[Bug]: 计算AccuracyAndF1出错

[xiehuanyi]

软件环境

paddle-bfloat                  0.1.7
paddle2onnx                    1.0.0
paddlefsl                      1.1.0
paddlehub                      2.3.0
paddlenlp                      2.5.2
paddlepaddle-gpu               2.3.2.post112
tb-paddle                      0.3.6

重复问题

• I have searched the existing issues

错误描述

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/tmp/ipykernel_3067/2457985955.py in <module>
     89     compute_metrics=paddlenlp.metrics.AccuracyAndF1,
     90 )
---> 91 trainer.train()

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlenlp/trainer/trainer.py in train(self, resume_from_checkpoint, ignore_keys_for_eval)
    714 
    715                     self.control = self.callback_handler.on_step_end(args, self.state, self.control)
--> 716                     self._maybe_log_save_evaluate(tr_loss, model, epoch, ignore_keys_for_eval, inputs=inputs)
    717                 else:
    718                     self.control = self.callback_handler.on_substep_end(args, self.state, self.control)

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlenlp/trainer/trainer.py in _maybe_log_save_evaluate(self, tr_loss, model, epoch, ignore_keys_for_eval, **kwargs)
    846                     )
    847             else:
--> 848                 metrics = self.evaluate(ignore_keys=ignore_keys_for_eval)
    849 
    850         if self.control.should_save:

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlenlp/trainer/trainer.py in evaluate(self, eval_dataset, ignore_keys, metric_key_prefix)
   1614             prediction_loss_only=True if self.compute_metrics is None else None,
   1615             ignore_keys=ignore_keys,
-> 1616             metric_key_prefix=metric_key_prefix,
   1617         )
   1618 

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlenlp/trainer/trainer.py in evaluation_loop(self, dataloader, description, prediction_loss_only, ignore_keys, metric_key_prefix, max_eval_iters)
   1787         # Metrics!
   1788         if self.compute_metrics is not None and all_preds is not None and all_labels is not None:
-> 1789             metrics = self.compute_metrics(EvalPrediction(predictions=all_preds, label_ids=all_labels))
   1790         else:
   1791             metrics = {}

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddlenlp/metrics/glue.py in __init__(self, topk, pos_label, name, *args, **kwargs)
     64         self.pos_label = pos_label
     65         self._name = name
---> 66         self.acc = Accuracy(self.topk, *args, **kwargs)
     67         self.precision = Precision(*args, **kwargs)
     68         self.recall = Recall(*args, **kwargs)

/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/metric/metrics.py in __init__(self, topk, name, *args, **kwargs)
    238         super(Accuracy, self).__init__(*args, **kwargs)
    239         self.topk = topk
--> 240         self.maxk = max(topk)
    241         self._init_name(name)
    242         self.reset()

ValueError: operands could not be broadcast together with shapes (1000,) (1000,2) 

### 稳定复现步骤 & 代码

这是一个二分类的任务，我的数据读取是这样的：

def collate_fn(data):
feats = [d[0] for d in data]
labels = paddle.to_tensor([d[1] for d in data], dtype='int64')
encodings = tokenizer.encode_batch(feats)
input_ids = paddle.to_tensor([enc.ids for enc in encodings], dtype='int64')
attn_mask = paddle.to_tensor([enc.attention_mask for enc in encodings], dtype='int64')
return {
"input_ids": input_ids,
"attention_mask": attn_mask,
"labels": labels
}

我使用了Trainer，相关的设置：

args = paddlenlp.trainer.TrainingArguments(
output_dir='models',
do_train=True,
do_eval=True,
per_device_train_batch_size=8,
gradient_accumulation_steps=32,
evaluation_strategy='steps',
per_device_eval_batch_size=16,
eval_steps=10,
save_total_limit=1,
report_to='visualdl',
logging_steps=50,
)

trainer = paddlenlp.trainer.Trainer(
model=model,
criterion=paddle.nn.CrossEntropyLoss(),
args=args,
data_collator=collate_fn,
train_dataset=MyDataset('train_data/0.txt'),
eval_dataset=MyDataset('train_data/1.txt', 'eval'),
optimizers=[opt, paddle.optimizer.lr.NoamDecay(512, 25000, 5)],
compute_metrics=paddlenlp.metrics.AccuracyAndF1,
)
trainer.train()

[w5688414]

这个格式有点乱，请提供一下数据和代码，方便我们快速定位。