啥时候支持batch推理

[chamwen]

if kwargs.get("batch_size", 1) > 1:
raise NotImplementedError("batch decoding is not implemented")

[AdilAdam]

if kwargs.get("batch_size", 1) > 1: raise NotImplementedError("batch decoding is not implemented")

def inference_prepare(
self,
data_in,
data_lengths=None,
key: list = None,
tokenizer=None,
frontend=None,
**kwargs,
):
meta_data = {}

    if kwargs.get("batch_size", 1) > 1:
        raise NotImplementedError("batch decoding is not implemented")

    batch_file_input = []

    for data_in_ in data_in:
        contents = self.data_template(data_in_)
        output = self.data_load_speech(
            contents, tokenizer, frontend, meta_data=meta_data, **kwargs
        )

        batch_file_input.append(output)

    padded_batch = {}
    for key in output:
        if key == "wav_utt":
            padded_batch[key] = [s[key] for s in batch_file_input]
        else:
            if key == "speech":
                padded_batch[key] = pad_sequence(
                    [s[key].transpose(0, 1) for s in batch_file_input],
                    batch_first=True,
                    padding_value=0,
                ).transpose(1, 2)
            else:
                padded_batch[key] = pad_sequence(
                    [s[key].squeeze(0) for s in batch_file_input],
                    batch_first=True,
                    padding_value=0,
                )

    output = padded_batch

    batch = to_device(output, kwargs["device"])

you can rewrite it in model.py

[chamwen]

if kwargs.get("batch_size", 1) > 1: raise NotImplementedError("batch decoding is not implemented")

def inference_prepare( self, data_in, data_lengths=None, key: list = None, tokenizer=None, frontend=None, **kwargs, ): meta_data = {}

    if kwargs.get("batch_size", 1) > 1:
        raise NotImplementedError("batch decoding is not implemented")

    batch_file_input = []

    for data_in_ in data_in:
        contents = self.data_template(data_in_)
        output = self.data_load_speech(
            contents, tokenizer, frontend, meta_data=meta_data, **kwargs
        )

        batch_file_input.append(output)

    padded_batch = {}
    for key in output:
        if key == "wav_utt":
            padded_batch[key] = [s[key] for s in batch_file_input]
        else:
            if key == "speech":
                padded_batch[key] = pad_sequence(
                    [s[key].transpose(0, 1) for s in batch_file_input],
                    batch_first=True,
                    padding_value=0,
                ).transpose(1, 2)
            else:
                padded_batch[key] = pad_sequence(
                    [s[key].squeeze(0) for s in batch_file_input],
                    batch_first=True,
                    padding_value=0,
                )

    output = padded_batch

    batch = to_device(output, kwargs["device"])

you can rewrite it in model.py

thx, how you define pad_sequence()?

[Shenkailai]

if kwargs.get("batch_size", 1) > 1: raise NotImplementedError("batch decoding is not implemented")

def inference_prepare( self, data_in, data_lengths=None, key: list = None, tokenizer=None, frontend=None, **kwargs, ): meta_data = {}

    if kwargs.get("batch_size", 1) > 1:
        raise NotImplementedError("batch decoding is not implemented")

    batch_file_input = []

    for data_in_ in data_in:
        contents = self.data_template(data_in_)
        output = self.data_load_speech(
            contents, tokenizer, frontend, meta_data=meta_data, **kwargs
        )

        batch_file_input.append(output)

    padded_batch = {}
    for key in output:
        if key == "wav_utt":
            padded_batch[key] = [s[key] for s in batch_file_input]
        else:
            if key == "speech":
                padded_batch[key] = pad_sequence(
                    [s[key].transpose(0, 1) for s in batch_file_input],
                    batch_first=True,
                    padding_value=0,
                ).transpose(1, 2)
            else:
                padded_batch[key] = pad_sequence(
                    [s[key].squeeze(0) for s in batch_file_input],
                    batch_first=True,
                    padding_value=0,
                )

    output = padded_batch

    batch = to_device(output, kwargs["device"])

you can rewrite it in model.py

thx, how you define pad_sequence()?

from torch.nn.utils.rnn import pad_sequence