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

33
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
@@ -503,7 +497,7 @@ 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)
@@ -521,13 +515,15 @@ 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:
print("Loading Train Dataset", args.train_file)
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)
num_train_optimization_steps = int(
len(train_dataset) / 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 = BertForPreTraining.from_pretrained(args.bert_model)
@@ -550,6 +546,7 @@ def main():
{'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
@@ -570,14 +567,14 @@ def main():
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
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)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
@@ -607,10 +604,12 @@ 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_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