Why trainer.logged_metrics is mixed of tensor and non tensor value.

[tshu-w]

When I print trainer.logged_metrics, I got

{
    'train_f1': tensor(1.),
    'train_prc': tensor(1.),
    'train_rec': tensor(1.),
    'train_loss': tensor(0.0152),
    'valid_f1': 0.861257791519165,
    'valid_prc': 0.9134419560432434,
    'valid_rec': 0.8147138953208923,
    'valid_loss': 0.16973154246807098
}

Why trainer.logged_metrics is mixed of tensor and non tensor value.

[rohitgr7]

would you mind sharing the log calls in your training_step and validation_step hooks?

[tshu-w]

Sure, it look like this.

def __init__(self) -> None:
    ...

    metrics_kwargs = {"ignore_index": 0}
    metrics = MetricCollection(
        {
            "f1": F1(**metrics_kwargs),
            "prc": Precision(**metrics_kwargs),
            "rec": Recall(**metrics_kwargs),
        }
    )
    self.train_metrics = metrics.clone(prefix="train_")
    self.valid_metrics = metrics.clone(prefix="valid_")
    self.test_metrics = metrics.clone(prefix="test_")

def common_step(self, batch, step: str) -> Optional[STEP_OUTPUT]:
    outputs = self.forward(batch)
    loss, logits = outputs.loss, outputs.logits
    probs = F.softmax(logits, dim=-1)

    metrics = getattr(self, f"{step}_metrics")
    metrics(probs, batch["labels"])

    self.log_dict(metrics, prog_bar=True)
    self.log(f"{step}_loss", loss, prog_bar=True)

    return loss

def training_step(self, batch, batch_idx: int) -> STEP_OUTPUT:
    return self.common_step(batch, "train")

def validation_step(self, batch, batch_idx: int) -> Optional[STEP_OUTPUT]:
    return self.common_step(batch, "valid")

[rohitgr7]

thanks!

opened an improvement here: #11270