HuggingFace Trainer无法向wandb上报问题求助

我按照文档配置Trainer的report_to参数为wandb，代码如下：

training_args = TrainingArguments(
    output_dir="test_trainer",
    evaluation_strategy="steps",
    learning_rate=config.learning_rate,
    num_train_epochs=config.epochs,
    weight_decay=config.weight_decay,
    logging_dir=config.logging_dir,
    report_to="wandb",
    save_total_limit=1,
    per_device_train_batch_size=config.batch_size,
    per_device_eval_batch_size=config.batch_size,
    fp16=True,
    load_best_model_at_end=True,
    seed=42
)

初始化Trainer时：

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    compute_metrics=compute_metrics
)

出现如下报错：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-68-b009351ab52d> in <module>
      4     train_dataset=train_dataset,
      5     eval_dataset=eval_dataset,
----> 6     compute_metrics=compute_metrics
      7 )

~/.virtualenvs/transformers_lab/lib/python3.7/site-packages/transformers/trainer.py in __init__(self, model, args, data_collator, train_dataset, eval_dataset, tokenizer, model_init, compute_metrics, callbacks, optimizers)
    286                 "You should subclass `Trainer` and override the `create_optimizer_and_scheduler` method."
    287             )
--> 288         default_callbacks = DEFAULT_CALLBACKS + get_reporting_integration_callbacks(self.args.report_to)
    289         callbacks = default_callbacks if callbacks is None else default_callbacks + callbacks
    290         self.callback_handler = CallbackHandler(

~/.virtualenvs/transformers_lab/lib/python3.7/site-packages/transformers/integrations.py in get_reporting_integration_callbacks(report_to)
    794         if integration not in INTEGRATION_TO_CALLBACK:
    795             raise ValueError(
--> 796                 f"{integration} is not supported, only {', '.join(INTEGRATION_TO_CALLBACK.keys())} are supported."
    797             )
    798     return [INTEGRATION_TO_CALLBACK[integration] for integration in report_to]

ValueError: w is not supported, only azure_ml, comet_ml, mlflow, tensorboard, wandb are supported.

请问有没有人遇到过相同的错误？

解决方法 #

• 错误原因：旧版本transformers库中，report_to参数被设计为接受列表类型，传入字符串时会被当成可迭代对象逐个遍历字符，导致把"wandb"拆成了"w"、"a"等单个字符，触发不支持的报错。
• 修复步骤：将report_to的值改为列表格式：

  training_args = TrainingArguments(
      # 其他参数保持不变
      report_to=["wandb"],
      # 其他参数保持不变
  )
  

• 验证：修改后重新初始化Trainer，即可正常加载wandb集成，不会再出现该错误。

内容的提问来源于stack exchange，提问作者Weber Huang