'pooler_output' 的 Electra 序列分类与 pytorch 闪电问题

Electra sequence classification with pytorch lightning issues with 'pooler_output'

我正在处理一个句子分类任务,每个句子都有多个二进制标签。我正在使用 Electra 和 pytorch lightning 来完成这项工作,但我 运行 遇到了问题。当我 运行 宁 trainer.fit(model, data) 我得到以下错误:

AttributeError: 'tuple' object has no attribute 'pooler_output'

错误指的是我定义的部分中的第 13 行 pl.LightningModule:

class CrowdCodedTagger(pl.LightningModule):

  def __init__(self, n_classes: int, n_training_steps=None, n_warmup_steps=None):
    super().__init__()
    self.electra = ElectraModel.from_pretrained(ELECTRA_MODEL_NAME, return_dict=False) #changed ElectraModel to ElectraForSequenceClassification
    self.classifier = nn.Linear(self.electra.config.hidden_size, n_classes)
    self.n_training_steps = n_training_steps
    self.n_warmup_steps = n_warmup_steps
    self.criterion = nn.BCELoss()

  def forward(self, input_ids, attention_mask, labels=None):
    output = self.electra(input_ids, attention_mask=attention_mask)
    output = self.classifier(output.pooler_output) # <---- this is the line the error is referring to.
    output = torch.sigmoid(output)
    loss = 0
    if labels is not None:
        loss = self.criterion(output, labels)
    return loss, output

  def training_step(self, batch, batch_idx):
    input_ids = batch["input_ids"]
    attention_mask = batch["attention_mask"]
    labels = batch["labels"]
    loss, outputs = self(input_ids, attention_mask, labels)
    self.log("train_loss", loss, prog_bar=True, logger=True)
    return {"loss": loss, "predictions": outputs, "labels": labels}

  def validation_step(self, batch, batch_idx):
    input_ids = batch["input_ids"]
    attention_mask = batch["attention_mask"]
    labels = batch["labels"]
    loss, outputs = self(input_ids, attention_mask, labels)
    self.log("val_loss", loss, prog_bar=True, logger=True)
    return loss

  def test_step(self, batch, batch_idx):
    input_ids = batch["input_ids"]
    attention_mask = batch["attention_mask"]
    labels = batch["labels"]
    loss, outputs = self(input_ids, attention_mask, labels)
    self.log("test_loss", loss, prog_bar=True, logger=True)
    return loss

  def training_epoch_end(self, outputs):
    
    labels = []
    predictions = []
    for output in outputs:
      for out_labels in output["labels"].detach().cpu():
        labels.append(out_labels)
      for out_predictions in output["predictions"].detach().cpu():
        predictions.append(out_predictions)

    labels = torch.stack(labels).int()
    predictions = torch.stack(predictions)

    for i, name in enumerate(LABEL_COLUMNS):
      class_roc_auc = auroc(predictions[:, i], labels[:, i])
      self.logger.experiment.add_scalar(f"{name}_roc_auc/Train", class_roc_auc, self.current_epoch)

  def configure_optimizers(self):

    optimizer = AdamW(self.parameters(), lr=2e-5)

    scheduler = get_linear_schedule_with_warmup(
      optimizer,
      num_warmup_steps=self.n_warmup_steps,
      num_training_steps=self.n_training_steps
    )

    return dict(
      optimizer=optimizer,
      lr_scheduler=dict(
        scheduler=scheduler,
        interval='step'
      )
    )

任何人都可以指出我修复错误的方向吗?

数据结构示例(CSV 格式):

sentence                      label_1          label_2          label_3
Lorem ipsum dolor sit amet    1                0                1
consectetur adipiscing elit   0                0                0
sed do eiusmod tempor         0                1                1
incididunt ut labore et       1                0                0
Lorem ipsum dolor sit amet    1                0                1

ELECTRA has no pooler layer like BERT(比较 return 部分以获得更多信息)。

如果您只想使用 [CLS] 标记进行序列分类,您可以简单地使用 last_hidden_state 的第一个元素(在没有 return_dict=False 的情况下初始化 electra):

output = self.classifier(output.last_hidden_state[:, 0])