/// Multi-label classification output adapted for multiple metrics.
///
/// Supported metrics:
/// - HammingScore
/// - Precision (via ConfusionStatsInput)
/// - Recall (via ConfusionStatsInput)
/// - FBetaScore (via ConfusionStatsInput)
/// - Loss
#[derive(new)]
pub struct MultiLabelSoftClassificationOutput<B: Backend> {
    /// The loss.
    pub loss: Tensor<B, 1>,

    /// The label logits or probabilities. Shape: \[batch_size, num_classes\].
    pub output: Tensor<B, 2>,

    /// The ground truth labels (target values). Shape: \[batch_size, num_classes\].
    pub targets: Tensor<B, 2>,
}

impl<B: Backend> ItemLazy for MultiLabelSoftClassificationOutput<B> {
    type ItemSync = MultiLabelSoftClassificationOutput<NdArray>;

    fn sync(self) -> Self::ItemSync {
        let [output, loss, targets] = Transaction::default()
            .register(self.output)
            .register(self.loss)
            .register(self.targets)
            .execute()
            .try_into()
            .expect("Correct amount of tensor data");

        let device = &Default::default();

        MultiLabelSoftClassificationOutput {
            output: Tensor::from_data(output, device),
            loss: Tensor::from_data(loss, device),
            targets: Tensor::from_data(targets, device),
        }
    }
}

impl<B: Backend> Adaptor<HammingScoreInput<B>> for MultiLabelSoftClassificationOutput<B> {
    fn adapt(&self) -> HammingScoreInput<B> {
        HammingScoreInput::new(self.output.clone(), self.targets.clone())
    }
}

impl<B: Backend> Adaptor<LossInput<B>> for MultiLabelSoftClassificationOutput<B> {
    fn adapt(&self) -> LossInput<B> {
        LossInput::new(self.loss.clone())
    }
}

impl<B: Backend> Adaptor<ConfusionStatsInput<B>> for MultiLabelSoftClassificationOutput<B> {
    fn adapt(&self) -> ConfusionStatsInput<B> {
        ConfusionStatsInput::new(self.output.clone(), self.targets.clone().bool())
    }
}