def optimizer_fn(param_group, max_grad_norm=None):
group0 = dict(params=[], weight_decay_rate=args.weight_decay, names=[])
group1 = dict(params=[], weight_decay_rate=0.00, names=[])
for (n, p) in param_group:
if not any(nd in n for nd in no_decay):
group0['params'].append(p)
group0['names'].append(n)
else:
group1['params'].append(p)
group1['names'].append(n)
optimizer_grouped_parameters = [group0, group1]
optimizer = BertAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
b1=args.adam_beta1,
b2=args.adam_beta2,
v1=args.qhadam_v1,
v2=args.qhadam_v2,
lr_ends=args.lr_schedule_ends,
warmup=args.warmup_proportion if args.warmup_proportion < 1 else
args.warmup_proportion / training_steps,
t_total=training_steps,
schedule=args.lr_schedule,
max_grad_norm=args.max_grad_norm
if max_grad_norm is None else max_grad_norm,
global_grad_norm=args.global_grad_norm,
init_spec=init_spec,
weight_decay_rate=args.weight_decay)
return optimizer
'weight_decay_rate':
args.weight_decay_rate
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
num_train_steps = int(
len(train_iter.dataset) / args.batch_size *
device_num) * args.num_epochs
num_train_steps = num_train_steps if args.t_total else -1
optimizer = BertAdam(params=optimizer_grouped_parameters,
lr=args.bert_lr,
warmup=args.warmup,
t_total=num_train_steps)
for epoch in range(1, args.num_epochs + 1):
logger.info(
"==========epoch {} fine tune start==========".format(epoch))
logger.info('train examples {}'.format(len(train_iter.dataset)))
logger.info('train batch size {}'.format(args.batch_size))
logger.info('train lr {}'.format(optimizer.get_lr()[0]))
writer.add_scalar('lr', optimizer.get_lr()[0], epoch)
train(model, train_iter, optimizer, epoch)
logger.info(
"==========epoch {} fine tune end==========".format(epoch))
logger.info(
"==========epoch {} eval start==========".format(epoch))
evaluate(model, eval_iter, epoch)
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."
)
parser.add_argument("--negative_weight", default=1., type=float)
parser.add_argument("--neutral_words_file", default='data/identity.csv')
# if true, use test data instead of val data
parser.add_argument("--test", action='store_true')
# Explanation specific arguments below
# whether run explanation algorithms
parser.add_argument("--explain",
action='store_true',
help='if true, explain test set predictions')
parser.add_argument("--debug", action='store_true')
# which algorithm to run
parser.add_argument("--algo", choices=['soc'])
# the output filename without postfix
parser.add_argument("--output_filename", default='temp.tmp')
# see utils/config.py
parser.add_argument("--use_padding_variant", action='store_true')
parser.add_argument("--mask_outside_nb", action='store_true')
parser.add_argument("--nb_range", type=int)
parser.add_argument("--sample_n", type=int)
# whether use explanation regularization
parser.add_argument("--reg_explanations", action='store_true')
parser.add_argument("--reg_strength", type=float)
parser.add_argument("--reg_mse", action='store_true')
# whether discard other neutral words during regularization. default: False
parser.add_argument("--discard_other_nw",
action='store_false',
dest='keep_other_nw')
# whether remove neutral words when loading datasets
parser.add_argument("--remove_nw", action='store_true')
# if true, generate hierarchical explanations instead of word level outputs.
# Only useful when the --explain flag is also added.
parser.add_argument("--hiex", action='store_true')
parser.add_argument("--hiex_tree_height", default=5, type=int)
# whether add the sentence itself to the sample set in SOC
parser.add_argument("--hiex_add_itself", action='store_true')
# the directory where the lm is stored
parser.add_argument("--lm_dir", default='runs/lm')
# if configured, only generate explanations for instances with given line numbers
parser.add_argument("--hiex_idxs", default=None)
# if true, use absolute values of explanations for hierarchical clustering
parser.add_argument("--hiex_abs", action='store_true')
# if either of the two is true, only generate explanations for positive / negative instances
parser.add_argument("--only_positive", action='store_true')
parser.add_argument("--only_negative", action='store_true')
# stop after generating x explanation
parser.add_argument("--stop", default=100000000, type=int)
# early stopping with decreasing learning rate. 0: direct exit when validation F1 decreases
parser.add_argument("--early_stop", default=5, type=int)
# other external arguments originally here in pytorch_transformers
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
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=32,
type=int,
help="Total batch size for eval.")
parser.add_argument("--validate_steps",
default=200,
type=int,
help="validate once for how many steps")
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")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
combine_args(configs, args)
args = configs
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {
'gab': GabProcessor,
'ws': WSProcessor,
'nyt': NytProcessor,
'MT': MTProcessor,
#'multi-label': multilabel_Processor,
}
output_modes = {
'gab': 'classification',
'ws': 'classification',
'nyt': 'classification'
}
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')
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))
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) and args.do_train:
# raise ValueError("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)
# save configs
f = open(os.path.join(args.output_dir, 'args.json'), 'w')
json.dump(args.__dict__, f, indent=4)
f.close()
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
processor = processors[task_name](configs, tokenizer=tokenizer)
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
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
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
if args.do_train:
model = BertForSequenceClassification.from_pretrained(
args.bert_model, cache_dir=cache_dir, num_labels=num_labels)
else:
model = BertForSequenceClassification.from_pretrained(
args.output_dir, num_labels=num_labels)
model.to(device)
if args.fp16:
model.half()
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
}]
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)
warmup_linear = WarmupLinearSchedule(
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
if args.do_train:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss, tr_reg_loss = 0, 0
tr_reg_cnt = 0
epoch = -1
val_best_f1 = -1
val_best_loss = 1e10
early_stop_countdown = args.early_stop
if args.reg_explanations:
train_lm_dataloder = processor.get_dataloader('train',
configs.train_batch_size)
dev_lm_dataloader = processor.get_dataloader('dev',
configs.train_batch_size)
explainer = SamplingAndOcclusionExplain(
model,
configs,
tokenizer,
device=device,
vocab=tokenizer.vocab,
train_dataloader=train_lm_dataloder,
dev_dataloader=dev_lm_dataloader,
lm_dir=args.lm_dir,
output_path=os.path.join(configs.output_dir,
configs.output_filename),
)
else:
explainer = None
if args.do_train:
epoch = 0
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer, output_mode,
configs)
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":
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)
class_weight = torch.FloatTensor([args.negative_weight, 1]).to(device)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) 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":
loss_fct = CrossEntropyLoss(class_weight)
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
tr_loss += loss.item()
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
# regularize explanations
# NOTE: backward performed inside this function to prevent OOM
if args.reg_explanations:
reg_loss, reg_cnt = explainer.compute_explanation_loss(
input_ids,
input_mask,
segment_ids,
label_ids,
do_backprop=True)
tr_reg_loss += reg_loss # float
tr_reg_cnt += reg_cnt
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
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.get_lr(
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 global_step % args.validate_steps == 0:
val_result = validate(args, model, processor, tokenizer,
output_mode, label_list, device,
num_labels, task_name, tr_loss,
global_step, epoch, explainer)
val_acc, val_f1 = val_result['acc'], val_result['f1']
if val_f1 > val_best_f1:
val_best_f1 = val_f1
if args.local_rank == -1 or torch.distributed.get_rank(
) == 0:
save_model(args, model, tokenizer, num_labels)
else:
# halve the learning rate
for param_group in optimizer.param_groups:
param_group['lr'] *= 0.5
early_stop_countdown -= 1
logger.info(
"Reducing learning rate... Early stop countdown %d"
% early_stop_countdown)
if early_stop_countdown < 0:
break
if early_stop_countdown < 0:
break
epoch += 1
# training finish ############################
# if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
# if not args.explain:
# args.test = True
# validate(args, model, processor, tokenizer, output_mode, label_list, device, num_labels,
# task_name, tr_loss, global_step=0, epoch=-1, explainer=explainer)
# else:
# args.test = True
# explain(args, model, processor, tokenizer, output_mode, label_list, device)
if not args.explain:
args.test = True
print('--Test_args.test: %s' % str(args.test)) #Test_args.test: True
validate(args,
model,
processor,
tokenizer,
output_mode,
label_list,
device,
num_labels,
task_name,
tr_loss,
global_step=888,
epoch=-1,
explainer=explainer)
args.test = False
else:
print('--Test_args.test: %s' % str(args.test)) # Test_args.test: True
args.test = True
explain(args, model, processor, tokenizer, output_mode, label_list,
device)
args.test = False
def main():
parser = ArgumentParser()
parser.add_argument('--pregenerated_data', type=Path, required=True)
parser.add_argument('--output_dir', type=Path, required=True)
parser.add_argument(
"--bert_model",
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--do_lower_case", action="store_true")
parser.add_argument(
"--reduce_memory",
action="store_true",
help=
"Store training data as on-disc memmaps to massively reduce memory usage"
)
parser.add_argument("--epochs",
type=int,
default=3,
help="Number of epochs to train for")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
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("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
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")
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("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--log_every',
type=int,
default=100,
help="Log every X batch")
parser.add_argument("--mlm_only",
action='store_true',
help="Only use MLM objective")
args = parser.parse_args()
assert args.pregenerated_data.is_dir(), \
"--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!"
if args.output_dir.is_dir() and list(args.output_dir.iterdir()):
print(
f"Output directory ({args.output_dir}) already exists and is not empty!"
)
args.output_dir.mkdir(parents=True, exist_ok=True)
logger = util.get_logger(f'{args.output_dir}/exp.txt')
for key, value in vars(args).items():
logger.info('command line argument: %s - %r', key, value)
samples_per_epoch = []
for i in range(args.epochs):
epoch_file = args.pregenerated_data / f"epoch_{i}.json"
metrics_file = args.pregenerated_data / f"epoch_{i}_metrics.json"
if epoch_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch.append(metrics['num_training_examples'])
else:
if i == 0:
exit("No training data was found!")
print(
f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs})."
)
print(
"This script will loop over the available data, but training diversity may be negatively impacted."
)
num_data_epochs = i
break
else:
num_data_epochs = args.epochs
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)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
total_train_examples = 0
for i in range(args.epochs):
# The modulo takes into account the fact that we may loop over limited epochs of data
total_train_examples += samples_per_epoch[i % len(samples_per_epoch)]
num_train_optimization_steps = int(total_train_examples /
args.train_batch_size /
args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
if args.mlm_only:
model = BertForMaskedLM.from_pretrained(args.bert_model)
else:
model = BertForPreTraining.from_pretrained(args.bert_model)
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())
if args.mlm_only:
param_optimizer = [
x for x in param_optimizer if 'bert.pooler' not in x[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
logger.info("***** Running training *****")
logger.info(f" Num examples = {total_train_examples}")
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
model.train()
for epoch in range(args.epochs):
epoch_dataset = PregeneratedDataset(
logger=logger,
epoch=epoch,
training_path=args.pregenerated_data,
tokenizer=tokenizer,
num_data_epochs=num_data_epochs,
mlm_only=args.mlm_only)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
losses = []
with tqdm(total=len(train_dataloader), desc=f"Epoch {epoch}") as pbar:
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
if args.mlm_only:
input_ids, input_mask, segment_ids, lm_label_ids = batch
loss = model(input_ids, segment_ids, input_mask,
lm_label_ids)
else:
input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
loss = model(input_ids, segment_ids, input_mask,
lm_label_ids, is_next)
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
pbar.update(1)
mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps
pbar.set_postfix_str(f"Loss: {mean_loss:.5f}")
losses.append(loss.item())
if step % args.log_every == 0:
logger.info(
f"loss at ep {epoch} batch {step}/{len(train_dataloader)} is {np.mean(losses):.5f}"
)
losses = []
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
# Save a trained model
logger.info("** ** * Saving fine-tuned model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = args.output_dir / f"epoch{epoch}_pytorch_model.bin"
torch.save(model_to_save.state_dict(), str(output_model_file))
def __init__(self, opt, state_dict=None, num_train_step=-1):
self.config = opt
self.updates = state_dict[
'updates'] if state_dict and 'updates' in state_dict else 0
self.train_loss = AverageMeter()
self.network = SANBertNetwork(opt)
# pdb.set_trace()
if state_dict:
new_state = set(self.network.state_dict().keys())
# change to a safer approach
old_keys = [k for k in state_dict['state'].keys()]
for k in old_keys:
if k not in new_state:
print('deleting state:', k)
del state_dict['state'][k]
for k, v in list(self.network.state_dict().items()):
if k not in state_dict['state']:
print('adding missing state:', k)
state_dict['state'][k] = v
# pdb.set_trace()
self.network.load_state_dict(state_dict['state'])
self.mnetwork = nn.DataParallel(
self.network) if opt['multi_gpu_on'] else self.network
self.total_param = sum([
p.nelement() for p in self.network.parameters() if p.requires_grad
])
no_decay = [
'bias', 'gamma', 'beta', 'LayerNorm.bias', 'LayerNorm.weight'
]
optimizer_parameters = [{
'params': [
p for n, p in self.network.named_parameters()
if n not in no_decay
],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in self.network.named_parameters() if n in no_decay],
'weight_decay_rate':
0.0
}]
# note that adamax are modified based on the BERT code
if opt['optimizer'] == 'sgd':
self.optimizer = optim.SGD(optimizer_parameters,
opt['learning_rate'],
weight_decay=opt['weight_decay'])
elif opt['optimizer'] == 'adamax':
self.optimizer = Adamax(optimizer_parameters,
opt['learning_rate'],
warmup=opt['warmup'],
t_total=num_train_step,
max_grad_norm=opt['grad_clipping'],
schedule=opt['warmup_schedule'])
if opt.get('have_lr_scheduler', False):
opt['have_lr_scheduler'] = False
elif opt['optimizer'] == 'adadelta':
self.optimizer = optim.Adadelta(optimizer_parameters,
opt['learning_rate'],
rho=0.95)
elif opt['optimizer'] == 'adam':
self.optimizer = Adam(optimizer_parameters,
lr=opt['learning_rate'],
warmup=opt['warmup'],
t_total=num_train_step,
max_grad_norm=opt['grad_clipping'],
schedule=opt['warmup_schedule'])
if opt.get('have_lr_scheduler', False):
opt['have_lr_scheduler'] = False
else:
raise RuntimeError('Unsupported optimizer: %s' % opt['optimizer'])
if state_dict and 'optimizer' in state_dict:
self.optimizer.load_state_dict(state_dict['optimizer'])
if opt.get('have_lr_scheduler', False):
if opt.get('scheduler_type', 'rop') == 'rop':
self.scheduler = ReduceLROnPlateau(self.optimizer,
mode='max',
factor=opt['lr_gamma'],
patience=3)
elif opt.get('scheduler_type', 'rop') == 'exp':
self.scheduler = ExponentialLR(self.optimizer,
gamma=opt.get('lr_gamma', 0.95))
else:
milestones = [
int(step)
for step in opt.get('multi_step_lr', '10,20,30').split(',')
]
self.scheduler = MultiStepLR(self.optimizer,
milestones=milestones,
gamma=opt.get('lr_gamma'))
else:
self.scheduler = None
self.ema = None
if opt['ema_opt'] > 0:
self.ema = EMA(self.config['ema_gamma'], self.network)
self.para_swapped = False
class MTDNNModel(object):
def __init__(self, opt, state_dict=None, num_train_step=-1):
self.config = opt
self.updates = state_dict[
'updates'] if state_dict and 'updates' in state_dict else 0
self.train_loss = AverageMeter()
self.network = SANBertNetwork(opt)
# pdb.set_trace()
if state_dict:
new_state = set(self.network.state_dict().keys())
# change to a safer approach
old_keys = [k for k in state_dict['state'].keys()]
for k in old_keys:
if k not in new_state:
print('deleting state:', k)
del state_dict['state'][k]
for k, v in list(self.network.state_dict().items()):
if k not in state_dict['state']:
print('adding missing state:', k)
state_dict['state'][k] = v
# pdb.set_trace()
self.network.load_state_dict(state_dict['state'])
self.mnetwork = nn.DataParallel(
self.network) if opt['multi_gpu_on'] else self.network
self.total_param = sum([
p.nelement() for p in self.network.parameters() if p.requires_grad
])
no_decay = [
'bias', 'gamma', 'beta', 'LayerNorm.bias', 'LayerNorm.weight'
]
optimizer_parameters = [{
'params': [
p for n, p in self.network.named_parameters()
if n not in no_decay
],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in self.network.named_parameters() if n in no_decay],
'weight_decay_rate':
0.0
}]
# note that adamax are modified based on the BERT code
if opt['optimizer'] == 'sgd':
self.optimizer = optim.SGD(optimizer_parameters,
opt['learning_rate'],
weight_decay=opt['weight_decay'])
elif opt['optimizer'] == 'adamax':
self.optimizer = Adamax(optimizer_parameters,
opt['learning_rate'],
warmup=opt['warmup'],
t_total=num_train_step,
max_grad_norm=opt['grad_clipping'],
schedule=opt['warmup_schedule'])
if opt.get('have_lr_scheduler', False):
opt['have_lr_scheduler'] = False
elif opt['optimizer'] == 'adadelta':
self.optimizer = optim.Adadelta(optimizer_parameters,
opt['learning_rate'],
rho=0.95)
elif opt['optimizer'] == 'adam':
self.optimizer = Adam(optimizer_parameters,
lr=opt['learning_rate'],
warmup=opt['warmup'],
t_total=num_train_step,
max_grad_norm=opt['grad_clipping'],
schedule=opt['warmup_schedule'])
if opt.get('have_lr_scheduler', False):
opt['have_lr_scheduler'] = False
else:
raise RuntimeError('Unsupported optimizer: %s' % opt['optimizer'])
if state_dict and 'optimizer' in state_dict:
self.optimizer.load_state_dict(state_dict['optimizer'])
if opt.get('have_lr_scheduler', False):
if opt.get('scheduler_type', 'rop') == 'rop':
self.scheduler = ReduceLROnPlateau(self.optimizer,
mode='max',
factor=opt['lr_gamma'],
patience=3)
elif opt.get('scheduler_type', 'rop') == 'exp':
self.scheduler = ExponentialLR(self.optimizer,
gamma=opt.get('lr_gamma', 0.95))
else:
milestones = [
int(step)
for step in opt.get('multi_step_lr', '10,20,30').split(',')
]
self.scheduler = MultiStepLR(self.optimizer,
milestones=milestones,
gamma=opt.get('lr_gamma'))
else:
self.scheduler = None
self.ema = None
if opt['ema_opt'] > 0:
self.ema = EMA(self.config['ema_gamma'], self.network)
self.para_swapped = False
def setup_ema(self):
if self.config['ema_opt']:
self.ema.setup()
def update_ema(self):
if self.config['ema_opt']:
self.ema.update()
def eval(self):
if self.config['ema_opt']:
self.ema.swap_parameters()
self.para_swapped = True
def train(self):
if self.para_swapped:
self.ema.swap_parameters()
self.para_swapped = False
def update(self, batch_meta, batch_data):
self.network.train()
labels = batch_data[batch_meta['label']]
# print('data size:',batch_data[batch_meta['token_id']].size())
if batch_meta['pairwise']:
labels = labels.contiguous().view(-1,
batch_meta['pairwise_size'])[:,
0]
if self.config['cuda']:
y = Variable(labels.cuda(async=True), requires_grad=False)
else:
y = Variable(labels, requires_grad=False)
task_id = batch_meta['task_id']
task_type = batch_meta['task_type']
inputs = batch_data[:batch_meta['input_len']]
if len(inputs) == 3:
inputs.append(None)
inputs.append(None)
inputs.append(task_id)
# pdb.set_trace()
logits = self.mnetwork(*inputs)
if batch_meta['pairwise']:
logits = logits.view(-1, batch_meta['pairwise_size'])
# pdb.set_trace()
if task_type > 0:
if self.config['answer_relu']:
logits = F.relu(logits)
loss = F.mse_loss(logits.squeeze(1), y)
else:
loss = F.cross_entropy(logits, y)
if self.config['mediqa_pairloss'] is not None and batch_meta[
'dataset_name'] in mediqa_name_list:
# print(logits)
# print(batch_data[batch_meta['rank_label']].size())
# input('ha')
logits = logits.squeeze().view(-1, 2)
# print(batch_data[batch_meta['rank_label']])
rank_y = batch_data[batch_meta['rank_label']].view(-1, 2)
# print(rank_y)
if self.config['mediqa_pairloss'] == 'hinge':
# print(logits)
first_logit, second_logit = logits.split(1, dim=1)
# print(first_logit,second_logit)
# pdb.set_trace()
rank_y = (2 * rank_y - 1).to(torch.float32)
rank_y = rank_y[:, 0]
pairwise_loss = F.margin_ranking_loss(
first_logit.squeeze(1),
second_logit.squeeze(1),
rank_y,
margin=self.config['hinge_lambda'])
else:
# pdb.set_trace()
pairwise_loss = F.cross_entropy(logits, rank_y[:, 1])
# print('pairwise_loss:',pairwise_loss,'mse loss:',loss)
loss += pairwise_loss
self.train_loss.update(loss.item(), logits.size(0))
self.optimizer.zero_grad()
loss.backward()
if self.config['global_grad_clipping'] > 0:
torch.nn.utils.clip_grad_norm_(self.network.parameters(),
self.config['global_grad_clipping'])
self.optimizer.step()
self.updates += 1
self.update_ema()
def predict(self, batch_meta, batch_data):
self.network.eval()
task_id = batch_meta['task_id']
task_type = batch_meta['task_type']
inputs = batch_data[:batch_meta['input_len']]
if len(inputs) == 3:
inputs.append(None)
inputs.append(None)
inputs.append(task_id)
score = self.mnetwork(*inputs)
gold_label = batch_meta['label']
if batch_meta['pairwise']:
score = score.contiguous().view(-1, batch_meta['pairwise_size'])
if task_type < 1:
score = F.softmax(score, dim=1)
score = score.data.cpu()
score = score.numpy()
predict = np.zeros(score.shape, dtype=int)
if task_type < 1:
positive = np.argmax(score, axis=1)
for idx, pos in enumerate(positive):
predict[idx, pos] = 1
predict = predict.reshape(-1).tolist()
score = score.reshape(-1).tolist()
return score, predict, batch_meta['true_label']
else:
if task_type < 1:
score = F.softmax(score, dim=1)
# pdb.set_trace()
score = score.data.cpu()
score = score.numpy()
if task_type < 1:
predict = np.argmax(score, axis=1).tolist()
else:
predict = np.greater(
score,
2.0 + self.config['mediqa_score_offset']).astype(int)
gold_label = np.greater(
batch_meta['label'],
2.00001 + self.config['mediqa_score_offset']).astype(int)
predict = predict.reshape(-1).tolist()
gold_label = gold_label.reshape(-1).tolist()
# print('predict:',predict,score)
score = score.reshape(-1).tolist()
return score, predict, gold_label
def save(self, filename):
network_state = dict([(k, v.cpu())
for k, v in self.network.state_dict().items()])
ema_state = dict([
(k, v.cpu()) for k, v in self.ema.model.state_dict().items()
]) if self.ema is not None else dict()
params = {
'state': network_state,
'optimizer': self.optimizer.state_dict(),
'ema': ema_state,
'config': self.config,
}
torch.save(params, filename)
logger.info('model saved to {}'.format(filename))
def cuda(self):
self.network.cuda()
if self.config['ema_opt']:
self.ema.cuda()
[n for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'names':
[n for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
args.steps_per_epoch = sum(num_batchs_per_task)
args.total_steps = args.steps_per_epoch * args.epoch_num
optimizer = BertAdam(params=optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup,
t_total=args.total_steps,
max_grad_norm=args.clip_grad,
schedule=args.schedule)
logging.info('Loading graph and entity linking...')
graph = pickle.load(open('graph/graph.pkl', 'rb'))
entity_linking = pickle.load(open('graph/entity_linking.pkl', 'rb'))
if args.do_train_and_eval:
# Train and evaluate
best_acc = 0
for epoch in range(args.epoch_num):
## Train
model.train()
t = trange(args.steps_per_epoch,
desc='Epoch {} -Train'.format(epoch))
def main():
parser = argparse.ArgumentParser(fromfile_prefix_chars="@")
parser.add_argument("--pregenerated_data",
type=Path,
required=True,
help="The input train corpus.")
parser.add_argument("--epochs", type=int, required=True)
parser.add_argument("--bert_model", type=str, required=True)
parser.add_argument("--bert_config_file",
type=str,
default="bert_config.json")
parser.add_argument("--vocab_file", type=str, default="senti_vocab.txt")
parser.add_argument('--output_dir', type=Path, required=True)
parser.add_argument("--model_name", type=str, default="senti_base_model")
parser.add_argument(
"--reduce_memory",
action="store_true",
help=
"Store training data as on-disc memmaps to massively reduce memory usage"
)
parser.add_argument("--world_size", type=int, default=4)
parser.add_argument("--start_rank", type=int, default=0)
parser.add_argument("--server", type=str, default="tcp://127.0.0.1:1234")
parser.add_argument("--load_model", action="store_true")
parser.add_argument("--load_model_name", type=str, default="large_model")
parser.add_argument("--save_step", type=int, default=100000)
parser.add_argument("--train_batch_size",
default=4,
type=int,
help="Total batch size for training.")
parser.add_argument("--learning_rate",
default=1e-4,
type=float,
help="The initial learning rate for Adam.")
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(
"--do_lower_case",
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--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 accumualte 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()
assert args.pregenerated_data.is_dir(), \
"--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!"
print("local_rank : ", args.local_rank)
samples_per_epoch = []
for i in range(args.epochs):
epoch_file = args.pregenerated_data / f"epoch_{i}.json"
metrics_file = args.pregenerated_data / f"epoch_{i}_metrics.json"
if epoch_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch.append(metrics['num_training_examples'])
else:
if i == 0:
exit("No training data was found!")
print(
f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs})."
)
print(
"This script will loop over the available data, but training diversity may be negatively impacted."
)
num_data_epochs = i
break
else:
num_data_epochs = args.epochs
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',
init_method=args.server,
rank=args.local_rank +
args.start_rank,
world_size=args.world_size)
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 args.output_dir.is_dir() and list(args.output_dir.iterdir()):
logger.warning(
f"Output directory ({args.output_dir}) already exists and is not empty!"
)
args.output_dir.mkdir(parents=True, exist_ok=True)
tokenizer = Tokenizer(
os.path.join(args.bert_model, "senti_vocab.txt"),
os.path.join(args.bert_model, "RoBERTa_Sentiment_kor"))
total_train_examples = 0
for i in range(args.epochs):
# The modulo takes into account the fact that we may loop over limited epochs of data
total_train_examples += samples_per_epoch[i % len(samples_per_epoch)]
num_train_optimization_steps = math.ceil(total_train_examples /
args.train_batch_size /
args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = math.ceil(
num_train_optimization_steps / torch.distributed.get_world_size())
# Prepare model
config = BertConfig.from_json_file(
os.path.join(args.bert_model, args.bert_config_file))
logger.info('{}'.format(config))
###############################################
# Load Model
if args.load_model:
load_model_name = os.path.join(args.output_dir, args.load_model_name)
model = BertForPreTraining.from_pretrained(
args.bert_model,
state_dict=torch.load(load_model_name)["state_dict"])
else:
model = BertForPreTraining(config)
###############################################
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
model = DDP(model)
except ImportError:
from torch.nn.parallel import DistributedDataParallel as DDP
model = DDP(model,
device_ids=[args.local_rank],
output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
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
}]
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)
warmup_linear = WarmupLinearSchedule(
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
epoch0 = 0
global_step = 0
if args.load_model:
###############################################
# Load Model
logger.info(f"***** Load Model {args.load_model_name} *****")
loaded_states = torch.load(os.path.join(args.output_dir,
args.load_model_name),
map_location=device)
optimizer.load_state_dict(loaded_states["optimizer"])
regex = re.compile(r'\d+epoch')
epoch0 = int(
regex.findall(args.load_model_name)[-1].replace('epoch', ''))
logger.info('extract {} -> epoch0 : {}'.format(args.load_model_name,
epoch0))
###############################################
logger.info("***** Running training *****")
logger.info(f" Num examples = {total_train_examples}")
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
model.train()
# model.eval()
for epoch in range(epoch0, args.epochs):
epoch_dataset = PregeneratedDataset(
epoch=epoch,
training_path=args.pregenerated_data,
tokenizer=tokenizer,
num_data_epochs=num_data_epochs,
reduce_memory=args.reduce_memory)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
with tqdm(total=len(train_dataloader), desc='training..') as pbar:
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, lm_label_ids = batch
loss = model(input_ids, input_mask, lm_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
pbar.update(1)
mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps
if (step + 1) % 50 == 0:
pbar.set_description(
"Epoch = {}, global_step = {}, loss = {:.5f}".format(
epoch, global_step + 1, mean_loss))
logger.info(
"Epoch = {}, global_step = {}, loss = {:.5f}".format(
epoch, global_step + 1, mean_loss))
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.get_lr(
global_step, 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 (step + 1) % args.save_step == 0:
if args.local_rank == -1 or args.local_rank == 0:
logger.info(
"** ** * Saving {} - step model ** ** * ".format(
global_step))
output_model_file = os.path.join(
args.output_dir,
args.model_name + "_{}step".format(global_step))
model_to_save = model.module if hasattr(
model, 'module') else model
state = {
"state_dict": model_to_save.state_dict(),
"optimizer": optimizer.state_dict()
}
torch.save(state, output_model_file)
if args.local_rank == -1 or args.local_rank == 0:
logger.info(
"** ** * Saving {} - epoch model ** ** * ".format(epoch))
output_model_file = os.path.join(
args.output_dir,
args.model_name + "_{}epoch".format(epoch + 1))
model_to_save = model.module if hasattr(model, 'module') else model
state = {
"state_dict": model_to_save.state_dict(),
"optimizer": optimizer.state_dict()
}
torch.save(state, output_model_file)
def main(*_, **kwargs):
use_cuda = torch.cuda.is_available() and kwargs["device"] >= 0
device = torch.device("cuda:" +
str(kwargs["device"]) if use_cuda else "cpu")
if use_cuda:
torch.cuda.set_device(device)
kwargs["use_cuda"] = use_cuda
neptune.create_experiment(
name="bert-span-parser",
upload_source_files=[],
params={
k: str(v) if isinstance(v, bool) else v
for k, v in kwargs.items()
},
)
logger.info("Settings: {}", json.dumps(kwargs,
indent=2,
ensure_ascii=False))
# For reproducibility
os.environ["PYTHONHASHSEED"] = str(kwargs["seed"])
random.seed(kwargs["seed"])
np.random.seed(kwargs["seed"])
torch.manual_seed(kwargs["seed"])
torch.cuda.manual_seed_all(kwargs["seed"])
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Prepare and load data
tokenizer = BertTokenizer.from_pretrained(kwargs["bert_model"],
do_lower_case=False)
logger.info("Loading data...")
train_treebank = load_trees(kwargs["train_file"])
dev_treebank = load_trees(kwargs["dev_file"])
test_treebank = load_trees(kwargs["test_file"])
logger.info(
"Loaded {:,} train, {:,} dev, and {:,} test examples!",
len(train_treebank),
len(dev_treebank),
len(test_treebank),
)
logger.info("Preprocessing data...")
train_parse = [tree.convert() for tree in train_treebank]
train_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in train_parse]
dev_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in dev_treebank]
test_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in test_treebank]
logger.info("Data preprocessed!")
logger.info("Preparing data for training...")
tags = []
labels = []
for tree in train_parse:
nodes = [tree]
while nodes:
node = nodes.pop()
if isinstance(node, InternalParseNode):
labels.append(node.label)
nodes.extend(reversed(node.children))
else:
tags.append(node.tag)
tag_encoder = LabelEncoder()
tag_encoder.fit(tags, reserved_labels=["[PAD]", "[UNK]"])
label_encoder = LabelEncoder()
label_encoder.fit(labels, reserved_labels=[()])
logger.info("Data prepared!")
# Settings
num_train_optimization_steps = kwargs["num_epochs"] * (
(len(train_parse) - 1) // kwargs["batch_size"] + 1)
kwargs["batch_size"] //= kwargs["gradient_accumulation_steps"]
logger.info("Creating dataloaders for training...")
train_dataloader, train_features = create_dataloader(
sentences=train_sentences,
batch_size=kwargs["batch_size"],
tag_encoder=tag_encoder,
tokenizer=tokenizer,
is_eval=False,
)
dev_dataloader, dev_features = create_dataloader(
sentences=dev_sentences,
batch_size=kwargs["batch_size"],
tag_encoder=tag_encoder,
tokenizer=tokenizer,
is_eval=True,
)
test_dataloader, test_features = create_dataloader(
sentences=test_sentences,
batch_size=kwargs["batch_size"],
tag_encoder=tag_encoder,
tokenizer=tokenizer,
is_eval=True,
)
logger.info("Dataloaders created!")
# Initialize model
model = ChartParser.from_pretrained(
kwargs["bert_model"],
tag_encoder=tag_encoder,
label_encoder=label_encoder,
lstm_layers=kwargs["lstm_layers"],
lstm_dim=kwargs["lstm_dim"],
tag_embedding_dim=kwargs["tag_embedding_dim"],
label_hidden_dim=kwargs["label_hidden_dim"],
dropout_prob=kwargs["dropout_prob"],
)
model.to(device)
# Prepare optimizer
param_optimizers = list(model.named_parameters())
if kwargs["freeze_bert"]:
for p in model.bert.parameters():
p.requires_grad = False
param_optimizers = [(n, p) for n, p in param_optimizers
if p.requires_grad]
# Hack to remove pooler, which is not used thus it produce None grad that break apex
param_optimizers = [n for n in param_optimizers if "pooler" not in n[0]]
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in param_optimizers
if not any(nd in n for nd in no_decay)
],
"weight_decay":
0.01,
},
{
"params": [
p for n, p in param_optimizers
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
optimizer = BertAdam(
optimizer_grouped_parameters,
lr=kwargs["learning_rate"],
warmup=kwargs["warmup_proportion"],
t_total=num_train_optimization_steps,
)
if kwargs["fp16"]:
model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
pretrained_model_file = os.path.join(kwargs["output_dir"], MODEL_FILENAME)
if kwargs["do_eval"]:
assert os.path.isfile(
pretrained_model_file), "Pretrained model file does not exist!"
logger.info("Loading pretrained model from {}", pretrained_model_file)
# Load model from file
params = torch.load(pretrained_model_file, map_location=device)
model.load_state_dict(params["model"])
logger.info(
"Loaded pretrained model (Epoch: {:,}, Fscore: {:.2f})",
params["epoch"],
params["fscore"],
)
eval_score = eval(
model=model,
eval_dataloader=test_dataloader,
eval_features=test_features,
eval_trees=test_treebank,
eval_sentences=test_sentences,
tag_encoder=tag_encoder,
device=device,
)
neptune.send_metric("test_eval_precision", eval_score.precision())
neptune.send_metric("test_eval_recall", eval_score.recall())
neptune.send_metric("test_eval_fscore", eval_score.fscore())
tqdm.write("Evaluation score: {}".format(str(eval_score)))
else:
# Training phase
global_steps = 0
start_epoch = 0
best_dev_fscore = 0
if kwargs["preload"] or kwargs["resume"]:
assert os.path.isfile(
pretrained_model_file), "Pretrained model file does not exist!"
logger.info("Resuming model from {}", pretrained_model_file)
# Load model from file
params = torch.load(pretrained_model_file, map_location=device)
model.load_state_dict(params["model"])
if kwargs["resume"]:
optimizer.load_state_dict(params["optimizer"])
torch.cuda.set_rng_state_all([
state.cpu()
for state in params["torch_cuda_random_state_all"]
])
torch.set_rng_state(params["torch_random_state"].cpu())
np.random.set_state(params["np_random_state"])
random.setstate(params["random_state"])
global_steps = params["global_steps"]
start_epoch = params["epoch"] + 1
best_dev_fscore = params["fscore"]
else:
assert not os.path.isfile(
pretrained_model_file
), "Please remove or move the pretrained model file to another place!"
for epoch in trange(start_epoch, kwargs["num_epochs"], desc="Epoch"):
model.train()
train_loss = 0
num_train_steps = 0
for step, (indices, *_) in enumerate(
tqdm(train_dataloader, desc="Iteration")):
ids, attention_masks, tags, sections, trees, sentences = prepare_batch_input(
indices=indices,
features=train_features,
trees=train_parse,
sentences=train_sentences,
tag_encoder=tag_encoder,
device=device,
)
loss = model(
ids=ids,
attention_masks=attention_masks,
tags=tags,
sections=sections,
sentences=sentences,
gold_trees=trees,
)
if kwargs["gradient_accumulation_steps"] > 1:
loss /= kwargs["gradient_accumulation_steps"]
if kwargs["fp16"]:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
train_loss += loss.item()
num_train_steps += 1
if (step + 1) % kwargs["gradient_accumulation_steps"] == 0:
optimizer.step()
optimizer.zero_grad()
global_steps += 1
# Write logs
neptune.send_metric("train_loss", epoch,
train_loss / num_train_steps)
neptune.send_metric("global_steps", epoch, global_steps)
tqdm.write(
"Epoch: {:,} - Train loss: {:.4f} - Global steps: {:,}".format(
epoch, train_loss / num_train_steps, global_steps))
# Evaluate
eval_score = eval(
model=model,
eval_dataloader=dev_dataloader,
eval_features=dev_features,
eval_trees=dev_treebank,
eval_sentences=dev_sentences,
tag_encoder=tag_encoder,
device=device,
)
neptune.send_metric("eval_precision", epoch,
eval_score.precision())
neptune.send_metric("eval_recall", epoch, eval_score.recall())
neptune.send_metric("eval_fscore", epoch, eval_score.fscore())
tqdm.write("Epoch: {:,} - Evaluation score: {}".format(
epoch, str(eval_score)))
# Save best model
if eval_score.fscore() > best_dev_fscore:
best_dev_fscore = eval_score.fscore()
tqdm.write("** Saving model...")
os.makedirs(kwargs["output_dir"], exist_ok=True)
torch.save(
{
"epoch":
epoch,
"global_steps":
global_steps,
"fscore":
best_dev_fscore,
"random_state":
random.getstate(),
"np_random_state":
np.random.get_state(),
"torch_random_state":
torch.get_rng_state(),
"torch_cuda_random_state_all":
torch.cuda.get_rng_state_all(),
"optimizer":
optimizer.state_dict(),
"model": (model.module if hasattr(model, "module") else
model).state_dict(),
},
pretrained_model_file,
)
tqdm.write(
"** Best evaluation fscore: {:.2f}".format(best_dev_fscore))
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}
]
num_train_steps = None
if args.do_train:
num_train_steps = int(len(data.train_data) / args.batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
args.batch_size = int(args.batch_size / args.gradient_accumulation_steps) * n_gpu
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
## Using half precision for faster training
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError("Haven't install apex!!!")
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_level)
# For distributed training
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
def main():
parser = argparse.ArgumentParser()
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,
help="The output directory where the model checkpoints will be written."
)
parser.add_argument("--train_file", default=None, type=str)
parser.add_argument("--val_file", default=None, type=str)
parser.add_argument("--test_file", default=None, type=str)
parser.add_argument("--test_output", default=None, type=str)
parser.add_argument("--label_vocab", default=None, type=str, required=True)
parser.add_argument("--punc_set", default='PU', type=str)
parser.add_argument("--has_confidence", action='store_true')
parser.add_argument("--only_save_bert", action='store_true')
parser.add_argument("--arc_space", default=512, type=int)
parser.add_argument("--type_space", default=128, type=int)
parser.add_argument("--log_file", default=None, type=str)
## 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_predict",
action='store_true',
help="Whether to run predict on the test set.")
parser.add_argument("--do_greedy_predict",
action='store_true',
help="Whether to run predict on the test set.")
parser.add_argument("--do_ensemble_predict",
action='store_true',
help="Whether to run predict on the test 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("--test_batch_size",
default=8,
type=int,
help="Total batch size for test.")
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()
if args.log_file is None:
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
else:
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
filename=args.log_file,
filemode='w',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO)
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_predict and not args.do_greedy_predict and not args.do_ensemble_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
assert args.output_dir is not None
if args.do_train and os.path.exists(args.output_dir) and os.listdir(
args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if args.do_train and not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
label_vocab, label_vocab2idx = load_label_vocab(args.label_vocab)
punc_set = set(
args.punc_set.split(',')) if args.punc_set is not None else None
train_examples = None
num_train_optimization_steps = None
if args.do_train:
assert args.train_file is not None
train_examples = read_conll_examples(
args.train_file,
is_training=True,
has_confidence=args.has_confidence)
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(
)
if args.do_train or args.do_predict or args.do_greedy_predict:
# load the pretrained model
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
model = BertForDependencyParsing.from_pretrained(
args.bert_model,
cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE),
'distributed_{}'.format(args.local_rank)),
arc_space=args.arc_space,
type_space=args.type_space,
num_labels=len(label_vocab))
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)
#
parser = model.module if hasattr(model, 'module') else model
elif args.do_ensemble_predict:
bert_models = args.bert_model.split(',')
assert len(bert_models) > 1
tokenizer = BertTokenizer.from_pretrained(
bert_models[0], do_lower_case=args.do_lower_case)
models = []
for bm in bert_models:
model = BertForDependencyParsing.from_pretrained(
bm,
cache_dir=os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE),
'distributed_{}'.format(args.local_rank)),
arc_space=args.arc_space,
type_space=args.type_space,
num_labels=len(label_vocab))
model.to(device)
model.eval()
models.append(model)
parser = models[0].module if hasattr(models[0],
'module') else models[0]
# Prepare optimizer
if args.do_train:
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
# !!! NOTE why?
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)
warmup_linear = WarmupLinearSchedule(
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
# start training loop
if args.do_train:
global_step = 0
train_features = convert_examples_to_features(
train_examples,
tokenizer,
args.max_seq_length,
label_vocab2idx,
True,
has_confidence=args.has_confidence)
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.float32)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_lengths = torch.tensor([f.seq_len for f in train_features],
dtype=torch.long)
all_heads = torch.tensor([f.heads for f in train_features],
dtype=torch.long)
all_labels = torch.tensor([f.labels for f in train_features],
dtype=torch.long)
if args.has_confidence:
all_confidence = torch.tensor(
[f.confidence for f in train_features], dtype=torch.float32)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_lengths, all_heads,
all_labels, all_confidence)
else:
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_lengths, all_heads,
all_labels)
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)
if args.do_eval:
assert args.val_file is not None
eval_examples = read_conll_examples(args.val_file,
is_training=False,
has_confidence=False)
eval_features = convert_examples_to_features(eval_examples,
tokenizer,
args.max_seq_length,
label_vocab2idx,
False,
has_confidence=False)
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_example_ids = torch.tensor(
[f.example_id for f in eval_features], dtype=torch.long)
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.float32)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features], dtype=torch.long)
all_lengths = torch.tensor([f.seq_len for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_lengths,
all_example_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)
best_uas = 0
best_las = 0
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
logger.info("Training epoch: {}".format(epoch))
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
model.train()
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
if args.has_confidence:
input_ids, input_mask, segment_ids, lengths, heads, label_ids, confidence = batch
else:
confidence = None
input_ids, input_mask, segment_ids, lengths, heads, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, heads,
label_ids, confidence)
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.get_lr(
global_step, 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 global_step % 100 == 0:
logger.info("Training loss: {}, global step: {}".format(
tr_loss / nb_tr_steps, global_step))
# we eval every epoch
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
logger.info("***** Running evaluation *****")
model.eval()
eval_predict_words, eval_predict_postags, eval_predict_heads, eval_predict_labels = [],[],[],[]
for input_ids, input_mask, segment_ids, lengths, example_ids in tqdm(
eval_dataloader, desc="Evaluating"):
example_ids = example_ids.numpy()
batch_words = [
eval_features[eid].example.sentence
for eid in example_ids
]
batch_postags = [
eval_features[eid].example.postags
for eid in example_ids
]
batch_word_index = [
eval_features[eid].word_index for eid in example_ids
] # token -> word
batch_token_starts = [
eval_features[eid].token_starts for eid in example_ids
] # word -> token start
batch_heads = [
eval_features[eid].example.heads for eid in example_ids
]
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
heads = heads.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
# tmp_eval_loss = model(input_ids, segment_ids, input_mask, heads, label_ids)
energy = model(input_ids, segment_ids, input_mask)
heads_pred, labels_pred = parser.decode_MST(
energy.cpu().numpy(),
lengths.numpy(),
leading_symbolic=0,
labeled=True)
# we convert the subword dependency parsing to word dependency parsing just the word and token start map
pred_heads = []
pred_labels = []
for i in range(len(batch_word_index)):
word_index = batch_word_index[i]
token_starts = batch_token_starts[i]
hpd = []
lpd = []
for j in range(len(token_starts)):
if j == 0: #[CLS]
continue
elif j == len(token_starts) - 1: # [SEP]
continue
else:
hpd.append(
word_index[heads_pred[i, token_starts[j]]])
lpd.append(
label_vocab[labels_pred[i,
token_starts[j]]])
pred_heads.append(hpd)
pred_labels.append(lpd)
eval_predict_words += batch_words
eval_predict_postags += batch_postags
eval_predict_heads += pred_heads
eval_predict_labels += pred_labels
eval_output_file = os.path.join(args.output_dir, 'eval.pred')
write_conll_examples(eval_predict_words, eval_predict_postags,
eval_predict_heads, eval_predict_labels,
eval_output_file)
eval_f = os.popen(
"python scripts/eval_nlpcc_dp.py " + args.val_file + " " +
eval_output_file, "r")
result_text = eval_f.read().strip()
logger.info("***** Eval results *****")
logger.info(result_text)
eval_f.close()
eval_res = re.findall(
r'UAS = \d+/\d+ = ([\d\.]+), LAS = \d+/\d+ = ([\d\.]+)',
result_text)
assert len(eval_res) > 0
eval_res = eval_res[0]
eval_uas = float(eval_res[0])
eval_las = float(eval_res[1])
# save model
if best_las < eval_las or (eval_las == best_las
and best_uas < eval_uas):
best_uas = eval_uas
best_las = eval_las
logger.info(
"new best uas %.2f%% las %.2f%%, saving models.",
best_uas, best_las)
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
# If we save using the predefined names, we can load using `from_pretrained`
output_model_file = os.path.join(args.output_dir,
WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir,
CONFIG_NAME)
model_dict = model_to_save.state_dict()
if args.only_save_bert:
model_dict = {
k: v
for k, v in model_dict.items() if 'bert.' in k
}
torch.save(model_dict, output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
# start predict
if args.do_predict:
model.eval()
assert args.test_file is not None
test_examples = read_conll_examples(args.test_file,
is_training=False,
has_confidence=False)
test_features = convert_examples_to_features(test_examples,
tokenizer,
args.max_seq_length,
label_vocab2idx,
False,
has_confidence=False)
logger.info("***** Running prediction *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.test_batch_size)
all_example_ids = torch.tensor([f.example_id for f in test_features],
dtype=torch.long)
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.float32)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
all_lengths = torch.tensor([f.seq_len for f in test_features],
dtype=torch.long)
test_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_lengths,
all_example_ids)
# Run prediction for full data
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.test_batch_size)
test_predict_words, test_predict_postags, test_predict_heads, test_predict_labels = [],[],[],[]
for batch_id, batch in enumerate(
tqdm(test_dataloader, desc="Predicting")):
input_ids, input_mask, segment_ids, lengths, example_ids = batch
example_ids = example_ids.numpy()
batch_words = [
test_features[eid].example.sentence for eid in example_ids
]
batch_postags = [
test_features[eid].example.postags for eid in example_ids
]
batch_word_index = [
test_features[eid].word_index for eid in example_ids
] # token -> word
batch_token_starts = [
test_features[eid].token_starts for eid in example_ids
] # word -> token start
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
lengths = lengths.numpy()
with torch.no_grad():
energy = model(input_ids, segment_ids, input_mask)
heads_pred, labels_pred = parser.decode_MST(energy.cpu().numpy(),
lengths,
leading_symbolic=0,
labeled=True)
pred_heads = []
pred_labels = []
for i in range(len(batch_word_index)):
word_index = batch_word_index[i]
token_starts = batch_token_starts[i]
hpd = []
lpd = []
for j in range(len(token_starts)):
if j == 0: #[CLS]
continue
elif j == len(token_starts) - 1: # [SEP]
continue
else:
hpd.append(word_index[heads_pred[i, token_starts[j]]])
lpd.append(label_vocab[labels_pred[i,
token_starts[j]]])
pred_heads.append(hpd)
pred_labels.append(lpd)
test_predict_words += batch_words
test_predict_postags += batch_postags
test_predict_heads += pred_heads
test_predict_labels += pred_labels
assert args.test_output is not None
write_conll_examples(test_predict_words, test_predict_postags,
test_predict_heads, test_predict_labels,
args.test_output)
if args.do_greedy_predict:
model.eval()
assert args.test_file is not None
test_examples = read_conll_examples(args.test_file,
is_training=False,
has_confidence=False)
test_features = convert_examples_to_features(test_examples,
tokenizer,
args.max_seq_length,
label_vocab2idx,
False,
has_confidence=False)
logger.info("***** Running prediction *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.test_batch_size)
all_example_ids = torch.tensor([f.example_id for f in test_features],
dtype=torch.long)
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.float32)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
all_lengths = torch.tensor([f.seq_len for f in test_features],
dtype=torch.long)
test_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_lengths,
all_example_ids)
# Run prediction for full data
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.test_batch_size)
test_predict_words, test_predict_postags, test_predict_heads, test_predict_labels = [],[],[],[]
for batch_id, batch in enumerate(
tqdm(test_dataloader, desc="Predicting")):
input_ids, input_mask, segment_ids, lengths, example_ids = batch
example_ids = example_ids.numpy()
batch_words = [
test_features[eid].example.sentence for eid in example_ids
]
batch_postags = [
test_features[eid].example.postags for eid in example_ids
]
batch_word_index = [
test_features[eid].word_index for eid in example_ids
] # token -> word
batch_token_starts = [
test_features[eid].token_starts for eid in example_ids
] # word -> token start
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
lengths = lengths.numpy()
with torch.no_grad():
heads_pred, labels_pred = model(input_ids,
segment_ids,
input_mask,
greedy_inference=True)
pred_heads = []
pred_labels = []
for i in range(len(batch_word_index)):
word_index = batch_word_index[i]
token_starts = batch_token_starts[i]
hpd = []
lpd = []
for j in range(len(token_starts)):
if j == 0: #[CLS]
continue
elif j == len(token_starts) - 1: # [SEP]
continue
else:
hpd.append(word_index[heads_pred[i, token_starts[j]]])
lpd.append(label_vocab[labels_pred[i,
token_starts[j]]])
pred_heads.append(hpd)
pred_labels.append(lpd)
test_predict_words += batch_words
test_predict_postags += batch_postags
test_predict_heads += pred_heads
test_predict_labels += pred_labels
assert args.test_output is not None
write_conll_examples(test_predict_words, test_predict_postags,
test_predict_heads, test_predict_labels,
args.test_output)
if args.do_ensemble_predict:
assert args.test_file is not None
test_examples = read_conll_examples(args.test_file,
is_training=False,
has_confidence=False)
test_features = convert_examples_to_features(test_examples,
tokenizer,
args.max_seq_length,
label_vocab2idx,
False,
has_confidence=False)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(test_examples))
logger.info(" Batch size = %d", args.test_batch_size)
all_example_ids = torch.tensor([f.example_id for f in test_features],
dtype=torch.long)
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.float32)
all_segment_ids = torch.tensor([f.segment_ids for f in test_features],
dtype=torch.long)
all_lengths = torch.tensor([f.seq_len for f in test_features],
dtype=torch.long)
test_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_lengths,
all_example_ids)
# Run prediction for full data
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.test_batch_size)
test_predict_words, test_predict_postags, test_predict_heads, test_predict_labels = [],[],[],[]
for batch_id, batch in enumerate(
tqdm(test_dataloader, desc="Predicting")):
input_ids, input_mask, segment_ids, lengths, example_ids = batch
example_ids = example_ids.numpy()
batch_words = [
test_features[eid].example.sentence for eid in example_ids
]
batch_postags = [
test_features[eid].example.postags for eid in example_ids
]
batch_word_index = [
test_features[eid].word_index for eid in example_ids
] # token -> word
batch_token_starts = [
test_features[eid].token_starts for eid in example_ids
] # word -> token start
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
lengths = lengths.numpy()
with torch.no_grad():
energy_sum = None
for model in models:
energy = model(input_ids, segment_ids, input_mask)
if energy_sum is None:
energy_sum = energy
else:
energy_sum = energy_sum + energy
energy_sum = energy_sum / len(models)
heads_pred, labels_pred = parser.decode_MST(
energy_sum.cpu().numpy(),
lengths,
leading_symbolic=0,
labeled=True)
pred_heads = []
pred_labels = []
for i in range(len(batch_word_index)):
word_index = batch_word_index[i]
token_starts = batch_token_starts[i]
hpd = []
lpd = []
for j in range(len(token_starts)):
if j == 0: #[CLS]
continue
elif j == len(token_starts) - 1: # [SEP]
continue
else:
hpd.append(word_index[heads_pred[i, token_starts[j]]])
lpd.append(label_vocab[labels_pred[i,
token_starts[j]]])
pred_heads.append(hpd)
pred_labels.append(lpd)
test_predict_words += batch_words
test_predict_postags += batch_postags
test_predict_heads += pred_heads
test_predict_labels += pred_labels
assert args.test_output is not None
write_conll_examples(test_predict_words, test_predict_postags,
test_predict_heads, test_predict_labels,
args.test_output)
def create_optimizer(model,
args,
num_train_steps=None,
init_spec=None,
no_decay=['bias', 'LayerNorm.weight']):
# Prepare optimizer
if args.fp16:
dcnt = torch.cuda.device_count()
if args.no_even_grad:
param_optimizer = [(n, param.detach().clone().type(torch.cuda.FloatTensor).\
requires_grad_()) for i,(n,param) in enumerate(model.named_parameters())]
else:
total_size = sum(np.prod(p.size()) for p in model.parameters())
quota = {i: 0 for i in range(dcnt)}
quota[0] = total_size // (dcnt * 2)
param_optimizer = []
for i, (n, param) in enumerate(model.named_parameters()):
ps = np.prod(param.size())
index = list(sorted(quota.items(), key=lambda x: x[1]))[0][0]
quota[index] += ps
cp = param.clone().type(torch.cuda.FloatTensor).detach().to(
'cuda:{}'.format(index)).requires_grad_()
param_optimizer += [(n, cp)]
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 = [(n, p) for n, p in model.named_parameters()]
group0 = dict(params=[], weight_decay_rate=args.weight_decay, names=[])
group1 = dict(params=[], weight_decay_rate=0.00, names=[])
for (n, p) in param_optimizer:
if not any(nd in n for nd in no_decay):
group0['params'].append(p)
group0['names'].append(n)
else:
group1['params'].append(p)
group1['names'].append(n)
optimizer_grouped_parameters = [group0, group1]
t_total = num_train_steps
optimizer = None
if t_total:
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
b1=args.adam_beta1,
b2=args.adam_beta2,
v1=args.qhadam_v1,
v2=args.qhadam_v2,
lr_ends=args.lr_schedule_ends,
e=args.epsilon,
warmup=args.warmup_proportion if args.warmup_proportion < 1 else
args.warmup_proportion / t_total,
t_total=t_total,
schedule=args.lr_schedule,
max_grad_norm=args.max_grad_norm,
global_grad_norm=args.global_grad_norm,
init_spec=init_spec,
weight_decay_rate=args.weight_decay)
return optimizer, param_optimizer, t_total
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--device",
default=None,
type=str,
required=True,
help="The GPU device you will run on.")
parser.add_argument(
"--features_file",
default=None,
type=str,
required=True,
help=
"The train features file. Should contain the .csv files (after tokenized) for the task."
"Format: example_id,input_ids,input_mask,segment_ids,label\n")
parser.add_argument(
"--teacher_model",
default=None,
type=str,
help=
"The teacher model dir. Should contain the config/vocab/checkpoint file."
)
parser.add_argument(
"--general_student_model",
default=None,
type=str,
required=True,
help="The student model (after general distillation) dir. "
"Should contain the config/vocab/checkpoint file.")
parser.add_argument(
"--output_student_dir",
default=None,
type=str,
required=True,
help=
"The output directory for the task-specific distilled student models.")
parser.add_argument("--cache_file_dir",
default='./cache',
type=str,
required=True,
help="The directory where cache the features.")
parser.add_argument(
"--distill_model",
default='simplified',
type=str,
help="The distill model type, choose in 'standard' and 'simplified'.")
parser.add_argument(
"--max_seq_length",
default=256,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization."
)
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=64,
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('--weight_decay',
'--wd',
default=1e-2,
type=float,
metavar='W',
help='weight decay')
parser.add_argument("--num_train_epochs",
default=2,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--alpha",
default=0.5,
type=float,
help="The weight of soft loss in standard kd method."
"Only use when '--distill_model' is set as 'standard'.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="Random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--train_loss_step',
type=int,
default=1000,
help="How many train step to record a training loss. ")
parser.add_argument('--save_model_step',
type=int,
default=3000,
help="How many train step to save a student model.")
parser.add_argument('--temperature',
type=float,
default=1.,
help="The temperature in soft loss.")
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(
'--fp16_opt_level',
type=str,
default='O1',
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()
logger.info('The args: {}'.format(args))
# Prepare device
os.environ["CUDA_VISIBLE_DEVICES"] = args.device
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: {}".format(device, n_gpu))
# Prepare seed
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)
# Prepare task settings
if os.path.exists(args.output_student_dir) and os.listdir(
args.output_student_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_student_dir))
if not os.path.exists(args.output_student_dir):
os.makedirs(args.output_student_dir)
if not os.path.exists(args.cache_file_dir):
os.makedirs(args.cache_file_dir)
# For save vocab file for all output models.
tokenizer = BertTokenizer.from_pretrained(args.general_student_model,
do_lower_case=args.do_lower_case)
# Model
teacher_model = TinyBertForSequenceClassification.from_pretrained(
args.teacher_model, num_labels=2)
if args.fp16:
teacher_model.half()
teacher_model.to(device)
student_model = TinyBertForSequenceClassification.from_pretrained(
args.general_student_model, num_labels=2)
student_model.to(device)
# Train Config
num_examples, train_dataloader = distill_dataloader(
args, RandomSampler, batch_size=args.train_batch_size)
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
num_train_optimization_steps = int(
num_examples / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
logger.info("***** Running Distilling *****")
logger.info(" Num examples = %d", num_examples)
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
# Prepare optimizer
param_optimizer = list(student_model.named_parameters())
size = 0
for n, p in student_model.named_parameters():
logger.info('n: {}'.format(n))
size += p.nelement()
logger.info('Total parameters of student_model: {}'.format(size))
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':
args.weight_decay
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
schedule = 'warmup_linear'
optimizer = BertAdam(optimizer_grouped_parameters,
schedule=schedule,
lr=args.learning_rate,
warmup=args.warmup_proportion,
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 fp16 training."
)
student_model, optimizer = amp.initialize(
student_model, optimizer, opt_level=args.fp16_opt_level)
logger.info('FP16 is activated, use amp')
else:
logger.info('FP16 is not activated, only use BertAdam')
if n_gpu > 1:
student_model = torch.nn.DataParallel(student_model)
teacher_model = torch.nn.DataParallel(teacher_model)
# Prepare loss functions
loss_mse = MSELoss()
def soft_cross_entropy(predicts, targets):
student_likelihood = torch.nn.functional.log_softmax(predicts, dim=-1)
targets_prob = torch.nn.functional.softmax(targets, dim=-1)
return (-targets_prob * student_likelihood).mean()
# Train
global_step = 0
output_loss_file = os.path.join(args.output_student_dir, "train_loss.txt")
tr_loss = 0.
tr_att_loss = 0.
tr_rep_loss = 0.
tr_cls_loss = 0.
for epoch in trange(int(args.num_train_epochs), desc="Epoch"):
student_model.train()
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration", ascii=True)):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
if input_ids.size()[0] != args.train_batch_size:
continue
student_logits, student_atts, student_reps = student_model(
input_ids, segment_ids, input_mask, is_student=True)
with torch.no_grad():
teacher_logits, teacher_atts, teacher_reps = teacher_model(
input_ids, segment_ids, input_mask)
soft_loss = soft_cross_entropy(student_logits / args.temperature,
teacher_logits / args.temperature)
hard_loss = torch.nn.functional.cross_entropy(student_logits,
label_ids,
reduction='mean')
if args.distill_model == 'standard':
cls_loss = args.alpha * soft_loss + (1 -
args.alpha) * hard_loss
tr_cls_loss += cls_loss.item()
loss = cls_loss
elif args.distill_model == 'simplified':
teacher_layer_num = len(teacher_atts)
student_layer_num = len(student_atts)
assert teacher_layer_num % student_layer_num == 0
layers_per_block = int(teacher_layer_num / student_layer_num)
new_teacher_atts = [
teacher_atts[i * layers_per_block + layers_per_block - 1]
for i in range(student_layer_num)
]
att_loss = 0.
rep_loss = 0.
# attention loss
for student_att, teacher_att in zip(student_atts,
new_teacher_atts):
student_att = torch.where(
student_att <= -1e2,
torch.zeros_like(student_att).to(device), student_att)
teacher_att = torch.where(
teacher_att <= -1e2,
torch.zeros_like(teacher_att).to(device), teacher_att)
tmp_loss = loss_mse(student_att, teacher_att)
att_loss += tmp_loss
# hidden states loss
new_teacher_reps = [
teacher_reps[i * layers_per_block]
for i in range(student_layer_num + 1)
]
new_student_reps = student_reps
for student_rep, teacher_rep in zip(new_student_reps,
new_teacher_reps):
tmp_loss = loss_mse(student_rep, teacher_rep)
rep_loss += tmp_loss
tr_att_loss += att_loss.item()
tr_rep_loss += rep_loss.item()
# classification loss
cls_loss = soft_loss + hard_loss
tr_cls_loss += cls_loss.item()
# total loss
loss = rep_loss + att_loss + cls_loss
else:
raise NotImplementedError
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
tr_loss += loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
if global_step % args.train_loss_step == 0:
loss = tr_loss / args.train_loss_step
cls_loss = tr_cls_loss / args.train_loss_step
att_loss = tr_att_loss / args.train_loss_step
rep_loss = tr_rep_loss / args.train_loss_step
loss_dict = {}
loss_dict['global_step'] = global_step
loss_dict['cls_loss'] = cls_loss
loss_dict['att_loss'] = att_loss
loss_dict['rep_loss'] = rep_loss
loss_dict['loss'] = loss
write_loss_to_file(loss_dict, output_loss_file)
tr_loss = 0.
tr_att_loss = 0.
tr_rep_loss = 0.
tr_cls_loss = 0.
if global_step % args.save_model_step == 0:
logger.info("***** Save model *****")
model_to_save = student_model.module if hasattr(
student_model, 'module') else student_model
model_name = WEIGHTS_NAME
checkpoint_name = 'checkpoint-' + str(global_step)
output_model_dir = os.path.join(args.output_dir,
checkpoint_name)
if not os.path.exists(output_model_dir):
os.makedirs(output_model_dir)
output_model_file = os.path.join(output_model_dir, model_name)
output_config_file = os.path.join(output_model_dir,
CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(output_model_dir)
if os.path.exists(args.cache_file_dir):
import shutil
shutil.rmtree(args.cache_file_dir)
