_file_name)}"
)
except:
train_examples = processor.get_train_examples(
graph_retriever_config)
train_features = convert_examples_to_features(
train_examples,
args.max_seq_length,
args.max_para_num,
graph_retriever_config,
tokenizer,
train=True)
logger.info(
f"Saving pre-processed features into oss: {oss_features_cache_dir}"
)
torch_save_to_oss(
train_examples,
os.path.join(oss_features_cache_dir,
_examples_cache_file_name))
torch_save_to_oss(
train_features,
os.path.join(oss_features_cache_dir,
_features_cache_file_name))
if args.do_label:
logger.info("Finished.")
return
# len(train_examples) and len(train_features) can be different, depending on the redundant setting
num_train_steps = int(
len(train_features) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight', 'layer_norm']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // dist.get_world_size()
optimizer = AdamW(optimizer_grouped_parameters,
betas=(0.9, 0.98),
lr=args.learning_rate)
scheduler = get_linear_schedule_with_warmup(
optimizer, int(t_total * args.warmup_proportion), t_total)
logger.info(optimizer)
if args.fp16:
from apex import amp
amp.register_half_function(torch, "einsum")
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
if args.local_rank != -1:
if args.fp16_opt_level == 'O2':
try:
import apex
model = apex.parallel.DistributedDataParallel(
model, delay_allreduce=True)
except ImportError:
model = torch.nn.parallel.DistributedDataParallel(
model, find_unused_parameters=True)
else:
model = torch.nn.parallel.DistributedDataParallel(
model, find_unused_parameters=True)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
if args.resume is not None:
_amp_state_dict = os.path.join(args.oss_cache_dir,
f"amp_{args.resume}.bin")
_optimizer_state_dict = os.path.join(
args.oss_cache_dir, f"optimizer_{args.resume}.pt")
_scheduler_state_dict = os.path.join(
args.oss_cache_dir, f"scheduler_{args.resume}.pt")
amp.load_state_dict(
torch.load(load_buffer_from_oss(_amp_state_dict)))
optimizer.load_state_dict(
torch.load(load_buffer_from_oss(_optimizer_state_dict)))
scheduler.load_state_dict(
torch.load(load_buffer_from_oss(_scheduler_state_dict)))
logger.info(f"Loaded resumed state dict of step {args.resume}")
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_features))
logger.info(" Instantaneous batch size per GPU = %d",
args.train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(dist.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
model.train()
epc = 0
# test
if args.local_rank in [-1, 0]:
if args.fp16:
amp_file = os.path.join(args.oss_cache_dir,
f"amp_{global_step}.bin")
torch_save_to_oss(amp.state_dict(), amp_file)
optimizer_file = os.path.join(args.oss_cache_dir,
f"optimizer_{global_step}.pt")
torch_save_to_oss(optimizer.state_dict(), optimizer_file)
scheduler_file = os.path.join(args.oss_cache_dir,
f"scheduler_{global_step}.pt")
torch_save_to_oss(scheduler.state_dict(), scheduler_file)
tr_loss = 0
for _ in range(int(args.num_train_epochs)):
logger.info('Epoch ' + str(epc + 1))
TOTAL_NUM = len(train_features)
train_start_index = 0
CHUNK_NUM = 8
train_chunk = TOTAL_NUM // CHUNK_NUM
chunk_index = 0
random.shuffle(train_features)
save_retry = False
while train_start_index < TOTAL_NUM:
train_end_index = min(train_start_index + train_chunk - 1,
TOTAL_NUM - 1)
chunk_len = train_end_index - train_start_index + 1
if args.resume is not None and global_step < args.resume:
_chunk_steps = int(
math.ceil(chunk_len * 1.0 / args.train_batch_size /
(1 if args.local_rank == -1 else
dist.get_world_size())))
_chunk_steps = _chunk_steps // args.gradient_accumulation_steps
if global_step + _chunk_steps <= args.resume:
global_step += _chunk_steps
train_start_index = train_end_index + 1
continue
train_features_ = train_features[
train_start_index:train_start_index + chunk_len]
all_input_ids = torch.tensor(
[f.input_ids for f in train_features_], dtype=torch.long)
all_input_masks = torch.tensor(
[f.input_masks for f in train_features_], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in train_features_], dtype=torch.long)
all_output_masks = torch.tensor(
[f.output_masks for f in train_features_],
dtype=torch.float)
all_num_paragraphs = torch.tensor(
[f.num_paragraphs for f in train_features_],
dtype=torch.long)
all_num_steps = torch.tensor(
[f.num_steps for f in train_features_], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_masks,
all_segment_ids, all_output_masks,
all_num_paragraphs, all_num_steps)
if args.local_rank != -1:
train_sampler = torch.utils.data.DistributedSampler(
train_data)
else:
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size,
pin_memory=True,
num_workers=4)
if args.local_rank != -1:
train_dataloader.sampler.set_epoch(epc)
logger.info('Examples from ' + str(train_start_index) +
' to ' + str(train_end_index))
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])):
if args.resume is not None and global_step < args.resume:
if (step + 1) % args.gradient_accumulation_steps == 0:
global_step += 1
continue
input_masks = batch[1]
batch_max_len = input_masks.sum(dim=2).max().item()
num_paragraphs = batch[4]
batch_max_para_num = num_paragraphs.max().item()
num_steps = batch[5]
batch_max_steps = num_steps.max().item()
# output_masks_cpu = (batch[3])[:, :batch_max_steps, :batch_max_para_num + 1]
batch = tuple(t.to(device) for t in batch)
input_ids, input_masks, segment_ids, output_masks, _, _ = batch
B = input_ids.size(0)
input_ids = input_ids[:, :batch_max_para_num, :
batch_max_len]
input_masks = input_masks[:, :batch_max_para_num, :
batch_max_len]
segment_ids = segment_ids[:, :batch_max_para_num, :
batch_max_len]
output_masks = output_masks[:, :batch_max_steps, :
batch_max_para_num +
1] # 1 for EOE
target = torch.zeros(output_masks.size()).fill_(
NEGATIVE) # (B, NUM_STEPS, |P|+1) <- 1 for EOE
for i in range(B):
output_masks[i, :num_steps[i], -1] = 1.0 # for EOE
for j in range(num_steps[i].item() - 1):
target[i, j, j].fill_(POSITIVE)
target[i, num_steps[i] - 1, -1].fill_(POSITIVE)
target = target.to(device)
neg_start = batch_max_steps - 1
while neg_start < batch_max_para_num:
neg_end = min(neg_start + args.neg_chunk - 1,
batch_max_para_num - 1)
neg_len = (neg_end - neg_start + 1)
input_ids_ = torch.cat(
(input_ids[:, :batch_max_steps - 1, :],
input_ids[:, neg_start:neg_start + neg_len, :]),
dim=1)
input_masks_ = torch.cat(
(input_masks[:, :batch_max_steps - 1, :],
input_masks[:, neg_start:neg_start + neg_len, :]),
dim=1)
segment_ids_ = torch.cat(
(segment_ids[:, :batch_max_steps - 1, :],
segment_ids[:, neg_start:neg_start + neg_len, :]),
dim=1)
output_masks_ = torch.cat(
(output_masks[:, :, :batch_max_steps - 1],
output_masks[:, :, neg_start:neg_start + neg_len],
output_masks[:, :, batch_max_para_num:
batch_max_para_num + 1]),
dim=2)
target_ = torch.cat(
(target[:, :, :batch_max_steps - 1],
target[:, :, neg_start:neg_start + neg_len],
target[:, :,
batch_max_para_num:batch_max_para_num +
1]),
dim=2)
if neg_start != batch_max_steps - 1:
output_masks_[:, :, :batch_max_steps - 1] = 0.0
output_masks_[:, :, -1] = 0.0
loss = model(input_ids_, segment_ids_, input_masks_,
output_masks_, target_, batch_max_steps)
if n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss,
optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
neg_start = neg_end + 1
# del input_ids_
# del input_masks_
# del segment_ids_
# del output_masks_
# del target_
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), 1.0)
else:
torch.nn.utils.clip_grad_norm_(
model.parameters(), 1.0)
optimizer.step()
scheduler.step()
# optimizer.zero_grad()
model.zero_grad()
global_step += 1
if global_step % 50 == 0:
_cur_steps = global_step if args.resume is None else global_step - args.resume
logger.info(
f"Training loss: {tr_loss / _cur_steps}\t"
f"Learning rate: {scheduler.get_lr()[0]}\t"
f"Global step: {global_step}")
if global_step % args.save_steps == 0:
if args.local_rank in [-1, 0]:
model_to_save = model.module if hasattr(
model, 'module') else model
output_model_file = os.path.join(
args.oss_cache_dir,
f"pytorch_model_{global_step}.bin")
torch_save_to_oss(model_to_save.state_dict(),
output_model_file)
_suffix = "" if args.local_rank == -1 else f"_{args.local_rank}"
if args.fp16:
amp_file = os.path.join(
args.oss_cache_dir,
f"amp_{global_step}{_suffix}.bin")
torch_save_to_oss(amp.state_dict(), amp_file)
optimizer_file = os.path.join(
args.oss_cache_dir,
f"optimizer_{global_step}{_suffix}.pt")
torch_save_to_oss(optimizer.state_dict(),
optimizer_file)
scheduler_file = os.path.join(
args.oss_cache_dir,
f"scheduler_{global_step}{_suffix}.pt")
torch_save_to_oss(scheduler.state_dict(),
scheduler_file)
logger.info(
f"checkpoint of step {global_step} is saved to oss."
)
# del input_ids
# del input_masks
# del segment_ids
# del output_masks
# del target
# del batch
chunk_index += 1
train_start_index = train_end_index + 1
# Save the model at the half of the epoch
if (chunk_index == CHUNK_NUM // 2
or save_retry) and args.local_rank in [-1, 0]:
status = save(model, args.output_dir, str(epc + 0.5))
save_retry = (not status)
del train_features_
del all_input_ids
del all_input_masks
del all_segment_ids
del all_output_masks
del all_num_paragraphs
del all_num_steps
del train_data
del train_sampler
del train_dataloader
gc.collect()
# Save the model at the end of the epoch
if args.local_rank in [-1, 0]:
save(model, args.output_dir, str(epc + 1))
# model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
# output_model_file = os.path.join(args.oss_cache_dir, "pytorch_model_" + str(epc + 1) + ".bin")
# torch_save_to_oss(model_to_save.state_dict(), output_model_file)
epc += 1
if do_train:
return
##############################
# Evaluation #
##############################
assert args.model_suffix is not None
if graph_retriever_config.db_save_path is not None:
import sys
sys.path.append('../')
from pipeline.tfidf_retriever import TfidfRetriever
tfidf_retriever = TfidfRetriever(graph_retriever_config.db_save_path,
None)
else:
tfidf_retriever = None
if args.oss_cache_dir is not None:
file_name = 'pytorch_model_' + args.model_suffix + '.bin'
model_state_dict = torch.load(
load_buffer_from_oss(os.path.join(args.oss_cache_dir, file_name)))
else:
model_state_dict = load(args.output_dir, args.model_suffix)
model = RobertaForGraphRetriever.from_pretrained(
args.bert_model,
state_dict=model_state_dict,
graph_retriever_config=graph_retriever_config)
model.to(device)
model.eval()
if args.pred_file is not None:
pred_output = []
eval_examples = processor.get_dev_examples(graph_retriever_config)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
TOTAL_NUM = len(eval_examples)
eval_start_index = 0
while eval_start_index < TOTAL_NUM:
eval_end_index = min(
eval_start_index + graph_retriever_config.eval_chunk - 1,
TOTAL_NUM - 1)
chunk_len = eval_end_index - eval_start_index + 1
eval_features = convert_examples_to_features(
eval_examples[eval_start_index:eval_start_index + chunk_len],
args.max_seq_length, args.max_para_num, graph_retriever_config,
tokenizer)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_masks = torch.tensor([f.input_masks for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_output_masks = torch.tensor(
[f.output_masks for f in eval_features], dtype=torch.float)
all_num_paragraphs = torch.tensor(
[f.num_paragraphs for f in eval_features], dtype=torch.long)
all_num_steps = torch.tensor([f.num_steps for f in eval_features],
dtype=torch.long)
all_ex_indices = torch.tensor([f.ex_index for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_masks,
all_segment_ids, all_output_masks,
all_num_paragraphs, all_num_steps,
all_ex_indices)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
for input_ids, input_masks, segment_ids, output_masks, num_paragraphs, num_steps, ex_indices in tqdm(
eval_dataloader, desc="Evaluating"):
batch_max_len = input_masks.sum(dim=2).max().item()
batch_max_para_num = num_paragraphs.max().item()
batch_max_steps = num_steps.max().item()
input_ids = input_ids[:, :batch_max_para_num, :batch_max_len]
input_masks = input_masks[:, :batch_max_para_num, :batch_max_len]
segment_ids = segment_ids[:, :batch_max_para_num, :batch_max_len]
output_masks = output_masks[:, :batch_max_para_num +
2, :batch_max_para_num + 1]
output_masks[:, 1:, -1] = 1.0 # Ignore EOE in the first step
input_ids = input_ids.to(device)
input_masks = input_masks.to(device)
segment_ids = segment_ids.to(device)
output_masks = output_masks.to(device)
examples = [
eval_examples[eval_start_index + ex_indices[i].item()]
for i in range(input_ids.size(0))
]
with torch.no_grad():
pred, prob, topk_pred, topk_prob = model.beam_search(
input_ids,
segment_ids,
input_masks,
examples=examples,
tokenizer=tokenizer,
retriever=tfidf_retriever,
split_chunk=args.split_chunk)
for i in range(len(pred)):
e = examples[i]
titles = [e.title_order[p] for p in pred[i]]
# Output predictions to a file
if args.pred_file is not None:
pred_output.append({})
pred_output[-1]['q_id'] = e.guid
pred_output[-1]['titles'] = titles
pred_output[-1]['probs'] = []
for prob_ in prob[i]:
entry = {'EOE': prob_[-1]}
for j in range(len(e.title_order)):
entry[e.title_order[j]] = prob_[j]
pred_output[-1]['probs'].append(entry)
topk_titles = [[e.title_order[p] for p in topk_pred[i][j]]
for j in range(len(topk_pred[i]))]
pred_output[-1]['topk_titles'] = topk_titles
topk_probs = []
for k in range(len(topk_prob[i])):
topk_probs.append([])
for prob_ in topk_prob[i][k]:
entry = {'EOE': prob_[-1]}
for j in range(len(e.title_order)):
entry[e.title_order[j]] = prob_[j]
topk_probs[-1].append(entry)
pred_output[-1]['topk_probs'] = topk_probs
# Output the selected paragraphs
context = {}
for ts in topk_titles:
for t in ts:
context[t] = e.all_paras[t]
pred_output[-1]['context'] = context
eval_start_index = eval_end_index + 1
del eval_features
del all_input_ids
del all_input_masks
del all_segment_ids
del all_output_masks
del all_num_paragraphs
del all_num_steps
del all_ex_indices
del eval_data
if args.pred_file is not None:
json.dump(pred_output, open(args.pred_file, 'w'))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model checkpoints and predictions will be written."
)
# Optimizer parameters
parser.add_argument("--adam_epsilon",
default=1e-6,
type=float,
help="Epsilon for Adam optimizer.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--adam_betas", default="(0.9, 0.999)", type=str)
parser.add_argument("--no_bias_correction",
default=False,
action='store_true')
# Other parameters
parser.add_argument("--train_file",
default=None,
type=str,
help="SQuAD-format json file for training.")
parser.add_argument("--predict_file",
default=None,
type=str,
help="SQuAD-format json file for evaluation.")
parser.add_argument(
"--max_seq_length",
default=384,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded."
)
parser.add_argument(
"--doc_stride",
default=128,
type=int,
help=
"When splitting up a long document into chunks, how much stride to take between chunks."
)
parser.add_argument(
"--max_query_length",
default=64,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_predict",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_label", action='store_true')
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--predict_batch_size",
default=8,
type=int,
help="Total batch size for predictions.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10%% "
"of training.")
parser.add_argument(
"--n_best_size",
default=20,
type=int,
help=
"The total number of n-best predictions to generate in the nbest_predictions.json "
"output file.")
parser.add_argument(
"--max_answer_length",
default=30,
type=int,
help=
"The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument(
"--verbose_logging",
action='store_true',
help=
"If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--fp16_opt_level', default='O1', type=str)
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument(
'--version_2_with_negative',
action='store_true',
help=
'If true, the SQuAD examples contain some that do not have an answer.')
parser.add_argument(
'--null_score_diff_threshold',
type=float,
default=0.0,
help=
"If null_score - best_non_null is greater than the threshold predict null."
)
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument(
'--no_masking',
action='store_true',
help='If true, we do not mask the span loss for no-answer examples.')
parser.add_argument(
'--skip_negatives',
action='store_true',
help=
'If true, we skip negative examples during training; this is mainly for ablation.'
)
# For Natural Questions
parser.add_argument(
'--max_answer_len',
type=int,
default=1000000,
help=
"maximum length of answer tokens (might be set to 5 for Natural Questions!)"
)
# balance the two losses.
parser.add_argument(
'--lambda_scale',
type=float,
default=1.0,
help=
"If you would like to change the two losses, please change the lambda scale."
)
# Save checkpoints more
parser.add_argument(
'--save_gran',
type=str,
default="10,3",
help='"10,5" means saving a checkpoint every 1/10 of the total updates,'
'but start saving from the 5th attempt')
parser.add_argument('--oss_cache_dir', default=None, type=str)
parser.add_argument('--cache_dir', default=None, type=str)
parser.add_argument('--dist', default=False, action='store_true')
args = parser.parse_args()
print(args)
if args.dist:
dist.init_process_group(backend='nccl')
print(f"local rank: {args.local_rank}")
print(f"global rank: {dist.get_rank()}")
print(f"world size: {dist.get_world_size()}")
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of
# synchronizing nodes/GPUs
dist.init_process_group(backend='nccl')
if args.dist:
global_rank = dist.get_rank()
world_size = dist.get_world_size()
if world_size > 1:
args.local_rank = global_rank
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# Prepare model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
model = IterBertForQuestionAnsweringConfidence.from_pretrained(
args.bert_model,
num_labels=4,
no_masking=args.no_masking,
lambda_scale=args.lambda_scale)
model.to(device)
train_examples = None
train_features = None
num_train_optimization_steps = None
if args.do_train:
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}_{4}'.format(
model.base_model_prefix, str(args.max_seq_length),
str(args.doc_stride), str(args.max_query_length),
tokenizer.do_lower_case)
cached_train_features_file_name = cached_train_features_file.split(
'/')[-1]
_oss_feature_save_path = os.path.join(oss_features_cache_dir,
cached_train_features_file_name)
try:
if args.cache_dir is not None and os.path.exists(
os.path.join(args.cache_dir,
cached_train_features_file_name)):
logger.info(
f"Loading pre-processed features from {os.path.join(args.cache_dir, cached_train_features_file_name)}"
)
train_features = torch.load(
os.path.join(args.cache_dir,
cached_train_features_file_name))
else:
logger.info(
f"Loading pre-processed features from oss: {_oss_feature_save_path}"
)
train_features = torch.load(
load_buffer_from_oss(_oss_feature_save_path))
except:
train_examples = read_squad_examples(
input_file=args.train_file,
is_training=True,
version_2_with_negative=args.version_2_with_negative,
max_answer_len=args.max_answer_len,
skip_negatives=args.skip_negatives)
train_features = convert_examples_to_features_yes_no(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank in [-1, 0]:
torch_save_to_oss(train_features, _oss_feature_save_path)
logger.info(
f"Saving train features into oss: {_oss_feature_save_path}"
)
num_train_optimization_steps = int(
len(train_features) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.do_label:
logger.info("finished.")
return
if args.do_train:
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight', 'layer_norm']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
t_total = num_train_optimization_steps
if args.local_rank != -1:
t_total = t_total // dist.get_world_size()
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
betas=eval(args.adam_betas),
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer, int(t_total * args.warmup_proportion),
num_train_optimization_steps)
if args.fp16:
from apex import amp
if args.fp16_opt_level == 'O1':
amp.register_half_function(torch, "einsum")
if args.loss_scale == 0:
model, optimizer = amp.initialize(
model, optimizer, opt_level=args.fp16_opt_level)
else:
model, optimizer = amp.initialize(
model,
optimizer,
opt_level=args.fp16_opt_level,
loss_scale=args.loss_scale)
if args.local_rank != -1:
if args.fp16_opt_level == 'O2':
try:
import apex
model = apex.parallel.DistributedDataParallel(
model, delay_allreduce=True)
except ImportError:
model = torch.nn.parallel.DistributedDataParallel(
model, find_unused_parameters=True)
else:
model = torch.nn.parallel.DistributedDataParallel(
model, find_unused_parameters=True)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
global_step = 0
logger.info("***** Running training *****")
if train_examples:
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Instantaneous batch size per GPU = %d",
args.train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(dist.get_world_size() if args.local_rank != -1 else 1),
)
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
all_switches = torch.tensor([f.switch for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_start_positions,
all_end_positions, all_switches)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size,
pin_memory=True,
num_workers=4)
if args.save_gran is not None:
save_chunk, save_start = args.save_gran.split(',')
save_chunk = t_total // int(save_chunk)
save_start = int(save_start)
model.train()
tr_loss = 0
for _epc in trange(int(args.num_train_epochs), desc="Epoch"):
if args.local_rank != -1:
train_dataloader.sampler.set_epoch(_epc)
for step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])):
if n_gpu == 1:
# multi-gpu does scattering it-self
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, start_positions, end_positions, switches = batch
loss = model(input_ids=input_ids,
token_type_ids=segment_ids,
attention_mask=input_mask,
start_positions=start_positions,
end_positions=end_positions,
switch_list=switches)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
scheduler.step()
optimizer.zero_grad()
global_step += 1
if global_step % 50 == 0:
logger.info(f"Training loss: {tr_loss / global_step}\t"
f"Learning rate: {scheduler.get_lr()[0]}\t"
f"Global step: {global_step}")
if args.save_gran is not None and args.local_rank in [
-1, 0
]:
if (global_step % save_chunk == 0) and (
global_step // save_chunk >= save_start):
logger.info('Saving a checkpoint....')
output_dir_per_epoch = os.path.join(
args.output_dir,
str(global_step) + 'steps')
os.makedirs(output_dir_per_epoch)
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(
model, 'module'
) else model # Only save the model it-self
if args.oss_cache_dir is not None:
_oss_model_save_path = os.path.join(
args.oss_cache_dir, f"{global_step}steps")
torch_save_to_oss(
model_to_save.state_dict(),
_oss_model_save_path +
"/pytorch_model.bin")
model_to_save.save_pretrained(output_dir_per_epoch)
tokenizer.save_pretrained(output_dir_per_epoch)
logger.info('Done')
if args.do_train and (args.local_rank == -1 or dist.get_rank() == 0):
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
model_to_save.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
torch_save_to_oss(
model_to_save.state_dict(),
os.path.join(args.oss_cache_dir, "pytorch_model.bin"))
# Load a trained model and vocabulary that you have fine-tuned
# model = IterBertForQuestionAnsweringConfidence.from_pretrained(
# args.output_dir, num_labels=4, no_masking=args.no_masking)
tokenizer = AutoTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
if args.do_train is False and args.do_predict is True:
model = IterBertForQuestionAnsweringConfidence.from_pretrained(
args.output_dir, num_labels=4, no_masking=args.no_masking)
tokenizer = AutoTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
elif args.do_train is True and args.do_predict is True:
model = IterBertForQuestionAnsweringConfidence.from_pretrained(
args.output_dir, num_labels=4, no_masking=args.no_masking)
tokenizer = AutoTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
else:
model = IterBertForQuestionAnsweringConfidence.from_pretrained(
args.bert_model,
num_labels=4,
no_masking=args.no_masking,
lambda_scale=args.lambda_scale)
model.to(device)
if args.do_predict and (args.local_rank == -1 or dist.get_rank() == 0):
eval_examples = read_squad_examples(
input_file=args.predict_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative,
max_answer_len=args.max_answer_length,
skip_negatives=args.skip_negatives)
eval_features = convert_examples_to_features_yes_no(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.predict_batch_size)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(
eval_dataloader,
desc="Evaluating",
disable=args.local_rank not in [-1, 0]):
if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits, batch_switch_logits = model(
input_ids, segment_ids, input_mask)
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist()
end_logits = batch_end_logits[i].detach().cpu().tolist()
switch_logits = batch_switch_logits[i].detach().cpu().tolist()
eval_feature = eval_features[example_index.item()]
unique_id = int(eval_feature.unique_id)
all_results.append(
RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits,
switch_logits=switch_logits))
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds.json")
write_predictions_yes_no_no_empty_answer(
eval_examples, eval_features, all_results, args.n_best_size,
args.max_answer_length, args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative, args.null_score_diff_threshold,
args.no_masking)