onnx_diagnostic.tasks.automatic_speech_recognition

onnx_diagnostic.tasks.automatic_speech_recognition.get_inputs(model: Module, config: Any | None, dummy_max_token_id: int, max_source_positions: int, d_model: int, num_hidden_layers: int, encoder_attention_heads: int, encoder_layers: int, decoder_layers: int, head_dim: int, batch_size: int = 2, sequence_length: int = 30, **kwargs)

Generates inputs for task text2text-generation. Example:

dict(
    cache_position:T7s4,
    past_key_values:EncoderDecoderCache(
        self_attention_cache=DynamicCache[serialized](#2[#0[],#0[]]),
        cross_attention_cache=DynamicCache[serialized](#2[#0[],#0[]])
    ),
    decoder_input_ids:T7s1x4,
    encoder_outputs:BaseModelOutput(last_hidden_state:T1s1x1500x384),
    use_cache:bool,return_dict:bool
)
dict(
    cache_position:T7s1,
    past_key_values:EncoderDecoderCache(
        self_attention_cache=DynamicCache[serialized](#2[
            #4[T1s1x6x4x64,T1s1x6x4x64,T1s1x6x4x64,T1s1x6x4x64],
            #4[T1s1x6x4x64,T1s1x6x4x64,T1s1x6x4x64,T1s1x6x4x64]
        ]),
        cross_attention_cache=DynamicCache[serialized](#2[
            #4[T1s1x6x1500x64,T1s1x6x1500x64,T1s1x6x1500x64,T1s1x6x1500x64],
            #4[T1s1x6x1500x64,T1s1x6x1500x64,T1s1x6x1500x64,T1s1x6x1500x64]
        ]),
    ),
    decoder_input_ids:T7s1x1,
    encoder_outputs:BaseModelOutput(last_hidden_state:T1s1x1500x384),
    use_cache:bool,return_dict:bool
)

onnx_diagnostic.tasks.automatic_speech_recognition.random_input_kwargs(config: Any, task: str) → Tuple[Dict[str, Any], Callable]

Inputs kwargs.

If the configuration is None, the function selects typical dimensions.

onnx_diagnostic.tasks.automatic_speech_recognition.reduce_model_config(config: Any, task: str) → Dict[str, Any]

Reduces a model size.