SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Merge pull request #218 from matej-svejda/master

Browse Source 
Fix learning rate problems in run_classifier.py
This commit is contained in:
Thomas Wolf
2019-02-05 15:40:44 +01:00
committed by GitHub
parent 8f8bbd4a4c 1579c53635
commit e9e77cd3c4
6 changed files with 80 additions and 94 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
examples/run_squad.py
View File

@@ -36,7 +36,7 @@ from torch.utils.data.distributed import DistributedSampler
from pytorch_pretrained_bert.tokenization import whitespace_tokenize, BasicTokenizer, BertTokenizer
from pytorch_pretrained_bert.modeling import BertForQuestionAnswering
from pytorch_pretrained_bert.optimization import BertAdam
from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear
from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
@@ -670,11 +670,6 @@ def _compute_softmax(scores):
probs.append(score / total_sum)
return probs
def warmup_linear(x, warmup=0.002):
if x < warmup:
return x/warmup
return 1.0 - x
def main():
parser = argparse.ArgumentParser()
@@ -762,7 +757,7 @@ def main():
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
@@ -789,12 +784,14 @@ def main():
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_squad_examples(
input_file=args.train_file, is_training=True)
num_train_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare model
model = BertForQuestionAnswering.from_pretrained(args.bert_model,
@@ -826,9 +823,6 @@ def main():
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
@@ -848,7 +842,7 @@ def main():
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
@@ -874,7 +868,7 @@ def main():
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
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)
@@ -905,10 +899,12 @@ def main():
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
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