espnet.utils.io_utils.LoadInputsAndTargets

source

class espnet.utils.io_utils.LoadInputsAndTargets(mode='asr', preprocess_conf=None, load_input=True, load_output=True, sort_in_input_length=True, use_speaker_embedding=False, use_second_target=False, preprocess_args=None, keep_all_data_on_mem=False)

Bases: object

Create a mini-batch from a list of dicts.

>>> batch = [('utt1',
...           dict(input=[dict(feat='some.ark:123',
...                            filetype='mat',
...                            name='input1',
...                            shape=[100, 80])],
...                output=[dict(tokenid='1 2 3 4',
...                             name='target1',
...                             shape=[4, 31])]]))
>>> l = LoadInputsAndTargets()
>>> feat, target = l(batch)

• Param: str mode: Specify the task mode, “asr” or “tts”
• Param: str preprocess_conf: The path of a json file for pre-processing
• Param: bool load_input: If False, not to load the input data
• Param: bool load_output: If False, not to load the output data
• Param: bool sort_in_input_length: Sort the mini-batch in descending order of the input length
• Param: bool use_speaker_embedding: Used for tts mode only
• Param: bool use_second_target: Used for tts mode only
• Param: dict preprocess_args: Set some optional arguments for preprocessing
• Param: Optional[dict] preprocess_args: Used for tts mode only

Initialize Load inputs and Targets.