Struct BartModelOutput

pub struct BartModelOutput {
    pub decoder_output: Tensor,
    pub encoder_hidden_state: Option<Tensor>,
    pub cache: Option<Vec<(Option<LayerState>, Option<LayerState>)>>,
    pub all_decoder_hidden_states: Option<Vec<Tensor>>,
    pub all_decoder_attentions: Option<Vec<Tensor>>,
    pub all_encoder_hidden_states: Option<Vec<Tensor>>,
    pub all_encoder_attentions: Option<Vec<Tensor>>,
}

Container holding a BART model output. The decoder output may hold the hidden state of the last layer of the decoder, or may hold logits for a custom head module after the decoder (e.g. for classification or language modeling tasks)

Fields

decoder_output: Tensor

Hidden state of the last layer of the decoder, or logits for a custom head module after the decoder (e.g. for classification or language modeling tasks)

encoder_hidden_state: Option<Tensor>

Hidden state for the last layer of the encoder if they are calculated (not provided), otherwise None

cache: Option<Vec<(Option<LayerState>, Option<LayerState>)>>

Cached outputs of the model (attention layers keys and values) if the model is used for generation

all_decoder_hidden_states: Option<Vec<Tensor>>

Hidden states for all layers of the decoder

all_decoder_attentions: Option<Vec<Tensor>>

Attention weights for all layers of the decoder

all_encoder_hidden_states: Option<Vec<Tensor>>

Hidden states for all layers of the encoder

all_encoder_attentions: Option<Vec<Tensor>>

Attention weights for all layers of the encoder