update paraformer streaming code

haoneng.lhn

2023-04-27 7584bbd6f3e321cc8bc970739a7cfce29ffcc18b

update paraformer streaming code

 funasr/bin/asr_inference_paraformer_streaming.py

@@ -19,7 +19,6 @@

import numpy as np
import torch
import torchaudio
from typeguard import check_argument_types

from funasr.fileio.datadir_writer import DatadirWriter
@@ -40,10 +39,11 @@
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
from funasr.models.frontend.wav_frontend import WavFrontend, WavFrontendOnline
from funasr.export.models.e2e_asr_paraformer import Paraformer as Paraformer_export

np.set_printoptions(threshold=np.inf)


class Speech2Text:
    """Speech2Text class
@@ -89,7 +89,7 @@
        )
        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)
            frontend = WavFrontendOnline(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))
@@ -189,8 +189,7 @@

    @torch.no_grad()
    def __call__(
            self, cache: dict, speech: Union[torch.Tensor, np.ndarray], speech_lengths: Union[torch.Tensor, np.ndarray] = None,
            begin_time: int = 0, end_time: int = None,
            self, cache: dict, speech: Union[torch.Tensor], speech_lengths: Union[torch.Tensor] = None
    ):
        """Inference

@@ -201,38 +200,57 @@

        """
        assert check_argument_types()

        # Input as audio signal
        if isinstance(speech, np.ndarray):
            speech = torch.tensor(speech)
        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
        results = []
        cache_en = cache["encoder"]
        if speech.shape[1] < 16 * 60 and cache["is_final"]:
            cache["last_chunk"] = True
            feats = cache["feats"]
            feats_len = torch.tensor([feats.shape[1]])
        else:
            feats = speech
            feats_len = speech_lengths
        lfr_factor = max(1, (feats.size()[-1] // 80) - 1)
        feats_len = cache["encoder"]["stride"] + cache["encoder"]["pad_left"] + cache["encoder"]["pad_right"]
        feats = feats[:,cache["encoder"]["start_idx"]:cache["encoder"]["start_idx"]+feats_len,:]
        feats_len = torch.tensor([feats_len])
        batch = {"speech": feats, "speech_lengths": feats_len, "cache": cache}
            if self.frontend is not None:
                feats, feats_len = self.frontend.forward(speech, speech_lengths, cache_en["is_final"])
                feats = to_device(feats, device=self.device)
                feats_len = feats_len.int()
                self.asr_model.frontend = None
            else:
                feats = speech
                feats_len = speech_lengths

        # a. To device
            if feats.shape[1] != 0:
                if cache_en["is_final"]:
                    if feats.shape[1] + cache_en["chunk_size"][2] < cache_en["chunk_size"][1]:
                        cache_en["last_chunk"] = True
                    else:
                        # first chunk
                        feats_chunk1 = feats[:, :cache_en["chunk_size"][1], :]
                        feats_len = torch.tensor([feats_chunk1.shape[1]])
                        results_chunk1 = self.infer(feats_chunk1, feats_len, cache)

                        # last chunk
                        cache_en["last_chunk"] = True
                        feats_chunk2 = feats[:, -(feats.shape[1] + cache_en["chunk_size"][2] - cache_en["chunk_size"][1]):, :]
                        feats_len = torch.tensor([feats_chunk2.shape[1]])
                        results_chunk2 = self.infer(feats_chunk2, feats_len, cache)

                        return results_chunk1 + results_chunk2

                results = self.infer(feats, feats_len, cache)

        return results

    @torch.no_grad()
    def infer(self, feats: Union[torch.Tensor], feats_len: Union[torch.Tensor], cache: List = None):
        batch = {"speech": feats, "speech_lengths": feats_len}
        batch = to_device(batch, device=self.device)

        # b. Forward Encoder
        enc, enc_len = self.asr_model.encode_chunk(feats, feats_len, cache)
        enc, enc_len = self.asr_model.encode_chunk(feats, feats_len, cache=cache)
        if isinstance(enc, tuple):
            enc = enc[0]
        # 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_chunk(enc, cache)
        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.floor().long()
        pre_acoustic_embeds, pre_token_length= predictor_outs[0], predictor_outs[1]
        if torch.max(pre_token_length) < 1:
            return []
        decoder_outs = self.asr_model.cal_decoder_with_predictor_chunk(enc, pre_acoustic_embeds, cache)
@@ -279,163 +297,9 @@
                    text = self.tokenizer.tokens2text(token)
                else:
                    text = None

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

        # assert check_return_type(results)
        return results


class Speech2TextExport:
    """Speech2TextExport class

    """

    def __init__(
            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,
            bpemodel: str = None,
            device: str = "cpu",
            maxlenratio: float = 0.0,
            minlenratio: float = 0.0,
            dtype: str = "float32",
            beam_size: int = 20,
            ctc_weight: float = 0.5,
            lm_weight: float = 1.0,
            ngram_weight: float = 0.9,
            penalty: float = 0.0,
            nbest: int = 1,
            frontend_conf: dict = None,
            hotword_list_or_file: str = None,
            **kwargs,
    ):

        # 1. Build ASR model
        asr_model, asr_train_args = ASRTask.build_model_from_file(
            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()

        token_list = asr_model.token_list

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

        # 5. [Optional] Build Text converter: e.g. bpe-sym -> Text
        if token_type is None:
            token_type = asr_train_args.token_type
        if bpemodel is None:
            bpemodel = asr_train_args.bpemodel

        if token_type is None:
            tokenizer = None
        elif token_type == "bpe":
            if bpemodel is not None:
                tokenizer = build_tokenizer(token_type=token_type, bpemodel=bpemodel)
            else:
                tokenizer = None
        else:
            tokenizer = build_tokenizer(token_type=token_type)
        converter = TokenIDConverter(token_list=token_list)
        logging.info(f"Text tokenizer: {tokenizer}")

        # self.asr_model = asr_model
        self.asr_train_args = asr_train_args
        self.converter = converter
        self.tokenizer = tokenizer

        self.device = device
        self.dtype = dtype
        self.nbest = nbest
        self.frontend = frontend

        model = Paraformer_export(asr_model, onnx=False)
        self.asr_model = model

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

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

        """
        assert check_argument_types()

        # Input as audio signal
        if isinstance(speech, np.ndarray):
            speech = torch.tensor(speech)

        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

        enc_len_batch_total = feats_len.sum()
        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)

        decoder_outs = self.asr_model(**batch)
        decoder_out, ys_pad_lens = decoder_outs[0], decoder_outs[1]

        results = []
        b, n, d = decoder_out.size()
        for i in range(b):
            am_scores = decoder_out[i, :ys_pad_lens[i], :]

            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(
                yseq.tolist(), device=yseq.device
            )
            nbest_hyps = [Hypothesis(yseq=yseq, score=score)]

            for hyp in nbest_hyps:
                assert isinstance(hyp, (Hypothesis)), type(hyp)

                # 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()

                # remove blank symbol id, which is assumed to be 0
                token_int = list(filter(lambda x: x != 0 and x != 2, token_int))

                # 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))

        return results


@@ -536,8 +400,6 @@
        **kwargs,
):
    assert check_argument_types()
    ncpu = kwargs.get("ncpu", 1)
    torch.set_num_threads(ncpu)

    if word_lm_train_config is not None:
        raise NotImplementedError("Word LM is not implemented")
@@ -580,11 +442,9 @@
        penalty=penalty,
        nbest=nbest,
    )
    if export_mode:
        speech2text = Speech2TextExport(**speech2text_kwargs)
    else:
        speech2text = Speech2Text(**speech2text_kwargs)
        

    speech2text = Speech2Text(**speech2text_kwargs)

    def _load_bytes(input):
        middle_data = np.frombuffer(input, dtype=np.int16)
        middle_data = np.asarray(middle_data)
@@ -599,7 +459,33 @@
        offset = i.min + abs_max
        array = np.frombuffer((middle_data.astype(dtype) - offset) / abs_max, dtype=np.float32)
        return array
    

    def _prepare_cache(cache: dict = {}, chunk_size=[5,10,5], batch_size=1):
        if len(cache) > 0:
            return cache

        cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, 320)),
                    "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
                    "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], 560))}
        cache["encoder"] = cache_en

        cache_de = {"decode_fsmn": None}
        cache["decoder"] = cache_de

        return cache

    def _cache_reset(cache: dict = {}, chunk_size=[5,10,5], batch_size=1):
        if len(cache) > 0:
            cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, 320)),
                        "cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
                        "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], 560))}
            cache["encoder"] = cache_en

            cache_de = {"decode_fsmn": None}
            cache["decoder"] = cache_de

        return cache

    def _forward(
            data_path_and_name_and_type,
            raw_inputs: Union[np.ndarray, torch.Tensor] = None,
@@ -610,123 +496,35 @@
    ):

        # 3. Build data-iterator
        if data_path_and_name_and_type is not None and data_path_and_name_and_type[2] == "bytes":
            raw_inputs = _load_bytes(data_path_and_name_and_type[0])
            raw_inputs = torch.tensor(raw_inputs)
        if data_path_and_name_and_type is None and raw_inputs is not None:
            if isinstance(raw_inputs, np.ndarray):
                raw_inputs = torch.tensor(raw_inputs)
        is_final = False
        cache = {}
        chunk_size = [5, 10, 5]
        if param_dict is not None and "cache" in param_dict:
            cache = param_dict["cache"]
        if param_dict is not None and "is_final" in param_dict:
            is_final = param_dict["is_final"]
        if param_dict is not None and "chunk_size" in param_dict:
            chunk_size = param_dict["chunk_size"]

        if data_path_and_name_and_type is not None and data_path_and_name_and_type[2] == "bytes":
            raw_inputs = _load_bytes(data_path_and_name_and_type[0])
            raw_inputs = torch.tensor(raw_inputs)
        if data_path_and_name_and_type is not None and data_path_and_name_and_type[2] == "sound":
            raw_inputs = torchaudio.load(data_path_and_name_and_type[0])[0][0]
            is_final = True
        if data_path_and_name_and_type is None and raw_inputs is not None:
            if isinstance(raw_inputs, np.ndarray):
                raw_inputs = torch.tensor(raw_inputs)
        # 7 .Start for-loop
        # FIXME(kamo): The output format should be discussed about
        raw_inputs = torch.unsqueeze(raw_inputs, axis=0)
        input_lens = torch.tensor([raw_inputs.shape[1]])
        asr_result_list = []
        results = []
        asr_result = ""
        wait = True
        if len(cache) == 0:
            cache["encoder"] = {"start_idx": 0, "pad_left": 0, "stride": 10, "pad_right": 5, "cif_hidden": None, "cif_alphas": None, "is_final": is_final, "left": 0, "right": 0}
            cache_de = {"decode_fsmn": None}
            cache["decoder"] = cache_de
            cache["first_chunk"] = True
            cache["speech"] = []
            cache["accum_speech"] = 0

        if raw_inputs is not None:
            if len(cache["speech"]) == 0:
                cache["speech"] = raw_inputs
            else:
                cache["speech"] = torch.cat([cache["speech"], raw_inputs], dim=0)
            cache["accum_speech"] += len(raw_inputs)
            while cache["accum_speech"] >= 960:
                if cache["first_chunk"]:
                    if cache["accum_speech"] >= 14400:
                        speech = torch.unsqueeze(cache["speech"], axis=0)
                        speech_length = torch.tensor([len(cache["speech"])])
                        cache["encoder"]["pad_left"] = 5 
                        cache["encoder"]["pad_right"] = 5 
                        cache["encoder"]["stride"] = 10
                        cache["encoder"]["left"] = 5
                        cache["encoder"]["right"] = 0
                        results = speech2text(cache, speech, speech_length)
                        cache["accum_speech"] -= 4800
                        cache["first_chunk"] = False
                        cache["encoder"]["start_idx"] = -5
                        cache["encoder"]["is_final"] = False
                        wait = False
                    else:
                        if is_final:
                            cache["encoder"]["stride"] = len(cache["speech"]) // 960
                            cache["encoder"]["pad_left"] = 0
                            cache["encoder"]["pad_right"] = 0
                            speech = torch.unsqueeze(cache["speech"], axis=0)
                            speech_length = torch.tensor([len(cache["speech"])])
                            results = speech2text(cache, speech, speech_length)
                            cache["accum_speech"] = 0
                            wait = False
                        else:
                            break
                else:
                    if cache["accum_speech"] >= 19200:
                        cache["encoder"]["start_idx"] += 10
                        cache["encoder"]["stride"] = 10
                        cache["encoder"]["pad_left"] = 5
                        cache["encoder"]["pad_right"] = 5
                        cache["encoder"]["left"] = 0
                        cache["encoder"]["right"] = 0
                        speech = torch.unsqueeze(cache["speech"], axis=0)
                        speech_length = torch.tensor([len(cache["speech"])])
                        results = speech2text(cache, speech, speech_length)
                        cache["accum_speech"] -= 9600
                        wait = False
                    else:
                        if is_final:
                            cache["encoder"]["is_final"] = True
                            if cache["accum_speech"] >= 14400:
                                cache["encoder"]["start_idx"] += 10
                                cache["encoder"]["stride"] = 10
                                cache["encoder"]["pad_left"] = 5
                                cache["encoder"]["pad_right"] = 5
                                cache["encoder"]["left"] = 0
                                cache["encoder"]["right"] = cache["accum_speech"] // 960 - 15
                                speech = torch.unsqueeze(cache["speech"], axis=0)
                                speech_length = torch.tensor([len(cache["speech"])])
                                results = speech2text(cache, speech, speech_length)
                                cache["accum_speech"] -= 9600
                                wait = False
                            else:
                                cache["encoder"]["start_idx"] += 10
                                cache["encoder"]["stride"] = cache["accum_speech"] // 960 - 5
                                cache["encoder"]["pad_left"] = 5
                                cache["encoder"]["pad_right"] = 0
                                cache["encoder"]["left"] = 0
                                cache["encoder"]["right"] = 0
                                speech = torch.unsqueeze(cache["speech"], axis=0)
                                speech_length = torch.tensor([len(cache["speech"])])
                                results = speech2text(cache, speech, speech_length)
                                cache["accum_speech"] = 0
                                wait = False
                        else:
                            break
                
                if len(results) >= 1:
                    asr_result += results[0][0]
            if asr_result == "":
                asr_result = "sil"
            if wait:
                asr_result = "waiting_for_more_voice"
            item = {'key': "utt", 'value': asr_result}
            asr_result_list.append(item)
        else:
            return []
        cache = _prepare_cache(cache, chunk_size=chunk_size, batch_size=1)
        cache["encoder"]["is_final"] = is_final
        asr_result = speech2text(cache, raw_inputs, input_lens)
        item = {'key': "utt", 'value': asr_result}
        asr_result_list.append(item)
        if is_final:
            cache = _cache_reset(cache, chunk_size=chunk_size, batch_size=1)
        return asr_result_list

    return _forward
@@ -920,5 +718,4 @@
    #
    # 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)



 funasr/models/e2e_asr_paraformer.py

@@ -712,9 +712,9 @@

    def calc_predictor_chunk(self, encoder_out, cache=None):

        pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = \
        pre_acoustic_embeds, pre_token_length = \
            self.predictor.forward_chunk(encoder_out, cache["encoder"])
        return pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index
        return pre_acoustic_embeds, pre_token_length

    def cal_decoder_with_predictor_chunk(self, encoder_out, sematic_embeds, cache=None):
        decoder_outs = self.decoder.forward_chunk(

 funasr/models/encoder/sanm_encoder.py

@@ -6,9 +6,11 @@
import logging
import torch
import torch.nn as nn
import torch.nn.functional as F
from funasr.modules.streaming_utils.chunk_utilis import overlap_chunk
from typeguard import check_argument_types
import numpy as np
from funasr.torch_utils.device_funcs import to_device
from funasr.modules.nets_utils import make_pad_mask
from funasr.modules.attention import MultiHeadedAttention, MultiHeadedAttentionSANM, MultiHeadedAttentionSANMwithMask
from funasr.modules.embedding import SinusoidalPositionEncoder, StreamSinusoidalPositionEncoder
@@ -349,6 +351,23 @@
            return (xs_pad, intermediate_outs), olens, None
        return xs_pad, olens, None

    def _add_overlap_chunk(self, feats: np.ndarray, cache: dict = {}):
        if len(cache) == 0:
            return feats
        # process last chunk
        cache["feats"] = to_device(cache["feats"], device=feats.device)
        overlap_feats = torch.cat((cache["feats"], feats), dim=1)
        if cache["is_final"]:
            cache["feats"] = overlap_feats[:, -cache["chunk_size"][0]:, :]
            if not cache["last_chunk"]:
               padding_length = sum(cache["chunk_size"]) - overlap_feats.shape[1]
               overlap_feats = overlap_feats.transpose(1, 2)
               overlap_feats = F.pad(overlap_feats, (0, padding_length))
               overlap_feats = overlap_feats.transpose(1, 2)
        else:
            cache["feats"] = overlap_feats[:, -(cache["chunk_size"][0] + cache["chunk_size"][2]):, :]
        return overlap_feats

    def forward_chunk(self,
                      xs_pad: torch.Tensor,
                      ilens: torch.Tensor,
@@ -360,7 +379,7 @@
            xs_pad = xs_pad
        else:
            xs_pad = self.embed(xs_pad, cache)

        xs_pad = self._add_overlap_chunk(xs_pad, cache)
        encoder_outs = self.encoders0(xs_pad, None, None, None, None)
        xs_pad, masks = encoder_outs[0], encoder_outs[1]
        intermediate_outs = []

 funasr/models/predictor/cif.py

@@ -2,6 +2,7 @@
from torch import nn

import logging

import numpy as np

from funasr.torch_utils.device_funcs import to_device

from funasr.modules.nets_utils import make_pad_mask

from funasr.modules.streaming_utils.utils import sequence_mask



@@ -200,7 +201,7 @@
        return acoustic_embeds, token_num, alphas, cif_peak



    def forward_chunk(self, hidden, cache=None):

        b, t, d = hidden.size()

        batch_size, len_time, hidden_size = hidden.shape

        h = hidden

        context = h.transpose(1, 2)

        queries = self.pad(context)

@@ -211,58 +212,81 @@
        alphas = torch.nn.functional.relu(alphas * self.smooth_factor - self.noise_threshold)



        alphas = alphas.squeeze(-1)

        mask_chunk_predictor = None

        if cache is not None:

            mask_chunk_predictor = None

            mask_chunk_predictor = torch.zeros_like(alphas)

            mask_chunk_predictor[:, cache["pad_left"]:cache["stride"] + cache["pad_left"]] = 1.0

       
        if mask_chunk_predictor is not None:

            alphas = alphas * mask_chunk_predictor

      
        if cache is not None:

            if cache["is_final"]:

                alphas[:, cache["stride"] + cache["pad_left"] - 1] += 0.45

            if cache["cif_hidden"] is not None:

                hidden = torch.cat((cache["cif_hidden"], hidden), 1)

            if cache["cif_alphas"] is not None:

                alphas = torch.cat((cache["cif_alphas"], alphas), -1)



        token_num = alphas.sum(-1)

        acoustic_embeds, cif_peak = cif(hidden, alphas, self.threshold)

        len_time = alphas.size(-1)

        last_fire_place = len_time - 1

        last_fire_remainds = 0.0

        pre_alphas_length = 0

        last_fire = False

 
        mask_chunk_peak_predictor = None

        if cache is not None:

            mask_chunk_peak_predictor = None

            mask_chunk_peak_predictor = torch.zeros_like(cif_peak)

            if cache["cif_alphas"] is not None:

                pre_alphas_length = cache["cif_alphas"].size(-1)

                mask_chunk_peak_predictor[:, :pre_alphas_length] = 1.0

            mask_chunk_peak_predictor[:, pre_alphas_length + cache["pad_left"]:pre_alphas_length + cache["stride"] + cache["pad_left"]] = 1.0

            
        if mask_chunk_peak_predictor is not None:

            cif_peak = cif_peak * mask_chunk_peak_predictor.squeeze(-1)

        
        for i in range(len_time):

            if cif_peak[0][len_time - 1 - i] > self.threshold or cif_peak[0][len_time - 1 - i] == self.threshold:

                last_fire_place = len_time - 1 - i

                last_fire_remainds = cif_peak[0][len_time - 1 - i] - self.threshold

                last_fire = True

                break

        if last_fire:

           last_fire_remainds = torch.tensor([last_fire_remainds], dtype=alphas.dtype).to(alphas.device)

           cache["cif_hidden"] = hidden[:, last_fire_place:, :]

           cache["cif_alphas"] = torch.cat((last_fire_remainds.unsqueeze(0), alphas[:, last_fire_place+1:]), -1)

        else:

           cache["cif_hidden"] = hidden

           cache["cif_alphas"] = alphas

        token_num_int = token_num.floor().type(torch.int32).item()

        return acoustic_embeds[:, 0:token_num_int, :], token_num, alphas, cif_peak

        token_length = []

        list_fires = []

        list_frames = []

        cache_alphas = []

        cache_hiddens = []



        if cache is not None and "chunk_size" in cache:

            alphas[:, :cache["chunk_size"][0]] = 0.0

            alphas[:, sum(cache["chunk_size"][:2]):] = 0.0

        if cache is not None and "cif_alphas" in cache and "cif_hidden" in cache:

            cache["cif_hidden"] = to_device(cache["cif_hidden"], device=hidden.device)

            cache["cif_alphas"] = to_device(cache["cif_alphas"], device=alphas.device)

            hidden = torch.cat((cache["cif_hidden"], hidden), dim=1)

            alphas = torch.cat((cache["cif_alphas"], alphas), dim=1)

        if cache is not None and "last_chunk" in cache and cache["last_chunk"]:

            tail_hidden = torch.zeros((batch_size, 1, hidden_size), device=hidden.device)

            tail_alphas = torch.tensor([[self.tail_threshold]], device=alphas.device)

            tail_alphas = torch.tile(tail_alphas, (batch_size, 1))

            hidden = torch.cat((hidden, tail_hidden), dim=1)

            alphas = torch.cat((alphas, tail_alphas), dim=1)



        len_time = alphas.shape[1]

        for b in range(batch_size):

            integrate = 0.0

            frames = torch.zeros((hidden_size), device=hidden.device)

            list_frame = []

            list_fire = []

            for t in range(len_time):

                alpha = alphas[b][t]

                if alpha + integrate < self.threshold:

                    integrate += alpha

                    list_fire.append(integrate)

                    frames += alpha * hidden[b][t]

                else:

                    frames += (self.threshold - integrate) * hidden[b][t]

                    list_frame.append(frames)

                    integrate += alpha

                    list_fire.append(integrate)

                    integrate -= self.threshold

                    frames = integrate * hidden[b][t]



            cache_alphas.append(integrate)

            if integrate > 0.0:

                cache_hiddens.append(frames / integrate)

            else:

                cache_hiddens.append(frames)



            token_length.append(torch.tensor(len(list_frame), device=alphas.device))

            list_fires.append(list_fire)

            list_frames.append(list_frame)



        cache["cif_alphas"] = torch.stack(cache_alphas, axis=0)

        cache["cif_alphas"] = torch.unsqueeze(cache["cif_alphas"], axis=0)

        cache["cif_hidden"] = torch.stack(cache_hiddens, axis=0)

        cache["cif_hidden"] = torch.unsqueeze(cache["cif_hidden"], axis=0)



        max_token_len = max(token_length)

        if max_token_len == 0:

             return hidden, torch.stack(token_length, 0)

        list_ls = []

        for b in range(batch_size):

            pad_frames = torch.zeros((max_token_len - token_length[b], hidden_size), device=alphas.device)

            if token_length[b] == 0:

                list_ls.append(pad_frames)

            else:

                list_frames[b] = torch.stack(list_frames[b])

                list_ls.append(torch.cat((list_frames[b], pad_frames), dim=0))



        cache["cif_alphas"] = torch.stack(cache_alphas, axis=0)

        cache["cif_alphas"] = torch.unsqueeze(cache["cif_alphas"], axis=0)

        cache["cif_hidden"] = torch.stack(cache_hiddens, axis=0)

        cache["cif_hidden"] = torch.unsqueeze(cache["cif_hidden"], axis=0)

        return torch.stack(list_ls, 0), torch.stack(token_length, 0)





    def tail_process_fn(self, hidden, alphas, token_num=None, mask=None):

        b, t, d = hidden.size()


 funasr/modules/embedding.py

@@ -425,21 +425,14 @@
        return encoding.type(dtype)

    def forward(self, x, cache=None):
        start_idx = 0
        pad_left = 0
        pad_right = 0
        batch_size, timesteps, input_dim = x.size()
        start_idx = 0
        if cache is not None:
            start_idx = cache["start_idx"]
            pad_left = cache["left"]
            pad_right = cache["right"]
            cache["start_idx"] += timesteps
        positions = torch.arange(1, timesteps+start_idx+1)[None, :]
        position_encoding = self.encode(positions, input_dim, x.dtype).to(x.device)
        outputs = x + position_encoding[:, start_idx: start_idx + timesteps]
        outputs = outputs.transpose(1, 2)
        outputs = F.pad(outputs, (pad_left, pad_right))
        outputs = outputs.transpose(1, 2)
        return outputs
        return x + position_encoding[:, start_idx: start_idx + timesteps]

class StreamingRelPositionalEncoding(torch.nn.Module):
    """Relative positional encoding.