训练LORA情感分析模型时无日志且报'eval_loss'键错误求助

我正在学术项目中构建用于情感分析的LORA模型，但训练时计算损失持续报错，尝试多种损失计算方法均无效。相关代码、报错信息如下：

训练与指标计算代码 #

training_args = TrainingArguments(
    output_dir='C:/Users/felix/Downloads',
    evaluation_strategy='epoch',
    save_strategy='epoch',
    learning_rate=.2,
    per_device_train_batch_size=18,
    per_device_eval_batch_size=18,
    num_train_epochs=1,
    load_best_model_at_end=True,
    weight_decay=0.1,
    remove_unused_columns=False,
    #label_names="labels"
)

def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    predictions = torch.from_numpy(predictions)  # Convert predictions to tensor
    labels = torch.from_numpy(labels).long()  # Convert labels to tensor
    loss = nn.CrossEntropyLoss()(predictions, labels)  # Calculate the evaluation loss
    accuracy = (torch.argmax(predictions, axis=1) == labels).float().mean()  # Calculate the accuracy

    # Print the metrics dictionary for debugging
    metrics = {"eval_loss": loss.item(), "accuracy": accuracy.item()}
    print("Metrics:", metrics)

    return metrics
trainer = Trainer(
    model=lora_model,
    args = training_args,
    train_dataset = random_train_samples,
    eval_dataset = new_dataset['test'],
    tokenizer=tokenizer,
    data_collator = DataCollatorWithPadding(tokenizer=tokenizer),
    compute_metrics = compute_metrics
)
trainer.train()

报错信息 #

KeyError                                  Traceback (most recent call last)
Cell In[45], line 10
      1 trainer = Trainer(
      2     model=lora_model,
      3     args = training_args,
   (...)
      8     compute_metrics = compute_metrics
      9 )
---> 10 trainer.train()

File ~\anaconda3\Lib\site-packages\transformers\trainer.py:1555, in Trainer.train(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)
   1553         hf_hub_utils.enable_progress_bars()
   1554 else:
-> 1555     return inner_training_loop(
   1556         args=args,
   1557         resume_from_checkpoint=resume_from_checkpoint,
   1558         trial=trial,
   1559         ignore_keys_for_eval=ignore_keys_for_eval,
   1560     )

File ~\anaconda3\Lib\site-packages\transformers\trainer.py:1944, in Trainer._inner_training_loop(self, batch_size, args, resume_from_checkpoint, trial, ignore_keys_for_eval)
   1941     self.control.should_training_stop = True
   1943 self.control = self.callback_handler.on_epoch_end(args, self.state, self.control)
-> 1944 self._maybe_log_save_evaluate(tr_loss, model, trial, epoch, ignore_keys_for_eval)
   1946 if DebugOption.TPU_METRICS_DEBUG in self.args.debug:
   1947     if is_torch_tpu_available():
   1948         # tpu-comment: Logging debug metrics for PyTorch/XLA (compile, execute times, ops, etc.)

File ~\anaconda3\Lib\site-packages\transformers\trainer.py:2267, in Trainer._maybe_log_save_evaluate(self, tr_loss, model, trial, epoch, ignore_keys_for_eval)
   2264         self.lr_scheduler.step(metrics[metric_to_check])
   2266 if self.control.should_save:
-> 2267     self._save_checkpoint(model, trial, metrics=metrics)
   2268     self.control = self.callback_handler.on_save(self.args, self.state, self.control)

File ~\anaconda3\Lib\site-packages\transformers\trainer.py:2383, in Trainer._save_checkpoint(self, model, trial, metrics)
   2381 if not metric_to_check.startswith("eval_"):
   2382     metric_to_check = f"eval_{metric_to_check}"
-> 2383 metric_value = metrics[metric_to_check]
   2385 operator = np.greater if self.args.greater_is_better else np.less
   2386 if (
   2387     self.state.best_metric is None
   2388     or self.state.best_model_checkpoint is None
   2389     or operator(metric_value, self.state.best_metric)
   2390 ):

KeyError: 'eval_loss'

完整代码可查看GitHub仓库：therrief87/udacity 仓库下的 v4 Udacity Lightweight Fine Tuning (GPT2) Smaller Training Data (1).ipynb 文件

恳请各位提供帮助。

内容的提问来源于stack exchange，提问作者Therrief