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 .csv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
parser.add_argument("--init_checkpoint",
default=None,
type=str,
required=True,
help="The checkpoint file from pretraining")
## 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",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=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("--max_steps",
default=-1.0,
type=float,
help="Total number of training steps to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--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(
'--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 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 = 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):
print(
"WARNING: Output directory ({}) already exists and is not empty.".
format(args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_swag_examples(os.path.join(
args.data_dir, 'train.csv'),
is_training=True)
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
model = BertForMultipleChoice.from_pretrained(
args.bert_model,
cache_dir=os.path.join(PYTORCH_PRETRAINED_BERT_CACHE,
'distributed_{}'.format(args.local_rank)),
num_choices=4)
model.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu'),
strict=False)
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())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
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
}]
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=num_train_optimization_steps)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
tokenizer,
args.max_seq_length,
True)
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(select_field(train_features, 'input_ids'),
dtype=torch.long)
all_input_mask = torch.tensor(select_field(train_features,
'input_mask'),
dtype=torch.long)
all_segment_ids = torch.tensor(select_field(train_features,
'segment_ids'),
dtype=torch.long)
all_label = torch.tensor([f.label for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label)
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)
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")):
# Terminate early for benchmarking
if args.max_steps > 0 and global_step > args.max_steps:
break
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
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_train:
# Save a trained model and the associated configuration
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, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForMultipleChoice(config, num_choices=4)
# noinspection PyUnresolvedReferences
model.load_state_dict(torch.load(output_model_file))
else:
model = BertForMultipleChoice.from_pretrained(args.bert_model,
num_choices=4)
model.load_state_dict(torch.load(args.init_checkpoint,
map_location='cpu'),
strict=False)
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = read_swag_examples(os.path.join(
args.data_dir, 'val.csv'),
is_training=True)
eval_features = convert_examples_to_features(eval_examples, tokenizer,
args.max_seq_length, True)
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(select_field(eval_features, 'input_ids'),
dtype=torch.long)
all_input_mask = torch.tensor(select_field(eval_features,
'input_mask'),
dtype=torch.long)
all_segment_ids = torch.tensor(select_field(eval_features,
'segment_ids'),
dtype=torch.long)
all_label = torch.tensor([f.label for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label)
# 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, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
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():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_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
# noinspection PyUnboundLocalVariable
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': tr_loss / nb_tr_steps
}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
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():
args = arguments.get_argparse("multiple_choice")
logger.info(json.dumps(args.__dict__))
if args.eval_on_train and not args.log_spec:
args.log_spec = "on_train"
processors = {
"race": dataset_processor.RaceProcessor,
"mctest": dataset_processor.MCTestProcessor,
"swag": dataset_processor.SwagProcessor,
"squad": dataset_processor.SquadProcessor,
"openbookqa": dataset_processor.OpenBookQAProcessor,
"multirc": dataset_processor.MultiRCProcessor,
"arc": dataset_processor.ARCProcessor,
"qa4mre": dataset_processor.QA4MREProcessor,
}
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:
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")
if args.fp16:
logger.info(
"""16-bits training currently not supported in distributed
training""")
# (see https://github.com/pytorch/pytorch/pull/13496)
args.fp16 = False
logger.info(
"device %s n_gpu %d distributed training %r",
device,
n_gpu,
bool(args.local_rank != -1),
)
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 and not args.do_test:
raise ValueError(
"""At least one of `do_train` or `do_eval` or `do_test` must
be True.""")
if (args.do_train or args.do_eval) and args.do_test:
raise ValueError(
"Runing test must be independent of running train and/or dev")
bert_config = BertConfig.from_json_file(args.bert_config_file)
if args.max_seq_length > bert_config.max_position_embeddings:
raise ValueError(
"""Cannot use sequence length {} because the BERT model was only
trained up to sequence length {}""".format(
args.max_seq_length, bert_config.max_position_embeddings))
if args.small_debug:
args.output_dir = 'debug'
if os.path.exists(args.output_dir):
if not os.listdir(args.output_dir) == ["args_log.txt"
] and not args.small_debug:
raise ValueError(
"Output directory already exists and is not empty.")
os.makedirs(args.output_dir, exist_ok=True)
os.makedirs(args.cache_dir, exist_ok=True)
args_log = os.path.join(args.output_dir, "args_log.txt")
if not os.path.exists(args_log):
with open(args_log, "w") as writer:
writer.write(json.dumps(args.__dict__))
else:
print("args_log.txt already exists")
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
if "{}" in args.corenlp_cache_dir:
args.corenlp_cache_dir = args.corenlp_cache_dir.format(task_name)
processor = processors[task_name](args.data_dir, args.dataset_option)
num_options = processor.get_num_options()
if args.convert_from_ans_extr:
if args.do_train:
if args.train_predictions:
processor.set_candidates("train", args.train_predictions)
else:
raise ValueError("train prediction file is missing")
if args.do_eval:
if args.eval_predictions:
processor.set_candidates("dev", args.eval_predictions)
else:
raise ValueError("eval prediction file is missing")
tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
def cache_features(examples, split_name):
cache_spec_cand = [
task_name,
args.dataset_option,
split_name,
args.input_ablation,
]
cache_spec = "_".join(
[str(x) for x in cache_spec_cand if x is not None])
cache_path = os.path.join(args.cache_dir, "{}.pkl".format(cache_spec))
if os.path.exists(cache_path) and not args.no_cache:
features = pickle.load(open(cache_path, "rb"))
else:
if args.input_ablation or args.output_statistics:
corenlp_cache_path = os.path.join(
args.corenlp_cache_dir,
"{}_{}.pkl".format(task_name, split_name))
corenlp_cache = pickle.load(open(corenlp_cache_path, "rb"))
else:
corenlp_cache = None
if args.output_statistics:
output_mcmrc.output_statistics(examples, corenlp_cache)
tokenized_examples = generate_tokenized_examples(
examples, tokenizer, args.input_ablation, corenlp_cache,
args.entity_anonymization)
if args.output_mturk or args.output_examples:
for ex in examples:
ex.input_ablation = "original"
original_examples = generate_tokenized_examples(
examples, tokenizer, None, None)
if args.output_examples:
output_mcmrc.output_examples(
tokenized_examples,
original_examples,
task_name,
'ent_anon'
if args.entity_anonymization else args.input_ablation,
)
if args.output_mturk:
output_mcmrc.output_mturk(tokenized_examples,
original_examples, task_name,
args.input_ablation)
exit(1)
features = convert_examples_to_features(
tokenized_examples,
num_options,
args.max_seq_length,
args.max_query_length,
args.max_option_length,
tokenizer,
)
if not args.no_cache:
with open(cache_path, "wb") as f:
pickle.dump(features, f)
# assert len(examples) == len(features)
return features
if args.do_train:
train_examples = processor.get_train_examples()
if args.small_debug:
train_examples = train_examples[:6000]
num_train_per_epoch = len(train_examples)
num_train_per_epoch /= args.train_batch_size
num_train_per_epoch /= args.gradient_accumulation_steps
num_train_steps = int(num_train_per_epoch * args.num_train_epochs)
train_features = cache_features(train_examples, "train")
if args.do_eval:
if args.eval_on_train:
eval_examples = processor.get_train_examples()
eval_features = cache_features(eval_examples, "train")
else:
eval_examples = processor.get_dev_examples()
if args.small_debug:
eval_examples = eval_examples[:1000]
eval_features = cache_features(eval_examples, "dev")
if args.do_test:
eval_examples = processor.get_test_examples()
eval_features = cache_features(eval_examples, "test")
global entity_set
if args.entity_anonymization:
if len(entity_set) == 0:
anon_tag_cache_file = os.path.join(
args.cache_dir,
f'{task_name}_anon_tags_{args.entity_anonymization}.pkl')
if not os.path.exists(anon_tag_cache_file):
raise ValueError("vocabulary cache cannot be loaded")
entity_set = pickle.load(open(anon_tag_cache_file, 'rb'))
tokenizer.vocab_update(sorted(entity_set))
else:
anon_tag_cache_file = os.path.join(
args.cache_dir,
f'{task_name}_anon_tags_{args.entity_anonymization}.pkl')
if not os.path.exists(anon_tag_cache_file):
with open(anon_tag_cache_file, 'wb') as f:
pickle.dump(entity_set, f)
# Prepare model
model = BertForMultipleChoice(bert_config, num_options)
if args.init_checkpoint is not None:
state_dict = torch.load(args.init_checkpoint, map_location="cpu")
if list(state_dict)[0].startswith("bert."):
# finetuned on some target dataset
model.load_state_dict(state_dict)
else:
# pretrained language model
model.bert.load_state_dict(state_dict)
if args.entity_anonymization and len(entity_set):
model.bert.embeddings.extend_word_embeddings(len(entity_set))
if args.limit_vocab_size or args.limit_vocab_freq:
use_vocab, train_features, eval_features = vocab_selection(
train_features,
eval_features,
args.cache_dir,
args.output_dir,
task_name,
tokenizer,
args.entity_anonymization,
args.limit_vocab_size,
args.limit_vocab_freq,
num_options=num_options,
)
id_to_token = {v: k for k, v in tokenizer.vocab.items()}
use_tokens = [id_to_token[i] for i in use_vocab]
logger.info(sorted(use_tokens))
logger.info(f'{len(use_tokens)}')
model.bert.embeddings.limit_vocab(use_vocab)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [
(n, param.clone().detach().to("cpu").float().requires_grad_())
for n, param in model.named_parameters()
]
elif args.optimize_on_cpu:
param_optimizer = [(n,
param.clone().detach().to("cpu").requires_grad_())
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ["bias", "gamma", "beta"]
optimizer_grouped_parameters = [
{
"params": [p for n, p in param_optimizer if n not in no_decay],
"weight_decay_rate": 0.01,
},
{
"params": [p for n, p in param_optimizer if n in no_decay],
"weight_decay_rate": 0.0,
},
]
optimizer = BERTAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps,
)
global_step = 0
if args.enter_debugger:
model.eval()
# features = convert_examples_to_features(
# eval_examples,
# num_options,
# args.max_seq_length,
# args.max_query_length,
# args.max_option_length,
# tokenizer,
# )
# output = get_predictions(
# model, eval_examples, features, args, device
# )
# output_logits, output_predictions, eval_loss, eval_accuracy = output
print("in debugger")
import pdb
pdb.set_trace()
def eval_func(num_epoch=-1, num_step=-1, log_spec=None):
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
model.eval()
output = get_predictions(model, eval_examples, eval_features, args,
device)
output_logits, output_predictions, eval_loss, eval_accuracy = output
model.train()
output_qids = [e.qid for e in eval_examples]
output_answers = [e.ans_idx for e in eval_examples]
result = {
"eval_loss": eval_loss,
"eval_accuracy": eval_accuracy,
"global_step": global_step,
}
output_spec = ""
if num_epoch > -1 and num_step > -1:
output_spec = "_{}_{}".format(num_epoch, num_step)
elif log_spec:
output_spec += "_{}".format(log_spec)
output_eval_file = os.path.join(
args.output_dir, "eval_results{}.json".format(output_spec))
result["spec"] = output_spec
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
with open(output_eval_file, "w") as writer:
writer.write(json.dumps(result))
output_pred_file = os.path.join(
args.output_dir, "eval_preds{}.jsonl".format(output_spec))
with open(output_pred_file, "w") as f:
for qid, ans, pred, logit in zip(output_qids, output_answers,
output_predictions,
output_logits):
result = {
"qid": qid,
"answer": chr(ans + ord("A")),
"prediction": chr(pred + ord("A")),
"logits": logit.tolist(),
}
f.write(json.dumps(result) + "\n")
if args.do_train:
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)
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)
model.train()
for i in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
tmp_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) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss, _ = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/..
# ..sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
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 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
param.grad.data /= args.loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info("""FP16 TRAINING: Nan in gradients,
reducing loss scaling""")
args.loss_scale /= 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
if global_step % args.save_model_steps == 0:
output_model_file = os.path.join(
args.output_dir,
"pytorch_model_step_{}.bin".format(global_step),
)
if n_gpu > 1:
torch.save(model.module.state_dict(),
output_model_file)
else:
torch.save(model.state_dict(), output_model_file)
tmp_loss += loss.item()
if (args.loss_report_steps > 0
and global_step % args.loss_report_steps == 0):
logger.info("Step loss: {}".format(
tmp_loss / args.loss_report_steps))
tmp_loss = 0
if (args.eval_steps > 0 and global_step > 0
and global_step % args.eval_steps == 0
and args.do_eval):
eval_func(i, global_step, args.log_spec)
output_model_file = os.path.join(
args.output_dir, "pytorch_model_epoch_{}.bin".format(i))
if n_gpu > 1:
torch.save(model.module.state_dict(), output_model_file)
else:
torch.save(model.state_dict(), output_model_file)
if args.do_eval:
eval_func(i, global_step, args.log_spec)
if not args.do_train and args.do_eval:
eval_func(log_spec=args.log_spec or "dev")
if args.do_test:
eval_func(log_spec=args.log_spec or "test")
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("--bert_model", default=None, type=str, required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
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",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=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",
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('--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()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"commonsenseqa": CommonsenseQaProcessor,
}
num_labels_task = {
"cola": 2,
"mnli": 3,
"mrpc": 2,
"commonsenseqa":4,
}
# 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')
device = "cuda:2"
n_gpu = 1
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
print("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)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
print("current task is " + str(task_name))
processor = processors[task_name]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = 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.bert_model,
# cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(
# args.local_rank),
# num_labels=num_labels)
model = BertForMultipleChoice.from_pretrained(args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.
format(args.local_rank),
num_choices=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_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
nb_tr_steps = 0
tr_loss = 0
best_eval_accuracy = 0.0
if args.do_train:
train_features = convert_examples_to_features_mc(
train_examples, label_list, args.max_seq_length, tokenizer)
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)
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)
model.train()
# 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")
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) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss, logits = model(input_ids, segment_ids, input_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()
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 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_mc(
eval_examples, label_list, args.max_seq_length, tokenizer)
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)
all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
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, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
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():
tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids)
# logits = model(input_ids, segment_ids, input_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_accuracy = eval_accuracy / nb_eval_examples
print("the current eval accuracy is: " + str(eval_accuracy))
if eval_accuracy > best_eval_accuracy:
best_eval_accuracy = eval_accuracy
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
model.train()
# # 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")
# if args.do_train:
# torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
model_state_dict = torch.load(output_model_file)
model = BertForMultipleChoice.from_pretrained(args.bert_model,
state_dict=model_state_dict,
num_choices=num_labels)
model.to(device)
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_mc(
eval_examples, label_list, args.max_seq_length, tokenizer)
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)
all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
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, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
all_pred_labels = []
all_anno_labels = []
all_logits = []
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():
tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids)
# logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
output_labels = np.argmax(logits, axis=1)
all_pred_labels.extend(output_labels.tolist())
all_logits.extend(list(logits))
all_anno_labels.extend(list(label_ids))
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
loss = tr_loss / nb_tr_steps if args.do_train else None
result = {'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'best_eval_accuracy': best_eval_accuracy,
'global_step': global_step,
'loss': loss}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
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])))
for i in range(len(all_pred_labels)):
writer.write(str(i) + "\t" + str(all_anno_labels[i]) + "\t" +
str(all_pred_labels[i]) + "\t" + str(all_logits[i]) + "\n")
def main(start_index=0, end_index=0):
num_train_epochs = 3
learning_rate = 3e-5
max_seq_length = 58
train_batch_size = 40
warmup_proportion = 0.1
seed = 1979
gradient_accumulation_steps = 1
margin = 0.37
l2_reg = 0.02
do_margin_loss = 1
train_batch_size = int(train_batch_size / gradient_accumulation_steps)
eval_batch_size = train_batch_size
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
train_examples = load_copa_data('./data/COPA/train.jsonl')
num_train_steps = int(
len(train_examples) / train_batch_size / gradient_accumulation_steps *
num_train_epochs)
eval_examples = load_copa_data('./data/COPA/val.jsonl')
test_examples = load_copa_data_from_csv('./data/COPA/test.csv')
# Prepare model
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased",
do_lower_case=True)
if do_margin_loss:
model = BertForMultipleChoice.from_pretrained("bert-base-uncased",
num_choices=2,
margin=margin)
else:
model = BertForMultipleChoice.from_pretrained("bert-base-uncased",
num_choices=2)
model.cuda()
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
l2_reg
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
t_total = num_train_steps
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=t_total,
b2=0.98)
# Prep Eval Data
eval_features = convert_examples_to_features(eval_examples, tokenizer,
max_seq_length, True)
eval_input_ids = torch.tensor(select_field(eval_features, 'input_ids'),
dtype=torch.long)
eval_input_mask = torch.tensor(select_field(eval_features, 'input_mask'),
dtype=torch.long)
eval_segment_ids = torch.tensor(select_field(eval_features, 'segment_ids'),
dtype=torch.long)
eval_label = torch.tensor([f.label for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(eval_input_ids, eval_input_mask,
eval_segment_ids, eval_label)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=eval_batch_size)
# Prep Test Data
test_features = convert_examples_to_features(test_examples, tokenizer,
max_seq_length, True)
test_input_ids = torch.tensor(select_field(test_features, 'input_ids'),
dtype=torch.long)
test_input_mask = torch.tensor(select_field(test_features, 'input_mask'),
dtype=torch.long)
test_segment_ids = torch.tensor(select_field(test_features, 'segment_ids'),
dtype=torch.long)
test_label = torch.tensor([f.label for f in test_features],
dtype=torch.long)
test_data = TensorDataset(test_input_ids, test_input_mask,
test_segment_ids, test_label)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=eval_batch_size)
# Prep Training Data
train_features = convert_examples_to_features(train_examples, tokenizer,
max_seq_length, True)
train_input_ids = torch.tensor(select_field(train_features, 'input_ids'),
dtype=torch.long)
train_input_mask = torch.tensor(select_field(train_features, 'input_mask'),
dtype=torch.long)
train_segment_ids = torch.tensor(select_field(train_features,
'segment_ids'),
dtype=torch.long)
train_label = torch.tensor([f.label for f in train_features],
dtype=torch.long)
train_data = TensorDataset(train_input_ids, train_input_mask,
train_segment_ids, train_label)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=train_batch_size)
print("***** Running training *****")
print(" Num examples = %d", len(train_examples))
print(" Batch size = %d", train_batch_size)
print(" Num steps = %d", num_train_steps)
global_step = 0
eval_acc_list = []
for epoch in range(num_train_epochs):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc=(f'Epoch: {epoch}'))):
model.train()
batch = tuple(t.cuda() for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if gradient_accumulation_steps > 1:
loss = loss / gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % gradient_accumulation_steps == 0:
optimizer.step()
model.zero_grad()
global_step += 1
logits_all, eval_accuracy = do_evaluation(model,
eval_dataloader,
is_training=False)
tqdm.write(f'\nEvaluation Accuracy: {eval_accuracy}\n')
eval_acc_list.append(eval_accuracy)
logits_all, best_test_acc = do_evaluation(model,
test_dataloader,
is_training=False)
print(f'Testing Accuracy: {best_test_acc}')