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

20
examples/run_lm_finetuning.py
View File

@@ -31,7 +31,7 @@ from torch.utils.data.distributed import DistributedSampler
from pytorch_pretrained_bert.tokenization import BertTokenizer
from pytorch_pretrained_bert.modeling import BertForPreTraining
from pytorch_pretrained_bert.optimization import BertAdam
from pytorch_pretrained_bert.optimization import BertAdam, warmup_linear
from torch.utils.data import Dataset
import random
@@ -42,12 +42,6 @@ logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message
logger = logging.getLogger(__name__)
def warmup_linear(x, warmup=0.002):
if x < warmup:
return x/warmup
return 1.0 - x
class BERTDataset(Dataset):
def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8", corpus_lines=None, on_memory=True):
self.vocab = tokenizer.vocab
@@ -527,7 +521,7 @@ def main():
train_dataset = BERTDataset(args.train_file, tokenizer, seq_len=args.max_seq_length,
corpus_lines=None, on_memory=args.on_memory)
num_train_steps = int(
len(train_dataset) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
len(train_dataset) / args.train_batch_size * args.num_train_epochs)
# Prepare model
model = BertForPreTraining.from_pretrained(args.bert_model)
@@ -607,10 +601,12 @@ def main():
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_steps, 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_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