readme

游雁

2023-02-11 0a4e3b7e64e9e095cfdcd4b3c28bde7aa58839e7

 funasr/bin/asr_inference_paraformer.py

old mode 100755
new mode 100644
@@ -3,11 +3,17 @@
import logging
import sys
import time
import copy
import os
import codecs
from pathlib import Path
from typing import Optional
from typing import Sequence
from typing import Tuple
from typing import Union
from typing import Dict
from typing import Any
from typing import List

import numpy as np
import torch
@@ -30,6 +36,19 @@
from funasr.utils.types import str2bool
from funasr.utils.types import str2triple_str
from funasr.utils.types import str_or_none
from funasr.utils import asr_utils, wav_utils, postprocess_utils
from funasr.models.frontend.wav_frontend import WavFrontend
from funasr.models.e2e_asr_paraformer import BiCifParaformer, ContextualParaformer


header_colors = '\033[95m'
end_colors = '\033[0m'

global_asr_language: str = 'zh-cn'
global_sample_rate: Union[int, Dict[Any, int]] = {
    'audio_fs': 16000,
    'model_fs': 16000
}


class Speech2Text:
@@ -48,6 +67,7 @@
            self,
            asr_train_config: Union[Path, str] = None,
            asr_model_file: Union[Path, str] = None,
            cmvn_file: Union[Path, str] = None,
            lm_train_config: Union[Path, str] = None,
            lm_file: Union[Path, str] = None,
            token_type: str = None,
@@ -62,6 +82,8 @@
            ngram_weight: float = 0.9,
            penalty: float = 0.0,
            nbest: int = 1,
            frontend_conf: dict = None,
            hotword_list_or_file: str = None,
            **kwargs,
    ):
        assert check_argument_types()
@@ -69,16 +91,23 @@
        # 1. Build ASR model
        scorers = {}
        asr_model, asr_train_args = ASRTask.build_model_from_file(
            asr_train_config, asr_model_file, device
            asr_train_config, asr_model_file, cmvn_file, device
        )
        frontend = None
        if asr_train_args.frontend is not None and asr_train_args.frontend_conf is not None:
            frontend = WavFrontend(cmvn_file=cmvn_file, **asr_train_args.frontend_conf)

        logging.info("asr_model: {}".format(asr_model))
        logging.info("asr_train_args: {}".format(asr_train_args))
        asr_model.to(dtype=getattr(torch, dtype)).eval()

        ctc = CTCPrefixScorer(ctc=asr_model.ctc, eos=asr_model.eos)
        if asr_model.ctc != None:
            ctc = CTCPrefixScorer(ctc=asr_model.ctc, eos=asr_model.eos)
            scorers.update(
                ctc=ctc
            )
        token_list = asr_model.token_list
        scorers.update(
            ctc=ctc,
            length_bonus=LengthBonus(len(token_list)),
        )

@@ -120,7 +149,7 @@
        for scorer in scorers.values():
            if isinstance(scorer, torch.nn.Module):
                scorer.to(device=device, dtype=getattr(torch, dtype)).eval()
        logging.info(f"Beam_search: {beam_search}")

        logging.info(f"Decoding device={device}, dtype={dtype}")

        # 5. [Optional] Build Text converter: e.g. bpe-sym -> Text
@@ -145,22 +174,58 @@
        self.asr_train_args = asr_train_args
        self.converter = converter
        self.tokenizer = tokenizer

        # 6. [Optional] Build hotword list from file or str
        if hotword_list_or_file is None:
            self.hotword_list = None
        elif os.path.exists(hotword_list_or_file):
            self.hotword_list = []
            hotword_str_list = []
            with codecs.open(hotword_list_or_file, 'r') as fin:
                for line in fin.readlines():
                    hw = line.strip()
                    hotword_str_list.append(hw)
                    self.hotword_list.append(self.converter.tokens2ids([i for i in hw]))
                self.hotword_list.append([1])
                hotword_str_list.append('<s>')
            logging.info("Initialized hotword list from file: {}, hotword list: {}."
                .format(hotword_list_or_file, hotword_str_list))
        else:
            logging.info("Attempting to parse hotwords as str...")
            self.hotword_list = []
            hotword_str_list = []
            for hw in hotword_list_or_file.strip().split():
                hotword_str_list.append(hw)
                self.hotword_list.append(self.converter.tokens2ids([i for i in hw]))
            self.hotword_list.append([1])
            hotword_str_list.append('<s>')
            logging.info("Hotword list: {}.".format(hotword_str_list))


        is_use_lm = lm_weight != 0.0 and lm_file is not None
        if (ctc_weight == 0.0 or asr_model.ctc == None) and not is_use_lm:
            beam_search = None
        self.beam_search = beam_search
        logging.info(f"Beam_search: {self.beam_search}")
        self.beam_search_transducer = beam_search_transducer
        self.maxlenratio = maxlenratio
        self.minlenratio = minlenratio
        self.device = device
        self.dtype = dtype
        self.nbest = nbest
        self.frontend = frontend
        self.encoder_downsampling_factor = 1
        if asr_train_args.encoder == "data2vec_encoder" or asr_train_args.encoder_conf["input_layer"] == "conv2d":
            self.encoder_downsampling_factor = 4

    @torch.no_grad()
    def __call__(
            self, speech: Union[torch.Tensor, np.ndarray]
            self, speech: Union[torch.Tensor, np.ndarray], speech_lengths: Union[torch.Tensor, np.ndarray] = None
    ):
        """Inference

        Args:
                data: Input speech data
                speech: Input speech data
        Returns:
                text, token, token_int, hyp

@@ -171,12 +236,16 @@
        if isinstance(speech, np.ndarray):
            speech = torch.tensor(speech)

        # data: (Nsamples,) -> (1, Nsamples)
        speech = speech.unsqueeze(0).to(getattr(torch, self.dtype))
        lfr_factor = max(1, (speech.size()[-1]//80)-1)
        # lengths: (1,)
        lengths = speech.new_full([1], dtype=torch.long, fill_value=speech.size(1))
        batch = {"speech": speech, "speech_lengths": lengths}
        if self.frontend is not None:
            feats, feats_len = self.frontend.forward(speech, speech_lengths)
            feats = to_device(feats, device=self.device)
            feats_len = feats_len.int()
            self.asr_model.frontend = None
        else:
            feats = speech
            feats_len = speech_lengths
        lfr_factor = max(1, (feats.size()[-1] // 80) - 1)
        batch = {"speech": feats, "speech_lengths": feats_len}

        # a. To device
        batch = to_device(batch, device=self.device)
@@ -185,78 +254,170 @@
        enc, enc_len = self.asr_model.encode(**batch)
        if isinstance(enc, tuple):
            enc = enc[0]
        assert len(enc) == 1, len(enc)
        # assert len(enc) == 1, len(enc)
        enc_len_batch_total = torch.sum(enc_len).item() * self.encoder_downsampling_factor

        predictor_outs = self.asr_model.calc_predictor(enc, enc_len)
        pre_acoustic_embeds, pre_token_length = predictor_outs[0], predictor_outs[1]
        pre_token_length = torch.tensor([pre_acoustic_embeds.size(1)], device=pre_acoustic_embeds.device)
        decoder_outs = self.asr_model.cal_decoder_with_predictor(enc, enc_len, pre_acoustic_embeds, pre_token_length)
        decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]
        pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = predictor_outs[0], predictor_outs[1], \
                                                                        predictor_outs[2], predictor_outs[3]
        pre_token_length = pre_token_length.round().long()
        if torch.max(pre_token_length) < 1:
            return []
        if not isinstance(self.asr_model, ContextualParaformer):
            if self.hotword_list:
                logging.warning("Hotword is given but asr model is not a ContextualParaformer.")
            decoder_outs = self.asr_model.cal_decoder_with_predictor(enc, enc_len, pre_acoustic_embeds, pre_token_length)
            decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]
        else:
            decoder_outs = self.asr_model.cal_decoder_with_predictor(enc, enc_len, pre_acoustic_embeds, pre_token_length, hw_list=self.hotword_list)
            decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]

        nbest_hyps = self.beam_search(
            x=enc[0], am_scores=decoder_out[0], maxlenratio=self.maxlenratio, minlenratio=self.minlenratio
        )

        nbest_hyps = nbest_hyps[: self.nbest]
        results = []
        for hyp in nbest_hyps:
            assert isinstance(hyp, (Hypothesis)), type(hyp)
        b, n, d = decoder_out.size()
        for i in range(b):
            x = enc[i, :enc_len[i], :]
            am_scores = decoder_out[i, :pre_token_length[i], :]
            if self.beam_search is not None:
                nbest_hyps = self.beam_search(
                    x=x, am_scores=am_scores, maxlenratio=self.maxlenratio, minlenratio=self.minlenratio
                )

            # remove sos/eos and get results
            last_pos = -1
            if isinstance(hyp.yseq, list):
                token_int = hyp.yseq[1:last_pos]
                nbest_hyps = nbest_hyps[: self.nbest]
            else:
                token_int = hyp.yseq[1:last_pos].tolist()
                yseq = am_scores.argmax(dim=-1)
                score = am_scores.max(dim=-1)[0]
                score = torch.sum(score, dim=-1)
                # pad with mask tokens to ensure compatibility with sos/eos tokens
                yseq = torch.tensor(
                    [self.asr_model.sos] + yseq.tolist() + [self.asr_model.eos], device=yseq.device
                )
                nbest_hyps = [Hypothesis(yseq=yseq, score=score)]

            # remove blank symbol id, which is assumed to be 0
            token_int = list(filter(lambda x: x != 0, token_int))
            for hyp in nbest_hyps:
                assert isinstance(hyp, (Hypothesis)), type(hyp)

            # Change integer-ids to tokens
            token = self.converter.ids2tokens(token_int)
                # remove sos/eos and get results
                last_pos = -1
                if isinstance(hyp.yseq, list):
                    token_int = hyp.yseq[1:last_pos]
                else:
                    token_int = hyp.yseq[1:last_pos].tolist()

            if self.tokenizer is not None:
                text = self.tokenizer.tokens2text(token)
            else:
                text = None
                # remove blank symbol id, which is assumed to be 0
                token_int = list(filter(lambda x: x != 0 and x != 2, token_int))

            results.append((text, token, token_int, hyp, speech.size(1), lfr_factor))
                # Change integer-ids to tokens
                token = self.converter.ids2tokens(token_int)

                if self.tokenizer is not None:
                    text = self.tokenizer.tokens2text(token)
                else:
                    text = None

                results.append((text, token, token_int, hyp, enc_len_batch_total, lfr_factor))

        # assert check_return_type(results)
        return results


def inference(
        output_dir: str,
        maxlenratio: float,
        minlenratio: float,
        batch_size: int,
        dtype: str,
        beam_size: int,
        ngpu: int,
        seed: int,
        ctc_weight: float,
        lm_weight: float,
        ngram_weight: float,
        penalty: float,
        nbest: int,
        num_workers: int,
        log_level: Union[int, str],
        data_path_and_name_and_type: Sequence[Tuple[str, str, str]],
        key_file: Optional[str],
        data_path_and_name_and_type,
        asr_train_config: Optional[str],
        asr_model_file: Optional[str],
        lm_train_config: Optional[str],
        lm_file: Optional[str],
        word_lm_train_config: Optional[str],
        token_type: Optional[str],
        bpemodel: Optional[str],
        allow_variable_data_keys: bool,
        cmvn_file: Optional[str] = None,
        raw_inputs: Union[np.ndarray, torch.Tensor] = None,
        lm_train_config: Optional[str] = None,
        lm_file: Optional[str] = None,
        token_type: Optional[str] = None,
        key_file: Optional[str] = None,
        word_lm_train_config: Optional[str] = None,
        bpemodel: Optional[str] = None,
        allow_variable_data_keys: bool = False,
        streaming: bool = False,
        output_dir: Optional[str] = None,
        dtype: str = "float32",
        seed: int = 0,
        ngram_weight: float = 0.9,
        nbest: int = 1,
        num_workers: int = 1,

        **kwargs,
):
    inference_pipeline = inference_modelscope(
        maxlenratio=maxlenratio,
        minlenratio=minlenratio,
        batch_size=batch_size,
        beam_size=beam_size,
        ngpu=ngpu,
        ctc_weight=ctc_weight,
        lm_weight=lm_weight,
        penalty=penalty,
        log_level=log_level,
        asr_train_config=asr_train_config,
        asr_model_file=asr_model_file,
        cmvn_file=cmvn_file,
        raw_inputs=raw_inputs,
        lm_train_config=lm_train_config,
        lm_file=lm_file,
        token_type=token_type,
        key_file=key_file,
        word_lm_train_config=word_lm_train_config,
        bpemodel=bpemodel,
        allow_variable_data_keys=allow_variable_data_keys,
        streaming=streaming,
        output_dir=output_dir,
        dtype=dtype,
        seed=seed,
        ngram_weight=ngram_weight,
        nbest=nbest,
        num_workers=num_workers,

        **kwargs,
    )
    return inference_pipeline(data_path_and_name_and_type, raw_inputs)


def inference_modelscope(
        maxlenratio: float,
        minlenratio: float,
        batch_size: int,
        beam_size: int,
        ngpu: int,
        ctc_weight: float,
        lm_weight: float,
        penalty: float,
        log_level: Union[int, str],
        # data_path_and_name_and_type,
        asr_train_config: Optional[str],
        asr_model_file: Optional[str],
        cmvn_file: Optional[str] = None,
        lm_train_config: Optional[str] = None,
        lm_file: Optional[str] = None,
        token_type: Optional[str] = None,
        key_file: Optional[str] = None,
        word_lm_train_config: Optional[str] = None,
        bpemodel: Optional[str] = None,
        allow_variable_data_keys: bool = False,
        dtype: str = "float32",
        seed: int = 0,
        ngram_weight: float = 0.9,
        nbest: int = 1,
        num_workers: int = 1,
        output_dir: Optional[str] = None,
        param_dict: dict = None,
        **kwargs,
):
    assert check_argument_types()
    if batch_size > 1:
        raise NotImplementedError("batch decoding is not implemented")

    if word_lm_train_config is not None:
        raise NotImplementedError("Word LM is not implemented")
    if ngpu > 1:
@@ -267,10 +428,16 @@
        format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
    )

    if ngpu >= 1:
    if param_dict is not None:
        hotword_list_or_file = param_dict.get('hotword')
    else:
        hotword_list_or_file = None

    if ngpu >= 1 and torch.cuda.is_available():
        device = "cuda"
    else:
        device = "cpu"
        batch_size = 1

    # 1. Set random-seed
    set_all_random_seed(seed)
@@ -279,6 +446,7 @@
    speech2text_kwargs = dict(
        asr_train_config=asr_train_config,
        asr_model_file=asr_model_file,
        cmvn_file=cmvn_file,
        lm_train_config=lm_train_config,
        lm_file=lm_file,
        token_type=token_type,
@@ -293,73 +461,104 @@
        ngram_weight=ngram_weight,
        penalty=penalty,
        nbest=nbest,
        hotword_list_or_file=hotword_list_or_file,
    )
    speech2text = Speech2Text(**speech2text_kwargs)

    # 3. Build data-iterator
    loader = ASRTask.build_streaming_iterator(
        data_path_and_name_and_type,
        dtype=dtype,
        batch_size=batch_size,
        key_file=key_file,
        num_workers=num_workers,
        preprocess_fn=ASRTask.build_preprocess_fn(speech2text.asr_train_args, False),
        collate_fn=ASRTask.build_collate_fn(speech2text.asr_train_args, False),
        allow_variable_data_keys=allow_variable_data_keys,
        inference=True,
    )
    def _forward(
            data_path_and_name_and_type,
            raw_inputs: Union[np.ndarray, torch.Tensor] = None,
            output_dir_v2: Optional[str] = None,
            fs: dict = None,
            param_dict: dict = None,
    ):
        # 3. Build data-iterator
        if data_path_and_name_and_type is None and raw_inputs is not None:
            if isinstance(raw_inputs, torch.Tensor):
                raw_inputs = raw_inputs.numpy()
            data_path_and_name_and_type = [raw_inputs, "speech", "waveform"]
        loader = ASRTask.build_streaming_iterator(
            data_path_and_name_and_type,
            dtype=dtype,
            fs=fs,
            batch_size=batch_size,
            key_file=key_file,
            num_workers=num_workers,
            preprocess_fn=ASRTask.build_preprocess_fn(speech2text.asr_train_args, False),
            collate_fn=ASRTask.build_collate_fn(speech2text.asr_train_args, False),
            allow_variable_data_keys=allow_variable_data_keys,
            inference=True,
        )

    forward_time_total = 0.0
    length_total = 0.0
    # 7 .Start for-loop
    # FIXME(kamo): The output format should be discussed about
    with DatadirWriter(output_dir) as writer:
        forward_time_total = 0.0
        length_total = 0.0
        finish_count = 0
        file_count = 1
        # 7 .Start for-loop
        # FIXME(kamo): The output format should be discussed about
        asr_result_list = []
        output_path = output_dir_v2 if output_dir_v2 is not None else output_dir
        if output_path is not None:
            writer = DatadirWriter(output_path)
        else:
            writer = None

        for keys, batch in loader:
            assert isinstance(batch, dict), type(batch)
            assert all(isinstance(s, str) for s in keys), keys
            _bs = len(next(iter(batch.values())))
            assert len(keys) == _bs, f"{len(keys)} != {_bs}"
            batch = {k: v[0] for k, v in batch.items() if not k.endswith("_lengths")}
            # batch = {k: v for k, v in batch.items() if not k.endswith("_lengths")}

            logging.info("decoding, utt_id: {}".format(keys))
            # N-best list of (text, token, token_int, hyp_object)

            try:
                time_beg = time.time()
                results = speech2text(**batch)
                time_end = time.time()
                forward_time = time_end - time_beg
                lfr_factor = results[0][-1]
                length = results[0][-2]
                results = [results[0][:-2]]
                forward_time_total += forward_time
                length_total += length
                logging.info(
                    "decoding, feature length: {}, forward_time: {:.4f}, rtf: {:.4f}".
                        format(length, forward_time, 100 * forward_time / (length*lfr_factor)))
            except TooShortUttError as e:
                logging.warning(f"Utterance {keys} {e}")
            time_beg = time.time()
            results = speech2text(**batch)
            if len(results) < 1:
                hyp = Hypothesis(score=0.0, scores={}, states={}, yseq=[])
                results = [[" ", ["<space>"], [2], hyp]] * nbest
                results = [[" ", ["sil"], [2], hyp, 10, 6]] * nbest
            time_end = time.time()
            forward_time = time_end - time_beg
            lfr_factor = results[0][-1]
            length = results[0][-2]
            forward_time_total += forward_time
            length_total += length
            rtf_cur = "decoding, feature length: {}, forward_time: {:.4f}, rtf: {:.4f}".format(length, forward_time, 100 * forward_time / (length * lfr_factor))
            logging.info(rtf_cur)

            # Only supporting batch_size==1
            key = keys[0]
            for n, (text, token, token_int, hyp) in zip(range(1, nbest + 1), results):
                # Create a directory: outdir/{n}best_recog
                ibest_writer = writer[f"{n}best_recog"]
            for batch_id in range(_bs):
                result = [results[batch_id][:-2]]

                # Write the result to each file
                ibest_writer["token"][key] = " ".join(token)
                ibest_writer["token_int"][key] = " ".join(map(str, token_int))
                ibest_writer["score"][key] = str(hyp.score)
                key = keys[batch_id]
                for n, (text, token, token_int, hyp) in zip(range(1, nbest + 1), result):
                    # Create a directory: outdir/{n}best_recog
                    if writer is not None:
                        ibest_writer = writer[f"{n}best_recog"]

                if text is not None:
                    ibest_writer["text"][key] = text
                        # Write the result to each file
                        ibest_writer["token"][key] = " ".join(token)
                        # ibest_writer["token_int"][key] = " ".join(map(str, token_int))
                        ibest_writer["score"][key] = str(hyp.score)
                        ibest_writer["rtf"][key] = rtf_cur

                logging.info("decoding, predictions: {}".format(text))
                    if text is not None:
                        text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
                        item = {'key': key, 'value': text_postprocessed}
                        asr_result_list.append(item)
                        finish_count += 1
                        # asr_utils.print_progress(finish_count / file_count)
                        if writer is not None:
                            ibest_writer["text"][key] = text_postprocessed

    logging.info("decoding, feature length total: {}, forward_time total: {:.4f}, rtf avg: {:.4f}".
                 format(length_total, forward_time_total, 100 * forward_time_total / (length_total*lfr_factor)))
                    logging.info("decoding, utt: {}, predictions: {}".format(key, text))
        rtf_avg = "decoding, feature length total: {}, forward_time total: {:.4f}, rtf avg: {:.4f}".format(length_total, forward_time_total, 100 * forward_time_total / (length_total * lfr_factor))
        logging.info(rtf_avg)
        if writer is not None:
            ibest_writer["rtf"]["rtf_avf"] = rtf_avg
        return asr_result_list

    return _forward


def get_parser():
@@ -398,12 +597,17 @@
        default=1,
        help="The number of workers used for DataLoader",
    )

    parser.add_argument(
        "--hotword",
        type=str_or_none,
        default=None,
        help="hotword file path or hotwords seperated by space"
    )
    group = parser.add_argument_group("Input data related")
    group.add_argument(
        "--data_path_and_name_and_type",
        type=str2triple_str,
        required=True,
        required=False,
        action="append",
    )
    group.add_argument("--key_file", type=str_or_none)
@@ -419,6 +623,11 @@
        "--asr_model_file",
        type=str,
        help="ASR model parameter file",
    )
    group.add_argument(
        "--cmvn_file",
        type=str,
        help="Global cmvn file",
    )
    group.add_argument(
        "--lm_train_config",
@@ -494,6 +703,8 @@
        default=None,
        help="",
    )
    group.add_argument("--raw_inputs", type=list, default=None)
    # example=[{'key':'EdevDEWdIYQ_0021','file':'/mnt/data/jiangyu.xzy/test_data/speech_io/SPEECHIO_ASR_ZH00007_zhibodaihuo/wav/EdevDEWdIYQ_0021.wav'}])

    group = parser.add_argument_group("Text converter related")
    group.add_argument(
@@ -519,10 +730,22 @@
    print(get_commandline_args(), file=sys.stderr)
    parser = get_parser()
    args = parser.parse_args(cmd)
    param_dict = {'hotword': args.hotword}
    kwargs = vars(args)
    kwargs.pop("config", None)
    kwargs['param_dict'] = param_dict
    inference(**kwargs)


if __name__ == "__main__":
    main()

    # from modelscope.pipelines import pipeline
    # from modelscope.utils.constant import Tasks
    #
    # inference_16k_pipline = pipeline(
    #     task=Tasks.auto_speech_recognition,
    #     model='damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch')
    #
    # rec_result = inference_16k_pipline(audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav')
    # print(rec_result)