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

fix learning rate/fp16 and warmup problem for all examples

Browse Source
This commit is contained in:
Matej Svejda
2019-01-27 14:07:24 +01:00
parent 01ff4f82ba
commit 9c6a48c8c3
5 changed files with 39 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
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()
@@ -794,7 +789,7 @@ def main():
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)
len(train_examples) / args.train_batch_size * args.num_train_epochs)
# Prepare model
model = BertForQuestionAnswering.from_pretrained(args.bert_model,
@@ -905,10 +900,12 @@ 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/t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1