def pretrain_model():
print("======MLM & NSP Pretraining======")
""" MLM & NSP Pretraining
You can modify this function by yourself.
This function does not affects your final score.
"""
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
comment_train_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'unlabeled'))
title_train_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'news_title'))
train_dataset = PretrainDataset(args.max_seq_length, comment_train_dataset,
title_train_dataset, tokenizer)
comment_val_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'unlabeled', 'dev'))
title_val_dataset = ParagraphDataset(
os.path.join('data', 'korean-hate-speech', 'news_title', 'dev'))
val_dataset = PretrainDataset(args.max_seq_length, comment_val_dataset,
title_val_dataset, tokenizer)
config = Config(args.config_file)
device = torch.device('cpu')
model = MlmNspModel(config).to(torch.device(device))
model.bert.load_state_dict(torch.load(args.checkpoint,
map_location=device))
model_name = 'pretrained'
MLM_train_losses, MLM_val_losses, NSP_train_losses, NSP_val_losses \
= pretraining(model, model_name, train_dataset, val_dataset)
torch.save(model.state_dict(), model_name + '_final.pth')
with open(model_name + '_result.pkl', 'wb') as f:
pickle.dump((MLM_train_losses, MLM_val_losses, NSP_train_losses,
NSP_val_losses), f)
utils_pretrain.plot_values(MLM_train_losses,
MLM_val_losses,
title=model_name + "_mlm")
utils_pretrain.plot_values(NSP_train_losses,
NSP_val_losses,
title=model_name + "_nsp")
print("Final MLM training loss: {:06.4f}".format(MLM_train_losses[-1]))
print("Final MLM validation loss: {:06.4f}".format(MLM_val_losses[-1]))
print("Final NSP training loss: {:06.4f}".format(NSP_train_losses[-1]))
print("Final NSP validation loss: {:06.4f}".format(NSP_val_losses[-1]))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint",
default='output/korquad_3.bin',
type=str,
help="checkpoint")
parser.add_argument(
"--output_dir",
default='debug',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
## Other parameters
parser.add_argument(
"--predict_file",
default='data/KorQuAD_v1.0_dev.json',
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument(
"--config_name",
default="data/bert_small.json",
type=str,
help="Pretrained config name or path if not the same as model_name")
parser.add_argument(
"--max_seq_length",
default=512,
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=64,
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=96,
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_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: %s, n_gpu: %s, 16-bits training: %s", device,
args.n_gpu, args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(vocab_file='data/wiki_vocab_32k.txt',
do_basic_tokenize=True,
max_len=args.max_seq_length)
config = Config.from_json_file(args.config_name)
model = QuestionAnswering(config)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
evaluate(args, model, examples, features)
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--config_file",
type=str,
required=True,
help="model configuration file")
parser.add_argument("--vocab_file",
type=str,
required=True,
help="tokenizer vocab file")
parser.add_argument("--checkpoint",
type=str,
required=True,
help="pretrained model checkpoint")
# Other parameters
parser.add_argument(
"--output_dir",
default="output",
type=str,
help=
"The output directory where the model predictions and checkpoints will be written."
)
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_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--train_batch_size",
default=16,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=32,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=10.0,
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("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--no_cuda",
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',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
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")
args = parser.parse_args()
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
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = n_gpu
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
set_seed(args)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
config = Config(args.config_file)
#
if output_mode == "classification":
model = SequenceClassification(config, num_labels=num_labels)
elif output_mode == "classification_multitask":
model = SequenceClassificationMultitask(config)
model.bert.load_state_dict(torch.load(args.checkpoint))
# Do not fine-tune BERT parameters
# for param in model.bert.parameters():
# param.requires_grad = False
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
global_step, tr_loss = 0, 0.
train_features = convert_examples_to_features(train_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
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_optimization_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)
if output_mode == "classification_multitask":
all_label_ids = all_label_ids = torch.tensor(
[f.label_id for f in train_features], dtype=torch.long)
num_labels = [2, 3, 3]
elif output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
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)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['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 = len(train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate)
scheduler = WarmupLinearSchedule(optimizer,
warmup_steps=t_total *
args.warmup_proportion,
t_total=t_total)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
if args.fp16_opt_level == "O2":
keep_batchnorm_fp32 = False
else:
keep_batchnorm_fp32 = True
model, optimizer = amp.initialize(
model,
optimizer,
opt_level=args.fp16_opt_level,
keep_batchnorm_fp32=keep_batchnorm_fp32)
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)
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
writer = open(output_eval_file, "w", encoding="utf-8")
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
epoch_iterator = tqdm(train_dataloader,
desc="Train(XX Epoch) Step(X/X) (loss=X.X)",
disable=args.local_rank not in [-1, 0])
model.train()
for step, batch in enumerate(epoch_iterator):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# define a new function to compute loss values for both output_modes
logits = model(input_ids, segment_ids, input_mask, labels=None)
if output_mode == "classification_multitask":
loss_fct0 = CrossEntropyLoss()
loss_fct1 = CrossEntropyLoss()
loss_fct2 = CrossEntropyLoss()
loss0 = loss_fct0(logits[0].view(-1, num_labels[0]),
label_ids[:, 0].view(-1))
loss1 = loss_fct1(logits[1].view(-1, num_labels[1]),
label_ids[:, 1].view(-1))
loss2 = loss_fct2(logits[2].view(-1, num_labels[2]),
label_ids[:, 2].view(-1))
loss = loss0 + loss1 + loss2
elif output_mode == "classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), label_ids.view(-1))
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()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scheduler.step() # Update learning rate schedule
optimizer.step()
optimizer.zero_grad()
global_step += 1
epoch_iterator.set_description(
"Train(%d Epoch) Step(%d / %d) (loss=%5.5f)" %
(_, global_step, t_total, loss.item()))
model_checkpoint = "%s_%d.bin" % (args.task_name, _)
logger.info(model_checkpoint)
output_model_file = os.path.join(args.output_dir, model_checkpoint)
if n_gpu > 1 or args.local_rank != -1:
logger.info("***** Saving file *****(module)")
torch.save(model.module.state_dict(), output_model_file)
else:
logger.info("***** Saving file *****")
torch.save(model.state_dict(), output_model_file)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer,
output_mode)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_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)
if output_mode == "classification_multitask":
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.long)
elif output_mode == "classification":
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor(
[f.label_id for f in eval_features], dtype=torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, 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)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
preds0 = []
preds1 = []
preds2 = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids,
segment_ids,
input_mask,
labels=None)
# create eval loss and other metric required by the task
if output_mode == "classification_multitask":
loss_fct0 = CrossEntropyLoss()
loss_fct1 = CrossEntropyLoss()
loss_fct2 = CrossEntropyLoss()
loss0 = loss_fct0(logits[0].view(-1, num_labels[0]),
label_ids[:, 0].view(-1))
loss1 = loss_fct1(logits[1].view(-1, num_labels[1]),
label_ids[:, 1].view(-1))
loss2 = loss_fct2(logits[2].view(-1, num_labels[2]),
label_ids[:, 2].view(-1))
tmp_eval_loss = loss0 + loss1 + loss2
elif output_mode == "classification":
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
tmp_eval_loss = loss_fct(logits.view(-1),
label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if output_mode == "classification_multitask":
if len(preds0) == 0:
preds0.append(logits[0].detach().cpu().numpy())
preds1.append(logits[1].detach().cpu().numpy())
preds2.append(logits[2].detach().cpu().numpy())
else:
preds0[0] = np.append(preds0[0],
logits[0].detach().cpu().numpy(),
axis=0)
preds1[0] = np.append(preds1[0],
logits[1].detach().cpu().numpy(),
axis=0)
preds2[0] = np.append(preds2[0],
logits[2].detach().cpu().numpy(),
axis=0)
else:
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
if output_mode == "classification_multitask":
preds0 = preds0[0]
preds1 = preds1[0]
preds2 = preds2[0]
preds0 = np.argmax(preds0, axis=1)
preds1 = np.argmax(preds1, axis=1)
preds2 = np.argmax(preds2, axis=1)
preds = [preds0, preds1, preds2]
elif output_mode == "classification":
preds = preds[0]
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = preds[0]
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / global_step
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
logger.info("***** Eval results *****")
writer.write("%d_epoch\n" % (_ + 1))
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
writer.write("\n")
writer.close()
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--config_file",
type=str,
required=True,
help="model configuration file")
parser.add_argument("--vocab_file",
type=str,
required=True,
help="tokenizer vocab file")
parser.add_argument("--checkpoint",
default=None,
type=str,
required=True,
help="fine-tuned model checkpoint")
# 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("--eval_batch_size",
default=32,
type=int,
help="Total batch size for eval.")
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('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
args.n_gpu = n_gpu
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: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
set_seed(args)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
test_examples = processor.get_test_examples(args.data_dir)
texts = [test_example.text_a for test_example in test_examples]
# Prepare model
config = Config(args.config_file)
model = SequenceClassificationMultitask(config)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
if args.fp16:
model = model.half()
test_features = convert_examples_to_features(
test_examples, label_list, args.max_seq_length, tokenizer, output_mode)
logger.info("***** Running Evaluation *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in test_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in test_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features], dtype=torch.long)
if output_mode == "classification_multitask":
all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.float)
test_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=args.eval_batch_size)
model.eval()
preds = []
preds0 = []
preds1 = []
preds2 = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(test_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask, labels=None)
if output_mode == "classification_multitask":
if len(preds0) == 0:
preds0.append(logits[0].detach().cpu().numpy())
preds1.append(logits[1].detach().cpu().numpy())
preds2.append(logits[2].detach().cpu().numpy())
else:
preds0[0] = np.append(preds0[0], logits[0].detach().cpu().numpy(), axis=0)
preds1[0] = np.append(preds1[0], logits[1].detach().cpu().numpy(), axis=0)
preds2[0] = np.append(preds2[0], logits[2].detach().cpu().numpy(), axis=0)
else:
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(
preds[0], logits.detach().cpu().numpy(), axis=0)
if output_mode == "classification_multitask":
preds0 = preds0[0]
preds1 = preds1[0]
preds2 = preds2[0]
preds0 = np.argmax(preds0, axis=1)
preds1 = np.argmax(preds1, axis=1)
preds2 = np.argmax(preds2, axis=1)
preds = [preds0, preds1, preds2]
label=[]
for p in preds0:
if p==0:
label.append('True')
elif p==1:
label.append('False')
elif output_mode == "classification":
preds = preds[0]
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = preds[0]
preds = np.squeeze(preds)
#TODO: write the csv
#result = compute_metrics(task_name, preds, all_label_ids.numpy())
data = {'comments':texts,
'label':label}
data_df = DataFrame(data)
csv_write_path='/mnt/sdc1/korLM/by_me/out/by_me_9_gender_0608.csv'
# data_df.to_csv(csv_write_path, sep=',',index=False)
with open(csv_write_path, 'w', encoding='utf-8') as f:
wr = csv.writer(f)
wr.writerow(['comments','label'])
for text, lab in zip(texts, label):
wr.writerow([text, lab])
logger.info("***** Eval results *****")
def main():
parser = argparse.ArgumentParser()
# Parameters
parser.add_argument(
"--output_dir",
default='output',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--checkpoint",
default='large_v1/large_v1_model.bin',
type=str,
help="pre-trained model checkpoint")
parser.add_argument("--config_file",
default='large_v1/large_config.json',
type=str,
help="model configuration file")
parser.add_argument("--vocab_file",
default='large_v1/large_v1_32k_vocab.txt',
type=str,
help="tokenizer vocab file")
parser.add_argument("--train_file",
default='data/korquad/KorQuAD_v1.0_train.json',
type=str,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument("--test",
default=False,
type=bool,
help="Test by sample")
parser.add_argument("--train_data_limit",
default=-1,
type=int,
help="For Check EDA performance limit train data")
######ADD for EDA#######
parser.add_argument(
"--same_random_index",
default=True,
type=bool,
help="for eda data limit, if True every eda pick same random index")
parser.add_argument("--eda_type",
default="no_eda",
type=str,
help="eda type e.g) sr, ri, rs, rd")
parser.add_argument("--alpha", default=0.1, type=float, help="For eda")
parser.add_argument("--num_aug", default=2, type=int, help="For eda")
parser.add_argument("--eda_model_name",
default="word2vec",
type=str,
help="For eda")
parser.add_argument("--min_score", default=0.7, type=float, help="For eda")
parser.add_argument("--dict_file_name",
default="data/syn/synonyms_file.pickle",
type=str)
########################
parser.add_argument(
"--max_seq_length",
default=512,
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(
"--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("--per_gpu_train_batch_size",
default=4,
type=int,
help="Total batch size for training.")
parser.add_argument(
'--gradient_accumulation_steps',
default=1,
type=int,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--weight_decay",
default=0.01,
type=float,
help="Weight deay if we apply some.")
parser.add_argument("--num_train_epochs",
default=4.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--adam_epsilon",
default=1e-6,
type=float,
help="Epsilon for Adam optimizer.")
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(
"--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',
default=42,
type=int,
help="random seed for initialization")
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='O2',
type=str,
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
parser.add_argument("--local_rank",
default=-1,
type=int,
help="For distributed training: local_rank")
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."
)
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
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")
args.n_gpu = torch.cuda.device_count()
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
torch.distributed.init_process_group(backend='nccl')
args.n_gpu = 1
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1),
args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(args.vocab_file,
max_len=args.max_seq_length,
do_basic_tokenize=True)
# Prepare model
config = Config.from_json_file(args.config_file)
model = QuestionAnswering(config)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
logger.info("Training hyper-parameters %s", args)
# Training
train_dataset = load_and_cache_examples(args, tokenizer)
train(args, train_dataset, model)
# Setup logging
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt = '%m/%d/%Y %H:%M:%S',
level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(args.vocab_file,
max_len=args.max_seq_length,
do_basic_tokenize=True)
# Prepare model
config = Config.from_json_file(args.config_file)
model = QuestionAnswering(config)
model.bert.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
logger.info("Training hyper-parameters %s", args)
# Training
train_dataset = load_and_cache_examples(args, tokenizer)
train(args, train_dataset, model)
if __name__ == "__main__":
main()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint",
default='output/korquad_3.bin',
type=str,
help="checkpoint")
parser.add_argument(
"--output_dir",
default='debug',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--model_config", type=str)
parser.add_argument("--vocab", type=str)
## Other parameters
parser.add_argument(
"--predict_file",
default='data/KorQuAD_v1.0_dev.json',
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument(
"--max_seq_length",
default=512,
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=64,
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=96,
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_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
parser.add_argument('--ensemble', default='false', type=str)
args = parser.parse_args()
args.ensemble = args.ensemble.lower() == 'true'
# Setup CUDA, GPU & distributed training
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: %s, n_gpu: %s, 16-bits training: %s", device,
args.n_gpu, args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(vocab_file=args.vocab,
do_basic_tokenize=True,
max_len=args.max_seq_length)
config = Config.from_json_file(args.model_config)
model = QuestionAnswering(config)
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
if not args.ensemble:
logger.info(" Load Model: %s ", args.checkpoint)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
results = evaluate(args, model, examples, features)
else:
list_ckpts = []
with open(os.path.join(args.output_dir, "ckpt_list.txt"), 'r') as f:
for line in f:
list_ckpts.append(line.strip())
list_results = []
for i, ckpt in enumerate(list_ckpts):
fn = os.path.join(args.output_dir, ckpt)
logger.info(" Load Model: %s ", fn)
model.load_state_dict(torch.load(fn))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
results = evaluate(args, model, examples, features)
list_results.append(results)
list_qid = []
for example in examples:
list_qid.append(example.qas_id)
all_predictions = collections.OrderedDict()
for qid in list_qid:
max_prob, answer = 0.0, ""
for results in list_results:
prob, text = 0.0, None
for output in results[qid]:
if output["text"]:
prob = output["probability"]
text = output["text"]
break
if prob > max_prob:
max_prob = prob
answer = text
all_predictions[qid] = answer
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
with open(output_prediction_file, "w") as writer:
writer.write(json.dumps(all_predictions, indent=4) + "\n")
expected_version = 'KorQuAD_v1.0'
with open(args.predict_file) as dataset_file:
dataset_json = json.load(dataset_file)
read_version = "_".join(dataset_json['version'].split("_")[:-1])
if (read_version != expected_version):
logger.info('Evaluation expects ' + expected_version +
', but got dataset with ' + read_version,
file=sys.stderr)
dataset = dataset_json['data']
with open(os.path.join(args.output_dir,
"predictions.json")) as prediction_file:
predictions = json.load(prediction_file)
logger.info(json.dumps(korquad_eval(dataset, predictions)))
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("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--config_file",
type=str,
required=True,
help="model configuration file")
parser.add_argument("--vocab_file",
type=str,
required=True,
help="tokenizer vocab file")
parser.add_argument("--checkpoint",
default=None,
type=str,
required=True,
help="fine-tuned model checkpoint")
# 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("--eval_batch_size",
default=32,
type=int,
help="Total batch size for eval.")
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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
args = parser.parse_args()
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
args.n_gpu = n_gpu
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: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
set_seed(args)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer(args.vocab_file, max_len=args.max_seq_length)
test_examples = processor.get_test_examples(args.data_dir)
# Prepare model
config = Config(args.config_file)
model = SequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
if args.fp16:
model = model.half()
test_features = convert_examples_to_features(test_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running Evaluation *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in test_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in test_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in test_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in test_features],
dtype=torch.float)
test_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_label_ids)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.eval_batch_size)
model.eval()
preds = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
test_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask, labels=None)
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
preds = preds[0]
if output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
def main():
parser = argparse.ArgumentParser()
# Required Parameters
parser.add_argument(
"--output_dir",
default='output',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--checkpoint",
default='pretrain_ckpt/bert_small_ckpt.bin',
type=str,
help="checkpoint")
parser.add_argument("--resume_checkpoint",
default=False,
type=bool,
help="resume")
parser.add_argument('--log_dir', default='./runs', type=str)
parser.add_argument("--model_config",
default='data/bert_small.json',
type=str)
# Other Parameters
parser.add_argument(
"--corpus",
default='./rsc/corpus/korquad_corpus_post_training.hdf5',
type=str,
help="SQuAD corpus for post-training.")
parser.add_argument(
"--max_seq_length",
default=512,
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("--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=4.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--adam_epsilon",
default=1e-6,
type=float,
help="Epsilon for Adam optimizer.")
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("--num_workers",
default=8,
type=int,
help="Proportion of workers of DataLoader")
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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
args = parser.parse_args()
summary_writer = SummaryWriter(args.log_dir)
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {} n_gpu: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
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 os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# Prepare model
config = Config.from_json_file(args.model_config)
model = BertForPostTraining(config)
if args.resume_checkpoint:
model.load_state_dict(torch.load(args.checkpoint))
else:
model.bert.load_state_dict(torch.load(args.checkpoint))
# Multi-GPU Setting
if n_gpu > 1:
model = nn.DataParallel(model)
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
post_training_dataset = BertPostTrainingDataset(args.corpus,
args.max_seq_length)
num_train_optimization_steps = int(
len(post_training_dataset) /
args.train_batch_size) * args.num_train_epochs
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['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
}]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = WarmupLinearSchedule(
optimizer,
warmup_steps=num_train_optimization_steps * 0.1,
t_total=num_train_optimization_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."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(post_training_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
num_train_step = num_train_optimization_steps
train_sampler = RandomSampler(post_training_dataset)
train_dataloader = DataLoader(post_training_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size,
num_workers=16,
pin_memory=True)
model.train()
global_step = 0
epoch = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
iter_bar = tqdm(
train_dataloader,
desc="Train(XX Epoch) Step(XX/XX) (Mean loss=X.X) (loss=X.X)")
tr_step, total_loss, mean_loss = 0, 0., 0.
for step, batch in enumerate(iter_bar):
# loss, mlm_loss, nsp_loss = model(batch)
output = model(batch)
loss, mlm_loss, nsp_loss = output[:3]
if n_gpu > 1:
loss = loss.mean()
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
global_step += 1
tr_step += 1
total_loss += loss
mean_loss = total_loss / tr_step
iter_bar.set_description(
"Train Step(%d / %d) (Mean loss=%5.5f) (loss=%5.5f)" %
(global_step, num_train_step, mean_loss, loss.item()))
if global_step % 100 == 0:
summary_writer.add_scalar('Train/Total_Mean_Loss', mean_loss,
global_step)
summary_writer.add_scalar('Train/Total_Loss', loss.item(),
global_step)
summary_writer.add_scalar('Train/MLM_Loss',
mlm_loss.mean().item(), global_step)
summary_writer.add_scalar('Train/NSP_Loss',
nsp_loss.mean().item(), global_step)
logger.info("***** Saving file *****")
if args.resume_checkpoint:
model_checkpoint = "pt_bert_from_checkpoint_%d.bin" % (epoch)
else:
model_checkpoint = "pt_bert_%d.bin" % (epoch)
logger.info(model_checkpoint)
output_model_file = os.path.join(args.output_dir, model_checkpoint)
if n_gpu > 1:
torch.save(model.module.state_dict(), output_model_file)
else:
torch.save(model.state_dict(), output_model_file)
epoch += 1
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--output_dir",
default='output',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--checkpoint",
default='pretrain_ckpt/bert_small_ckpt.bin',
type=str,
help="checkpoint")
parser.add_argument("--model_config",
default='data/bert_small.json',
type=str)
parser.add_argument("--vocab", type=str)
# Other parameters
parser.add_argument("--train_file",
default='data/KorQuAD_v1.0_train.json',
type=str,
help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument(
"--max_seq_length",
default=512,
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=96,
type=int,
help=
"The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
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=4.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--max_grad_norm",
default=1.0,
type=float,
help="Max gradient norm.")
parser.add_argument("--adam_epsilon",
default=1e-6,
type=float,
help="Epsilon for Adam optimizer.")
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(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--fp16_opt_level',
type=str,
default='O2',
help=
"For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
"See details at https://nvidia.github.io/apex/amp.html")
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('--grad_noise', default='false', type=str)
parser.add_argument('--gs_noise', default='false', type=str)
parser.add_argument('--max_num_example', default=80000, type=int)
args = parser.parse_args()
args.grad_noise = args.grad_noise.lower() == 'true'
args.gs_noise = args.gs_noise.lower() == 'true'
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {} n_gpu: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
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 os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
ckpt_list_fn = os.path.join(args.output_dir, "ckpt_list.txt")
if os.path.isfile(ckpt_list_fn):
os.remove(ckpt_list_fn)
tokenizer = BertTokenizer(args.vocab,
max_len=args.max_seq_length,
do_basic_tokenize=True)
# Prepare model
config = Config.from_json_file(args.model_config)
model = QuestionAnswering(config)
if 'bert' in args.checkpoint:
model.bert.load_state_dict(torch.load(args.checkpoint))
logger.info("bert parameters are loaded from %s" % args.checkpoint)
else:
model.load_state_dict(torch.load(args.checkpoint))
logger.info("all parameter is loaded from %s" % args.checkpoint)
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
model.to(device)
train_examples = read_squad_examples(input_file=args.train_file,
is_training=True,
version_2_with_negative=False)
num_train_examples = len(train_examples)
cached_train_features_file = args.train_file + '_{0}_{1}_{2}'.format(
str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length))
is_dynamic = False
if num_train_examples <= args.max_num_example:
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
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,
is_dynamic=False)
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
del train_examples
num_train_features = len(train_features)
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)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_start_positions,
all_end_positions)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
else:
# For being able to get data dynamically (for large datasets)
is_dynamic = True
pkl_list = []
num_train_features = 0
if not os.path.isfile(cached_train_features_file + "_0"):
batch = 10000
unique_id = 1000000000
for start, end in zip(
range(0, len(train_examples), batch),
range(batch,
len(train_examples) + batch, batch)):
file_path = cached_train_features_file + "_%d" % start
unique_id = convert_examples_to_features(
examples=train_examples[start:end],
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True,
is_dynamic=True,
file_path=file_path,
init_unique_id=unique_id)
pkl_list.append(file_path)
num_train_features = unique_id - 1000000000
with open(cached_train_features_file + "_info", 'w') as fw:
fw.write("%d\n" % num_train_features)
for pkl in pkl_list:
fw.write(pkl + '\n')
else:
with open(cached_train_features_file + "_info", 'r') as f:
line_cnt = 0
for line in f:
if line_cnt == 0:
num_train_features = int(line.strip())
else:
pkl_list.append(line.strip())
line_cnt += 1
del train_examples
train_data = KorquadDataset(pkl_list, args.train_batch_size, tokenizer)
num_train_optimization_steps = int(
num_train_features / args.train_batch_size) * args.num_train_epochs
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['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
}]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = WarmupLinearSchedule(
optimizer,
warmup_steps=num_train_optimization_steps * 0.1,
t_total=num_train_optimization_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."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", num_train_examples)
logger.info(" Num split examples = %d", num_train_features)
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
num_train_step = num_train_optimization_steps
model.train()
global_step = 0
epoch = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
if is_dynamic:
num_step = int(num_train_step / args.num_train_epochs)
if num_step * args.train_batch_size != num_train_features:
num_step += 1
iter_bar = tqdm(
list(range(num_step)),
desc="Train(XX Epoch) Step(XX/XX) (Mean loss=X.X) (loss=X.X)")
else:
iter_bar = tqdm(
train_dataloader,
desc="Train(XX Epoch) Step(XX/XX) (Mean loss=X.X) (loss=X.X)")
tr_step, total_loss, mean_loss = 0, 0., 0.
for step, batch in enumerate(iter_bar):
if is_dynamic:
input_ids, input_mask, segment_ids, start_positions, end_positions = train_data.next_batch(
)
if n_gpu == 1:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
start_positions = start_positions.to(device)
end_positions = end_positions.to(device)
else:
if n_gpu == 1:
batch = tuple(
t.to(device)
for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch
if args.grad_noise:
loss = model(input_ids,
segment_ids,
input_mask,
start_positions,
end_positions,
cal_grad=True)
else:
loss = model(input_ids, segment_ids, input_mask,
start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
args.max_grad_norm)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
if args.grad_noise or args.gs_noise:
if args.grad_noise:
adv_grad = model.bert.embedding_output.grad
adv_loss = model(input_ids,
segment_ids,
input_mask,
start_positions,
end_positions,
adv_grad,
cal_grad=False,
is_gs_noise=False)
else:
adv_loss = model(input_ids,
segment_ids,
input_mask,
start_positions,
end_positions,
adv_grad=None,
cal_grad=False,
is_gs_noise=True)
if n_gpu > 1:
adv_loss = adv_loss.mean(
) # mean() to average on multi-gpu.
if args.fp16:
with amp.scale_loss(adv_loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
adv_loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scheduler.step()
optimizer.step()
optimizer.zero_grad()
global_step += 1
tr_step += 1
total_loss += loss.item()
mean_loss = total_loss / tr_step
iter_bar.set_description(
"Train Step(%d / %d) (Mean loss=%5.5f) (loss=%5.5f)" %
(global_step, num_train_step, mean_loss, loss.item()))
logger.info("** ** * Saving file * ** **")
model_checkpoint = "korquad_%d.bin" % (epoch)
logger.info(model_checkpoint)
output_model_file = os.path.join(args.output_dir, model_checkpoint)
if n_gpu > 1:
torch.save(model.module.state_dict(), output_model_file)
else:
torch.save(model.state_dict(), output_model_file)
with open(os.path.join(args.output_dir, "ckpt_list.txt"), 'a') as fw:
fw.write(model_checkpoint + '\n')
epoch += 1
def main(input, output):
max_seq_length = 512
doc_stride = 64
max_query_length = 64
batch_size = 16
n_best_size = 20
max_answer_length = 30
seed = 42
fp16 = False
device = torch.device("cpu")
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {} n_gpu: {}".format(device, n_gpu))
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(seed)
tokenizer = BertTokenizer(vocab_file=VOCAB_PATH,
max_len=max_seq_length,
do_basic_tokenize=True)
eval_examples = read_squad_examples(input_file=input,
is_training=False,
version_2_with_negative=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=max_seq_length,
doc_stride=doc_stride,
max_query_length=max_query_length,
is_training=False)
# Prepare model
config = Config.from_json_file(CONFIG_PATH)
model = QuestionAnswering(config)
if fp16 is True:
model.half()
model.load_state_dict(torch.load(CHK_PATH, map_location=device))
model.to(device)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", 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)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=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"):
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 = 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()
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))
output_nbest_file = os.path.join("nbest_predictions.json")
write_predictions(eval_examples, eval_features, all_results, n_best_size,
max_answer_length, False, output, output_nbest_file,
None, False, False, 0.0)
def main():
parser = argparse.ArgumentParser()
# Parameters
parser.add_argument(
"--output_dir",
default='debug',
type=str,
help=
"The output directory where the model checkpoints and predictions will be written."
)
parser.add_argument("--checkpoint",
default='output/korquad_3.bin',
type=str,
help="fine-tuned model checkpoint")
parser.add_argument("--config_file",
default='data/large_config.json',
type=str,
help="model configuration file")
parser.add_argument("--vocab_file",
default='data/large_v1_32k_vocab.txt',
type=str,
help="tokenizer vocab file")
parser.add_argument("--check_list", default=False, type=bool)
parser.add_argument("--eda_type",
default="no_eda",
type=str,
help="sr, ri , rd, rs")
parser.add_argument(
"--predict_file",
default='data/korquad/KorQuAD_v1.0_dev.json',
type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json"
)
parser.add_argument(
"--max_seq_length",
default=512,
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=64,
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(
"--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("--batch_size",
default=16,
type=int,
help="Batch size per GPU/CPU for evaluation.")
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(
"--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("--just_eval", default=False, type=bool)
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(
'--fp16',
action='store_true',
help=
"Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit"
)
args = parser.parse_args()
# Setup CUDA, GPU & distributed training
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
args.n_gpu = torch.cuda.device_count()
args.device = device
# Setup logging
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
logger.info("device: %s, n_gpu: %s, 16-bits training: %s", device,
args.n_gpu, args.fp16)
# Set seed
set_seed(args)
tokenizer = BertTokenizer(vocab_file=args.vocab_file,
do_basic_tokenize=True,
max_len=args.max_seq_length)
config = Config.from_json_file(args.config_file)
model = QuestionAnswering(config)
if args.check_list:
model_list = os.listdir(args.checkpoint)
model_list.sort()
result_dict = defaultdict(list)
eda_list = ["st_rs", "st_rd", "rs", "rd", "sr", "ri", "no_eda"]
for m in model_list:
model.load_state_dict(torch.load(args.checkpoint + '/' + m))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
for eda in eda_list:
if eda in m:
args.eda_type = eda
break
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
r = evaluate(args, model, examples, features)
em = r["exact_match"]
f1 = r["f1"]
result_dict[args.eda_type].append([em, f1, m])
print(result_dict)
else:
model.load_state_dict(torch.load(args.checkpoint))
num_params = count_parameters(model)
logger.info("Total Parameter: %d" % num_params)
if args.fp16:
model.half()
model.to(args.device)
logger.info("Evaluation parameters %s", args)
# Evaluate
examples, features = load_and_cache_examples(args, tokenizer)
evaluate(args, model, examples, features)