Merge branch 'main' of https://github.com/alibaba-damo-academy/FunASR into ...

雾聪

2023-05-17 8706e767affc6bdc8cb7a67ca3a20a62779ff048

 funasr/bin/asr_inference_launch.py

old mode 100755
new mode 100644
@@ -1,11 +1,10 @@
#!/usr/bin/env python3
# Copyright ESPnet (https://github.com/espnet/espnet). All Rights Reserved.
#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)

import argparse
import logging
import os
import sys
from typing import Union, Dict, Any

from funasr.utils import config_argparse
from funasr.utils.cli_utils import get_commandline_args
@@ -72,8 +71,29 @@
    )
    group.add_argument("--key_file", type=str_or_none)
    group.add_argument("--allow_variable_data_keys", type=str2bool, default=False)

    group.add_argument(
            "--mc",
            type=bool,
            default=False,
            help="MultiChannel input",
        )
        
    group = parser.add_argument_group("The model configuration related")
    group.add_argument(
        "--vad_infer_config",
        type=str,
        help="VAD infer configuration",
    )
    group.add_argument(
        "--vad_model_file",
        type=str,
        help="VAD model parameter file",
    )
    group.add_argument(
        "--cmvn_file",
        type=str,
        help="Global CMVN file",
    )
    group.add_argument(
        "--asr_train_config",
        type=str,
@@ -115,6 +135,11 @@
        help="Pretrained model tag. If specify this option, *_train_config and "
             "*_file will be overwritten",
    )
    group.add_argument(
        "--beam_search_config",
        default={},
        help="The keyword arguments for transducer beam search.",
    )

    group = parser.add_argument_group("Beam-search related")
    group.add_argument(
@@ -146,12 +171,47 @@
    group.add_argument(
        "--ctc_weight",
        type=float,
        default=0.5,
        default=0.0,
        help="CTC weight in joint decoding",
    )
    group.add_argument("--lm_weight", type=float, default=1.0, help="RNNLM weight")
    group.add_argument("--ngram_weight", type=float, default=0.9, help="ngram weight")
    group.add_argument("--streaming", type=str2bool, default=False)
    group.add_argument("--simu_streaming", type=str2bool, default=False)
    group.add_argument("--chunk_size", type=int, default=16)
    group.add_argument("--left_context", type=int, default=16)
    group.add_argument("--right_context", type=int, default=0)
    group.add_argument(
        "--display_partial_hypotheses",
        type=bool,
        default=False,
        help="Whether to display partial hypotheses during chunk-by-chunk inference.",
    )    
   
    group = parser.add_argument_group("Dynamic quantization related")
    group.add_argument(
        "--quantize_asr_model",
        type=bool,
        default=False,
        help="Apply dynamic quantization to ASR model.",
    )
    group.add_argument(
        "--quantize_modules",
        nargs="*",
        default=None,
        help="""Module names to apply dynamic quantization on.
        The module names are provided as a list, where each name is separated
        by a comma (e.g.: --quantize-config=[Linear,LSTM,GRU]).
        Each specified name should be an attribute of 'torch.nn', e.g.:
        torch.nn.Linear, torch.nn.LSTM, torch.nn.GRU, ...""",
    )
    group.add_argument(
        "--quantize_dtype",
        type=str,
        default="qint8",
        choices=["float16", "qint8"],
        help="Dtype for dynamic quantization.",
    )    

    group = parser.add_argument_group("Text converter related")
    group.add_argument(
@@ -181,6 +241,72 @@
    return parser



def inference_launch(**kwargs):
    if 'mode' in kwargs:
        mode = kwargs['mode']
    else:
        logging.info("Unknown decoding mode.")
        return None
    if mode == "asr":
        from funasr.bin.asr_inference import inference_modelscope
        return inference_modelscope(**kwargs)
    elif mode == "uniasr":
        from funasr.bin.asr_inference_uniasr import inference_modelscope
        return inference_modelscope(**kwargs)
    elif mode == "paraformer":
        from funasr.bin.asr_inference_paraformer import inference_modelscope
        return inference_modelscope(**kwargs)
    elif mode == "paraformer_streaming":
        from funasr.bin.asr_inference_paraformer_streaming import inference_modelscope
        return inference_modelscope(**kwargs)
    elif mode.startswith("paraformer_vad"):
        from funasr.bin.asr_inference_paraformer import inference_modelscope_vad_punc
        return inference_modelscope_vad_punc(**kwargs)
    elif mode == "mfcca":
        from funasr.bin.asr_inference_mfcca import inference_modelscope
        return inference_modelscope(**kwargs)
    elif mode == "rnnt":
        from funasr.bin.asr_inference_rnnt import inference_modelscope
        return inference_modelscope(**kwargs)
    else:
        logging.info("Unknown decoding mode: {}".format(mode))
        return None

def inference_launch_funasr(**kwargs):
    if 'mode' in kwargs:
        mode = kwargs['mode']
    else:
        logging.info("Unknown decoding mode.")
        return None
    if mode == "asr":
        from funasr.bin.asr_inference import inference
        return inference(**kwargs)
    elif mode == "sa_asr":
        from funasr.bin.sa_asr_inference import inference
        return inference(**kwargs)
    elif mode == "uniasr":
        from funasr.bin.asr_inference_uniasr import inference
        return inference(**kwargs)
    elif mode == "paraformer":
        from funasr.bin.asr_inference_paraformer import inference_modelscope
        inference_pipeline = inference_modelscope(**kwargs)
        return inference_pipeline(kwargs["data_path_and_name_and_type"], hotword=kwargs.get("hotword", None))
    elif mode.startswith("paraformer_vad"):
        from funasr.bin.asr_inference_paraformer import inference_modelscope_vad_punc
        inference_pipeline = inference_modelscope_vad_punc(**kwargs)
        return inference_pipeline(kwargs["data_path_and_name_and_type"], hotword=kwargs.get("hotword", None))
    elif mode == "mfcca":
        from funasr.bin.asr_inference_mfcca import inference_modelscope
        return inference_modelscope(**kwargs)
    elif mode == "rnnt":
        from funasr.bin.asr_inference_rnnt import inference
        return inference(**kwargs)
    else:
        logging.info("Unknown decoding mode: {}".format(mode))
        return None


def main(cmd=None):
    print(get_commandline_args(), file=sys.stderr)
    parser = get_parser()
@@ -208,18 +334,8 @@
        os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
        os.environ["CUDA_VISIBLE_DEVICES"] = gpuid

    if args.mode == "asr":
        from funasr.bin.asr_inference import inference
        inference(**kwargs)
    elif args.mode == "uniasr":
        from funasr.bin.asr_inference_uniasr import inference
        inference(**kwargs)
    elif args.mode == "paraformer":
        from funasr.bin.asr_inference_paraformer import inference
        inference(**kwargs)
    else:
        logging.info("Unknown decoding mode: {}".format(args.mode))
    inference_launch_funasr(**kwargs)


if __name__ == "__main__":
    main()
    main()