[Bug]: float division by zero when training #5679
Description
软件环境
paddle-bfloat 0.1.7
paddle2onnx 1.0.0
paddlefsl 1.1.0
paddlehub 2.3.0
paddlenlp 2.4.2
paddlepaddle-gpu 2.4.0.post112
tb-paddle 0.3.6重复问题
- I have searched the existing issues
错误描述
我在使用paddlenlp的时候遇到了如下问题
---------------------------------------------------------------------------
ZeroDivisionError Traceback (most recent call last)
/tmp/ipykernel_595/4032920361.py in <module>
----> 1 trainer.train()
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlenlp/trainer/trainer_base.py in train(self, resume_from_checkpoint, ignore_keys_for_eval)
650
651 self._total_loss_scalar += tr_loss.item()
--> 652 train_loss = self._total_loss_scalar / self.state.global_step
653
654 metrics = speed_metrics("train",
ZeroDivisionError: float division by zero```
我的一些参数是这样的args = TrainingArguments(
output_dir='output',
do_train=True,
do_eval=True
)
trainer = Trainer(
model=model,
criterion=paddle.nn.CrossEntropyLoss(),
data_collator=collate_fn,
train_dataset=train_set,
eval_dataset=None,
tokenizer=tokenzier,
args=args
)
trainer.train()
稳定复现步骤 & 代码
这个没法给，比赛的项目代码不好给吧