def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--ernie_model",
default=None,
type=str,
required=True,
help="Ernie pre-trained model")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
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_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
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.")
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("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
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(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
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('--threshold', type=float, default=.3)
parser.add_argument("--vec_file",
default=None,
type=str,
required=True,
help="File with embeddings")
parser.add_argument("--qid_file",
default=None,
type=str,
required=True,
help="File with qid mapping")
parser.add_argument("--use_lim_ents",
default=None,
type=str,
required=True,
help="Whether to use limited entities")
args = parser.parse_args()
processors = FewrelProcessor
num_labels_task = 80
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 sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
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))
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_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
processor = processors()
num_labels = num_labels_task
label_list = None
tokenizer = BertTokenizer.from_pretrained(args.ernie_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
train_examples, label_list = processor.get_train_examples(args.data_dir)
# check for limited ents
lim_ents = []
lim_check = (args.use_lim_ents == "y")
if lim_check:
lim_ents = lim_ent_map(0, "kg_embeddings/dbp_eid_2_wd_eid.txt")
logger.info(
"Limited entities flag is on. Count of unique entities considered: "
+ str(len(lim_ents)))
vecs = []
vecs.append([0] * 100) # CLS
lineindex = 1
uid_map = {}
logger.info("Reading embeddings file.")
with open(args.vec_file, 'r') as fin:
for line in fin:
vec = line.strip().split('\t')
# first element is unique id
uniqid = int(vec[0])
# map line index to unique id
uid_map[uniqid] = lineindex
# increment line index
lineindex = lineindex + 1
if (lim_check and (uniqid in lim_ents)) or not lim_check:
vec = [float(x) for x in vec[1:101]]
else:
vec = vecs[0]
vecs.append(vec)
embed = torch.FloatTensor(vecs)
embed = torch.nn.Embedding.from_pretrained(embed)
#embed = torch.nn.Embedding(5041175, 100)
logger.info("Shape of entity embedding: " + str(embed.weight.size()))
del vecs
filenames = os.listdir(args.output_dir)
filenames = [x for x in filenames if "pytorch_model.bin_" in x]
file_mark = []
for x in filenames:
file_mark.append([x, True])
file_mark.append([x, False])
eval_examples = processor.get_dev_examples(args.data_dir)
dev = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length, tokenizer,
args.threshold, args.qid_file)
eval_examples = processor.get_test_examples(args.data_dir)
test = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length, tokenizer,
args.threshold, args.qid_file)
for x, mark in file_mark:
print(x, mark)
output_model_file = os.path.join(args.output_dir, x)
model_state_dict = torch.load(output_model_file)
model, _ = BertForSequenceClassification.from_pretrained(
args.ernie_model,
state_dict=model_state_dict,
num_labels=len(label_list))
model.to(device)
if mark:
eval_features = dev
else:
eval_features = test
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
# zeros = [0 for _ in range(args.max_seq_length)]
# zeros_ent = [0 for _ in range(100)]
# zeros_ent = [zeros_ent for _ in range(args.max_seq_length)]
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_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
all_ent = torch.tensor([f.input_ent for f in eval_features],
dtype=torch.long)
all_ent_masks = torch.tensor([f.ent_mask for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_ent, all_ent_masks,
all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
if mark:
output_eval_file = os.path.join(
args.output_dir,
"eval_results_{}.txt".format(x.split("_")[-1]))
output_file_pred = os.path.join(
args.output_dir, "eval_pred_{}.txt".format(x.split("_")[-1]))
output_file_glod = os.path.join(
args.output_dir, "eval_gold_{}.txt".format(x.split("_")[-1]))
else:
output_eval_file = os.path.join(
args.output_dir,
"test_results_{}.txt".format(x.split("_")[-1]))
output_file_pred = os.path.join(
args.output_dir, "test_pred_{}.txt".format(x.split("_")[-1]))
output_file_glod = os.path.join(
args.output_dir, "test_gold_{}.txt".format(x.split("_")[-1]))
fpred = open(output_file_pred, "w")
fgold = open(output_file_glod, "w")
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, input_ent, ent_mask, label_ids in eval_dataloader:
input_ent = embed(input_ent + 1) # -1 -> 0
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
input_ent = input_ent.to(device)
ent_mask = ent_mask.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
input_ent, ent_mask, label_ids)
logits = model(input_ids, segment_ids, input_mask, input_ent,
ent_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy, pred = accuracy(logits, label_ids)
for a, b in zip(pred, label_ids):
fgold.write("{}\n".format(b))
fpred.write("{}\n".format(a))
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy}
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--ernie_model",
default=None,
type=str,
required=True,
help="Ernie pre-trained model")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
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_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
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("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
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(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
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('--threshold', type=float, default=.1)
args = parser.parse_args()
processors = SemevalProcessor
num_labels_task = 3
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 sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
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 = 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 not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
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.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
processor = processors()
num_labels = num_labels_task
label_list = None
tokenizer = BertTokenizer.from_pretrained(args.ernie_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
train_examples, label_list = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model, _ = BertForSequenceClassification.from_pretrained(
args.ernie_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank),
num_labels=num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_grad = [
'bert.encoder.layer.11.output.dense_ent',
'bert.encoder.layer.11.output.LayerNorm_ent'
]
param_optimizer = [(n, p) for n, p in param_optimizer
if not any(nd in n for nd in no_grad)]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
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 // torch.distributed.get_world_size()
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer,
args.threshold)
vecs = []
vecs.append([0] * 100)
logger.info("Loading entity embedding.")
with open("kg_embed/entity2vec.vec", 'r') as fin:
for line in fin:
vec = line.strip().split('\t')
vec = [float(x) for x in vec]
vecs.append(vec)
embed = torch.FloatTensor(vecs)
embed = torch.nn.Embedding.from_pretrained(embed)
# embed = torch.nn.Embedding(5041175, 100)
logger.info("Shape of entity embedding: " + str(embed.weight.size()))
del vecs
if args.do_eval:
eval_examples = processor.get_dev_examples(args.data_dir)
dev = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length, tokenizer,
args.threshold)
eval_features = dev
logger.info("Eval Num examples = %d", len(eval_examples))
logger.info("Eval Batch size = %d", args.train_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_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
all_ent = torch.tensor([f.input_ent for f in eval_features],
dtype=torch.long)
all_ent_masks = torch.tensor([f.ent_mask for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_ent, all_ent_masks,
all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.train_batch_size)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
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_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
all_ent = torch.tensor([f.input_ent for f in train_features],
dtype=torch.long)
all_ent_masks = torch.tensor([f.ent_mask for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_ent, all_ent_masks,
all_label_ids)
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)
output_loss_file = os.path.join(args.output_dir, "loss")
loss_fout = open(output_loss_file, 'w')
model.train()
max_acc = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(train_dataloader):
batch = tuple(
t.to(device) if i != 3 else t for i, t in enumerate(batch))
input_ids, input_mask, segment_ids, input_ent, ent_mask, label_ids = batch
input_ent = embed(input_ent + 1).to(device) # -1 -> 0
loss = model(input_ids, segment_ids, input_mask,
input_ent.half(), ent_mask, label_ids)
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:
optimizer.backward(loss)
else:
loss.backward()
loss_fout.write("{}\n".format(loss.item()))
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.do_eval:
logger.info("***** Running evaluation *****")
output_eval_file = os.path.join(
args.output_dir, "eval_results_{}.txt".format(global_step))
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, input_ent, ent_mask, label_ids in eval_dataloader:
input_ent = embed(input_ent + 1) # -1 -> 0
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
input_ent = input_ent.to(device)
ent_mask = ent_mask.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids,
input_mask, input_ent, ent_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask,
input_ent, ent_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
max_acc = max(max_acc, eval_accuracy)
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'max_accuracy': max_acc
}
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument("--ernie_model",
default=None,
type=str,
required=True,
help="Ernie pre-trained model")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
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_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=16,
type=int,
help="Total batch size for training.")
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("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
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(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
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('--threshold', type=float, default=.3)
parser.add_argument("--vec_file",
default=None,
type=str,
required=True,
help="File with embeddings")
parser.add_argument("--qid_file",
default=None,
type=str,
required=True,
help="File with qid mapping")
parser.add_argument("--use_lim_ents",
default=None,
type=str,
required=True,
help="Whether to use limited entities")
args = parser.parse_args()
processors = FewrelProcessor
num_labels_task = 80
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 sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
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 = 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 not args.do_train:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
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.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
processor = processors()
num_labels = num_labels_task
label_list = None
tokenizer = BertTokenizer.from_pretrained(args.ernie_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
train_examples, label_list = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model, _ = BertForSequenceClassification.from_pretrained(
args.ernie_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank),
num_labels=num_labels)
# if args.fp16:
# model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_grad = [
'bert.encoder.layer.11.output.dense_ent',
'bert.encoder.layer.11.output.LayerNorm_ent'
]
param_optimizer = [(n, p) for n, p in param_optimizer
if not any(nd in n for nd in no_grad)]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
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 // torch.distributed.get_world_size()
if args.fp16:
try:
#from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False)
if args.loss_scale == 0:
model, optimizer = amp.initialize(model, optimizer, opt_level="O2")
# optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(
train_examples, label_list, args.max_seq_length, tokenizer,
args.threshold, args.qid_file)
# check for limited ents
lim_ents = []
lim_check = (args.use_lim_ents == "y")
if lim_check:
lim_ents = lim_ent_map(0, "kg_embeddings/dbp_eid_2_wd_eid.txt")
logger.info(
"Limited entities flag is on. Count of unique entities considered: "
+ str(len(lim_ents)))
vecs = []
vecs.append([0] * 100) # CLS
lineindex = 1
uid_map = {}
logger.info("Reading embeddings file.")
with open(args.vec_file, 'r') as fin:
for line in fin:
vec = line.strip().split('\t')
# first element is unique id
uniqid = int(vec[0])
# map line index to unique id
uid_map[uniqid] = lineindex
# increment line index
lineindex = lineindex + 1
if (lim_check and (uniqid in lim_ents)) or not lim_check:
vec = [float(x) for x in vec[1:101]]
else:
vec = vecs[0]
vecs.append(vec)
embed = torch.FloatTensor(vecs)
embed = torch.nn.Embedding.from_pretrained(embed)
#embed = torch.nn.Embedding(5041175, 100)
logger.info("Shape of entity embedding: " + str(embed.weight.size()))
del vecs
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
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_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
all_ent = torch.tensor([f.input_ent for f in train_features],
dtype=torch.long)
all_ent_masks = torch.tensor([f.ent_mask for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_ent, all_ent_masks,
all_label_ids)
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)
output_loss_file = os.path.join(args.output_dir, "loss")
loss_fout = open(output_loss_file, 'w')
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(
t.to(device) if i != 3 else t for i, t in enumerate(batch))
input_ids, input_mask, segment_ids, input_ent, ent_mask, label_ids = batch
input_ent = embed(input_ent + 1).to(device) # -1 -> 0
loss = model(input_ids, segment_ids, input_mask,
input_ent.half(), ent_mask, label_ids)
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:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
loss_fout.write("{}\n".format(loss.item()))
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
model_to_save = model.module if hasattr(model, 'module') else model
output_model_file = os.path.join(
args.output_dir, "pytorch_model.bin_{}".format(global_step))
torch.save(model_to_save.state_dict(), output_model_file)
# Save a trained model
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
torch.save(model_to_save.state_dict(), output_model_file)
def main():
device = get_device()
if NUM_LABELS == 2:
class_names = ['True', 'Fake']
else:
class_names = [
'True', 'Mostly-true', 'Half-true', 'Barely-true', 'False',
'Pants-fire'
]
data_processor = CovidDataProcessor()
labels, statements = data_processor.load_dataset()
labels = {
'train':
CovidDataProcessor.convert_labels(NUM_LABELS, labels['train']),
'test':
CovidDataProcessor.convert_labels(NUM_LABELS, labels['test']),
'validation':
CovidDataProcessor.convert_labels(NUM_LABELS, labels['validation'])
}
# Load pre-trained model tokenizer
tokenizer = BertTokenizer.from_pretrained(ERNIE_BASE_PATH,
do_lower_case=True)
with open('embed.txt', 'rb') as f:
embed = pickle.load(f)
with open('entity2id.txt', 'rb') as f:
entity2id = pickle.load(f)
with open('ent_map.txt', 'rb') as f:
ent_map = pickle.load(f)
# # currently all saved dataloader is generated with batch_size = 4, if change to other batch_size, need to regenerate
# if Path('covid_train_dataloader.txt').is_file():
# with open('covid_train_dataloader.txt', 'rb') as ff:
# train_dataloader = pickle.load(ff)
# else:
# print('generating train_dataloader')
# train_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['train'], labels['train'], MAX_LEN,
# tokenizer, BATCH_SIZE, entity2id, ent_map)
# with open('covid_train_dataloader.txt', 'wb') as ff:
# pickle.dump(train_dataloader, ff)
#
# if Path('covid_test_dataloader.txt').is_file():
# with open('covid_test_dataloader.txt', 'rb') as ff:
# test_dataloader = pickle.load(ff)
# else:
# print('generating test_dataloader')
# test_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['test'], labels['test'], MAX_LEN,
# tokenizer, BATCH_SIZE, entity2id, ent_map)
# with open('covid_test_dataloader.txt', 'wb') as ff:
# pickle.dump(test_dataloader, ff)
#
# if Path('covid_val_dataloader.txt').is_file():
# with open('covid_val_dataloader.txt', 'rb') as ff:
# validation_dataloader = pickle.load(ff)
# else:
# print('generating validation_dataloader')
# validation_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['validation'], labels['validation'],
# MAX_LEN, tokenizer, BATCH_SIZE, entity2id,
# ent_map)
# with open('covid_val_dataloader.txt', 'wb') as ff:
# pickle.dump(validation_dataloader, ff)
# currently all saved dataloader is generated with batch_size = 4, if change to other batch_size, need to regenerate
# we now try to give RandomSampler to dataloader
# if Path('covid_train_dataloader_random.txt').is_file():
# with open('covid_train_dataloader_random.txt', 'rb') as ff:
# train_dataloader = pickle.load(ff)
# else:
# print('generating train_dataloader')
# train_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['train'], labels['train'], MAX_LEN,
# tokenizer, BATCH_SIZE, entity2id, ent_map)
# with open('covid_train_dataloader_random.txt', 'wb') as ff:
# pickle.dump(train_dataloader, ff)
#
# if Path('covid_test_dataloader_random.txt').is_file():
# with open('covid_test_dataloader_random.txt', 'rb') as ff:
# test_dataloader = pickle.load(ff)
# else:
# print('generating test_dataloader')
# test_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['test'], labels['test'], MAX_LEN,
# tokenizer, BATCH_SIZE, entity2id, ent_map)
# with open('covid_test_dataloader_random.txt', 'wb') as ff:
# pickle.dump(test_dataloader, ff)
#
# if Path('covid_val_dataloader.txt_random').is_file():
# with open('covid_val_dataloader.txt_random', 'rb') as ff:
# validation_dataloader = pickle.load(ff)
# else:
# print('generating validation_dataloader')
# validation_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['validation'],
# labels['validation'],
# MAX_LEN, tokenizer, BATCH_SIZE, entity2id,
# ent_map)
# with open('covid_val_dataloader.txt_random', 'wb') as ff:
# pickle.dump(validation_dataloader, ff)
# # all liar saved dataloader is generated with batch_size = 2, if change to other batch_size, need to regenerate
# if Path('liar_train_dataloader.txt').is_file():
# with open('liar_train_dataloader.txt', 'rb') as ff:
# train_dataloader = pickle.load(ff)
# else:
# print('generating train_dataloader')
# train_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['train'], labels['train'], MAX_LEN,
# tokenizer, BATCH_SIZE, entity2id, ent_map)
# with open('liar_train_dataloader.txt', 'wb') as ff:
# pickle.dump(train_dataloader, ff)
#
# if Path('liar_test_dataloader.txt').is_file():
# with open('liar_test_dataloader.txt', 'rb') as ff:
# test_dataloader = pickle.load(ff)
# else:
# print('generating test_dataloader')
# test_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['test'], labels['test'], MAX_LEN,
# tokenizer, BATCH_SIZE, entity2id, ent_map)
# with open('liar_test_dataloader.txt', 'wb') as ff:
# pickle.dump(test_dataloader, ff)
#
# if Path('liar_val_dataloader.txt').is_file():
# with open('liar_val_dataloader.txt', 'rb') as ff:
# validation_dataloader = pickle.load(ff)
# else:
# print('generating validation_dataloader')
# validation_dataloader = CovidDataProcessor.get_ernie_dataloader(statements['validation'], labels['validation'],
# MAX_LEN, tokenizer, BATCH_SIZE, entity2id,
# ent_map)
# with open('liar_val_dataloader.txt', 'wb') as ff:
# pickle.dump(validation_dataloader, ff)
# above is when num_labels = 6, now consider binary cases for liar
if Path('binary_liar_train_dataloader.txt').is_file():
with open('binary_liar_train_dataloader.txt', 'rb') as ff:
train_dataloader = pickle.load(ff)
else:
print('generating train_dataloader')
train_dataloader = CovidDataProcessor.get_ernie_dataloader(
statements['train'], labels['train'], MAX_LEN, tokenizer,
BATCH_SIZE, entity2id, ent_map)
with open('binary_liar_train_dataloader.txt', 'wb') as ff:
pickle.dump(train_dataloader, ff)
if Path('binary_liar_test_dataloader.txt').is_file():
with open('binary_liar_test_dataloader.txt', 'rb') as ff:
test_dataloader = pickle.load(ff)
else:
print('generating test_dataloader')
test_dataloader = CovidDataProcessor.get_ernie_dataloader(
statements['test'], labels['test'], MAX_LEN, tokenizer, BATCH_SIZE,
entity2id, ent_map)
with open('binary_liar_test_dataloader.txt', 'wb') as ff:
pickle.dump(test_dataloader, ff)
if Path('binary_liar_val_dataloader.txt').is_file():
with open('binary_liar_val_dataloader.txt', 'rb') as ff:
validation_dataloader = pickle.load(ff)
else:
print('generating validation_dataloader')
validation_dataloader = CovidDataProcessor.get_ernie_dataloader(
statements['validation'], labels['validation'], MAX_LEN, tokenizer,
BATCH_SIZE, entity2id, ent_map)
with open('binary_liar_val_dataloader.txt', 'wb') as ff:
pickle.dump(validation_dataloader, ff)
loss_fn = torch.nn.CrossEntropyLoss().to(device)
# Train model
model, _ = BertForSequenceClassification.from_pretrained(
ERNIE_BASE_PATH, num_labels=NUM_LABELS)
model.to(device)
ErnieModel.train_model(model, train_dataloader, validation_dataloader,
EPOCHS, device, loss_fn, embed)
# evaluate model on test dataset
test_acc, test_loss = ErnieModel.eval_model(model, test_dataloader, device,
embed)
print('test accuracy: ', test_acc.item())
# predictions
pred, test_labels = ErnieModel.get_predictions(model, test_dataloader,
device, embed)
print(
classification_report(test_labels,
pred,
target_names=class_names,
digits=4))
