Struct SequenceOutput 

pub struct SequenceOutput {
    pub loss: Tensor<1>,
    pub logits: Tensor<3>,
    pub predictions: Option<Tensor<2, Int>>,
    pub targets: Tensor<2, Int>,
}

Sequence prediction output adapted for multiple metrics.

Supported metrics:

• Accuracy
• TopKAccuracy
• Perplexity
• Loss
• CER
• WER

Fields

loss: Tensor<1>

The loss.

logits: Tensor<3>

Raw logits. Shape: [batch_size, seq_len, vocab_size]

predictions: Option<Tensor<2, Int>>

Optional predicted token indices. Shape: [batch_size, seq_length]. If not provided, predictions default to argmax of logits along the last dimension.

targets: Tensor<2, Int>

The target token indices. Shape: [batch_size, seq_length]

Implementations

impl SequenceOutput

pub fn new( loss: Tensor<1>, logits: Tensor<3>, predictions: Option<Tensor<2, Int>>, targets: Tensor<2, Int>, ) -> SequenceOutput

Constructs a new SequenceOutput.

Trait Implementations

impl Adaptor<AccuracyInput> for SequenceOutput

fn adapt(&self) -> AccuracyInput

Adapt the type to be passed to a metric.

impl Adaptor<CerInput> for SequenceOutput

fn adapt(&self) -> CerInput

Adapt the type to be passed to a metric.

impl Adaptor<LossInput> for SequenceOutput

fn adapt(&self) -> LossInput

Adapt the type to be passed to a metric.

impl Adaptor<PerplexityInput> for SequenceOutput

fn adapt(&self) -> PerplexityInput

Adapt the type to be passed to a metric.

impl Adaptor<TopKAccuracyInput> for SequenceOutput

fn adapt(&self) -> TopKAccuracyInput

Adapt the type to be passed to a metric.

impl Adaptor<WerInput> for SequenceOutput

fn adapt(&self) -> WerInput

Adapt the type to be passed to a metric.

impl ItemLazy for SequenceOutput

fn sync(self) -> Result<SequenceOutput, ExecutionError>

Sync the item.