Dev gzf decoding (#1695)

zhifu gao

2024-05-06 00d0df3a1018c63ec8c5d13e611f53c564c0a7e2

Dev gzf decoding (#1695)

* resume from step

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* train_loss_avg train_acc_avg

* train_loss_avg train_acc_avg

* train_loss_avg train_acc_avg

* log step

* wav is not exist

* wav is not exist

* decoding

* decoding

 examples/industrial_data_pretraining/conformer/demo.py

@@ -8,6 +8,7 @@
model = AutoModel(model="iic/speech_conformer_asr_nat-zh-cn-16k-aishell2-vocab5212-pytorch")

res = model.generate(
    input="https://modelscope.oss-cn-beijing.aliyuncs.com/test/audios/asr_example.wav"
    input="https://modelscope.oss-cn-beijing.aliyuncs.com/test/audios/asr_example.wav",
    decoding_ctc_weight=0.0,
)
print(res)

 examples/industrial_data_pretraining/sense_voice/demo_fsmn.py

New file
@@ -0,0 +1,27 @@
#!/usr/bin/env python3
# -*- encoding: utf-8 -*-
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
#  MIT License  (https://opensource.org/licenses/MIT)

from funasr import AutoModel

model = AutoModel(
    model="/Users/zhifu/Downloads/modelscope_models/SenseVoiceModelscopeFSMN",
    vad_model="iic/speech_fsmn_vad_zh-cn-16k-common-pytorch",
    vad_kwargs={"max_single_segment_time": 30000},
)


input_wav = (
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"
)

DecodingOptions = {
    "task": ("ASR", "AED", "SER"),
    "language": "auto",
    "fp16": True,
    "gain_event": True,
}

res = model.generate(input=input_wav, batch_size_s=0, DecodingOptions=DecodingOptions, beam_size=5)
print(res)

 funasr/datasets/audio_datasets/update_jsonl.py

@@ -46,16 +46,17 @@
    data = json.loads(line.strip())

    wav_path = data["source"].replace("/cpfs01", "/cpfs_speech/data")
    waveform, _ = librosa.load(wav_path, sr=16000)
    sample_num = len(waveform)
    source_len = int(sample_num / 16000 * 1000 / 10)
    source_len_old = data["source_len"]
    # if (source_len_old - source_len) > 100 or (source_len - source_len_old) > 100:
    #     logging.info(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
    data["source_len"] = source_len
    data["source"] = wav_path
    jsonl_line = json.dumps(data, ensure_ascii=False)
    lines[i] = jsonl_line
    if os.path.exists(wav_path):
        waveform, _ = librosa.load(wav_path, sr=16000)
        sample_num = len(waveform)
        source_len = int(sample_num / 16000 * 1000 / 10)
        source_len_old = data["source_len"]
        # if (source_len_old - source_len) > 100 or (source_len - source_len_old) > 100:
        #     logging.info(f"old: {source_len_old}, new: {source_len}, wav: {wav_path}")
        data["source_len"] = source_len
        data["source"] = wav_path
        jsonl_line = json.dumps(data, ensure_ascii=False)
        lines[i] = jsonl_line


def update_wav_len(jsonl_file_list_in, jsonl_file_out_dir, ncpu=1):

 funasr/datasets/sense_voice_datasets/datasets.py

@@ -2,7 +2,7 @@

import torch
import random

import traceback
from funasr.register import tables
from funasr.utils.load_utils import extract_fbank, load_audio_text_image_video

@@ -73,15 +73,17 @@
            if idx == 0:
                index_cur = index
            else:
                if index <= self.retry:
                    index_cur = index + idx
                else:
                    index_cur = torch.randint(0, index, ()).item()
                index_cur = torch.randint(0, len(self.index_ds), ()).item()

            item = self.index_ds[index_cur]

            source = item["source"]
            data_src = load_audio_text_image_video(source, fs=self.fs)
            try:
                data_src = load_audio_text_image_video(source, fs=self.fs)
            except Exception as e:
                logging.error(f"Loading wav failed! {str(e)}, {traceback.format_exc()}")
                continue

            if self.preprocessor_speech:
                data_src = self.preprocessor_speech(data_src, fs=self.fs)
            speech, speech_lengths = extract_fbank(
@@ -186,7 +188,7 @@
                )

        if self.batch_type != "example":
            for i in range(3):
            for i in range(10):
                outputs = self._filter_badcase(outputs, i=i)

        return outputs

 funasr/models/sense_voice/decoder.py

@@ -15,6 +15,7 @@
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from funasr.models.transformer.utils.mask import subsequent_mask


class LayerNorm(nn.LayerNorm):
@@ -336,6 +337,29 @@

        return x

    def init_state(self, x):
        state = {}

        return state

    def final_score(self, state) -> float:
        """Score eos (optional).

        Args:
            state: Scorer state for prefix tokens

        Returns:
            float: final score

        """
        return 0.0

    def score(self, ys, state, x):
        """Score."""
        ys_mask = subsequent_mask(len(ys), device=x.device).unsqueeze(0)
        logp = self.forward(ys.unsqueeze(0), x.unsqueeze(0), cache=state)
        return logp.squeeze(0)[-1, :], state


class MultiHeadedAttentionSANMDecoder(nn.Module):
    """Multi-Head Attention layer.
@@ -443,9 +467,19 @@
        kv_cache: Optional[dict] = None,
        **kwargs,
    ):
        cache = kwargs.get("cache", {})
        layer = kwargs.get("layer", 0)
        is_pad_mask = kwargs.get("is_pad_mask", False)
        is_pad_memory_mask = kwargs.get("is_pad_memory_mask", False)
        x = x + self.attn(self.attn_ln(x), mask=None, kv_cache=kv_cache, is_pad_mask=is_pad_mask)[0]

        fsmn_cache = cache[layer]["fsmn_cache"] if len(cache) > 0 else None
        # if fsmn_cache is not None:
        #     x = x[:, -1:]
        att_res, fsmn_cache = self.attn(self.attn_ln(x), mask=None, cache=fsmn_cache)
        # if len(cache)>1:
        #     cache[layer]["fsmn_cache"] = fsmn_cache
        #     x = x[:, -1:]
        x = x + att_res
        if self.cross_attn:
            x = (
                x
@@ -510,10 +544,9 @@

        ys_in_lens = kwargs.get("ys_in_lens", None)

        offset = next(iter(kv_cache.values())).shape[1] if kv_cache else 0
        tgt, memory = x, xa
        tgt[tgt == -1] = 0
        tgt = self.token_embedding(tgt) + self.positional_embedding[offset : offset + tgt.size(1)]
        tgt = self.token_embedding(tgt) + self.positional_embedding[: tgt.size(1)]
        # tgt = self.dropout(tgt)

        x = tgt.to(memory.dtype)
@@ -531,9 +564,40 @@
                memory_mask=memory_mask,
                is_pad_mask=False,
                is_pad_memory_mask=True,
                cache=kwargs.get("cache", None),
                layer=layer,
            )

        x = self.ln(x)
        x = (x @ torch.transpose(self.token_embedding.weight.to(x.dtype), 0, 1)).float()

        return x

    def init_state(self, x):
        state = {}
        for layer, block in enumerate(self.blocks):
            state[layer] = {
                "fsmn_cache": None,
                "memory_key": None,
                "memory_value": None,
            }

        return state

    def final_score(self, state) -> float:
        """Score eos (optional).

        Args:
            state: Scorer state for prefix tokens

        Returns:
            float: final score

        """
        return 0.0

    def score(self, ys, state, x):
        """Score."""
        ys_mask = subsequent_mask(len(ys), device=x.device).unsqueeze(0)
        logp = self.forward(ys.unsqueeze(0), x.unsqueeze(0), cache=state)
        return logp.squeeze(0)[-1, :], state

 funasr/models/sense_voice/model.py

@@ -15,6 +15,7 @@
from funasr.train_utils.device_funcs import force_gatherable
from . import whisper_lib as whisper
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils.datadir_writer import DatadirWriter

from funasr.register import tables

@@ -395,6 +396,42 @@

        return loss_att, acc_att, None, None

    def init_beam_search(
        self,
        **kwargs,
    ):
        from .search import BeamSearch

        from funasr.models.transformer.scorers.length_bonus import LengthBonus

        # 1. Build ASR model
        scorers = {}

        scorers.update(
            decoder=self.model.decoder,
            length_bonus=LengthBonus(self.vocab_size),
        )

        weights = dict(
            decoder=1.0,
            ctc=0.0,
            lm=0.0,
            ngram=0.0,
            length_bonus=kwargs.get("penalty", 0.0),
        )
        beam_search = BeamSearch(
            beam_size=kwargs.get("beam_size", 5),
            weights=weights,
            scorers=scorers,
            sos=None,
            eos=None,
            vocab_size=self.vocab_size,
            token_list=None,
            pre_beam_score_key="full",
        )

        self.beam_search = beam_search

    def inference(
        self,
        data_in,
@@ -406,6 +443,12 @@
    ):
        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

        # init beamsearch
        if not hasattr(self, "beam_search") or self.beam_search is None:
            logging.info("enable beam_search")
            self.init_beam_search(**kwargs)
            self.nbest = kwargs.get("nbest", 1)

        if frontend is None and not hasattr(self, "frontend"):
            frontend_class = tables.frontend_classes.get("WhisperFrontend")
@@ -455,25 +498,65 @@
            task = [task]
        task = "".join([f"<|{x}|>" for x in task])
        initial_prompt = kwargs.get("initial_prompt", f"<|startoftranscript|>{task}")
        DecodingOptions["initial_prompt"] = initial_prompt

        language = DecodingOptions.get("language", None)
        language = None if language == "auto" else language
        DecodingOptions["language"] = language

        DecodingOptions["vocab_path"] = kwargs["tokenizer_conf"].get("vocab_path", None)
        sos = f"{initial_prompt}<|{language}|>" if language is not None else initial_prompt
        sos_int = tokenizer.encode(sos, allowed_special="all")
        eos = kwargs.get("model_conf").get("eos")
        eos_int = tokenizer.encode(eos, allowed_special="all")
        self.beam_search.sos = sos_int
        self.beam_search.eos = eos_int[0]

        if "without_timestamps" not in DecodingOptions:
            DecodingOptions["without_timestamps"] = True
        encoder_out, encoder_out_lens = self.encode(
            speech[None, :, :].permute(0, 2, 1), speech_lengths
        )

        options = whisper.DecodingOptions(**DecodingOptions)
        # c. Passed the encoder result and the beam search
        nbest_hyps = self.beam_search(
            x=encoder_out[0],
            maxlenratio=kwargs.get("maxlenratio", 0.0),
            minlenratio=kwargs.get("minlenratio", 0.0),
        )

        result = whisper.decode(self.model, speech, options)
        text = f"{result.text}"
        nbest_hyps = nbest_hyps[: self.nbest]

        results = []
        result_i = {"key": key[0], "text": text}
        b, n, d = encoder_out.size()
        for i in range(b):

        results.append(result_i)
            for nbest_idx, hyp in enumerate(nbest_hyps):
                ibest_writer = None
                if kwargs.get("output_dir") is not None:
                    if not hasattr(self, "writer"):
                        self.writer = DatadirWriter(kwargs.get("output_dir"))
                    ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]

                # 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 != self.eos and x != self.sos and x != self.blank_id, token_int
                #     )
                # )

                # Change integer-ids to tokens
                # token = tokenizer.ids2tokens(token_int)
                text = tokenizer.decode(token_int)

                result_i = {"key": key[i], "text": text}
                results.append(result_i)

                if ibest_writer is not None:
                    # ibest_writer["token"][key[i]] = " ".join(token)
                    ibest_writer["text"][key[i]] = text

        return results, meta_data

@@ -497,12 +580,14 @@
        # decoder
        del model.decoder
        decoder = kwargs.get("decoder", "SenseVoiceDecoder")
        decoder_conf = kwargs.get("decoder_conf", {})
        decoder_class = tables.decoder_classes.get(decoder)
        decoder = decoder_class(
            vocab_size=dims.n_vocab,
            encoder_output_size=dims.n_audio_state,
            **decoder_conf,
            n_vocab=dims.n_vocab,
            n_ctx=dims.n_text_ctx,
            n_state=dims.n_text_state,
            n_head=dims.n_text_head,
            n_layer=dims.n_text_layer,
            **kwargs.get("decoder_conf"),
        )
        model.decoder = decoder

@@ -512,7 +597,7 @@

        self.activation_checkpoint = kwargs.get("activation_checkpoint", False)
        self.ignore_id = kwargs.get("ignore_id", -1)
        self.vocab_size = kwargs.get("vocab_size", -1)
        self.vocab_size = dims.n_vocab
        self.length_normalized_loss = kwargs.get("length_normalized_loss", True)
        self.criterion_att = LabelSmoothingLoss(
            size=self.vocab_size,
@@ -630,6 +715,42 @@

        return loss_att, acc_att, None, None

    def init_beam_search(
        self,
        **kwargs,
    ):
        from .search import BeamSearch

        from funasr.models.transformer.scorers.length_bonus import LengthBonus

        # 1. Build ASR model
        scorers = {}

        scorers.update(
            decoder=self.model.decoder,
            length_bonus=LengthBonus(self.vocab_size),
        )

        weights = dict(
            decoder=1.0,
            ctc=0.0,
            lm=0.0,
            ngram=0.0,
            length_bonus=kwargs.get("penalty", 0.0),
        )
        beam_search = BeamSearch(
            beam_size=kwargs.get("beam_size", 5),
            weights=weights,
            scorers=scorers,
            sos=None,
            eos=None,
            vocab_size=self.vocab_size,
            token_list=None,
            pre_beam_score_key="full",
        )

        self.beam_search = beam_search

    def inference(
        self,
        data_in,
@@ -641,6 +762,12 @@
    ):
        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

        # init beamsearch
        if not hasattr(self, "beam_search") or self.beam_search is None:
            logging.info("enable beam_search")
            self.init_beam_search(**kwargs)
            self.nbest = kwargs.get("nbest", 1)

        if frontend is None and not hasattr(self, "frontend"):
            frontend_class = tables.frontend_classes.get("WhisperFrontend")
@@ -690,24 +817,64 @@
            task = [task]
        task = "".join([f"<|{x}|>" for x in task])
        initial_prompt = kwargs.get("initial_prompt", f"<|startoftranscript|>{task}")
        DecodingOptions["initial_prompt"] = initial_prompt

        language = DecodingOptions.get("language", None)
        language = None if language == "auto" else language
        DecodingOptions["language"] = language

        DecodingOptions["vocab_path"] = kwargs["tokenizer_conf"].get("vocab_path", None)
        sos = f"{initial_prompt}<|{language}|>" if language is not None else initial_prompt
        sos_int = tokenizer.encode(sos, allowed_special="all")
        eos = kwargs.get("model_conf").get("eos")
        eos_int = tokenizer.encode(eos, allowed_special="all")
        self.beam_search.sos = sos_int
        self.beam_search.eos = eos_int[0]

        if "without_timestamps" not in DecodingOptions:
            DecodingOptions["without_timestamps"] = True
        encoder_out, encoder_out_lens = self.encode(
            speech[None, :, :].permute(0, 2, 1), speech_lengths
        )

        options = whisper.DecodingOptions(**DecodingOptions)
        # c. Passed the encoder result and the beam search
        nbest_hyps = self.beam_search(
            x=encoder_out[0],
            maxlenratio=kwargs.get("maxlenratio", 0.0),
            minlenratio=kwargs.get("minlenratio", 0.0),
        )

        result = whisper.decode(self.model, speech, options)
        text = f"{result.text}"
        nbest_hyps = nbest_hyps[: self.nbest]

        results = []
        result_i = {"key": key[0], "text": text}
        b, n, d = encoder_out.size()
        for i in range(b):

        results.append(result_i)
            for nbest_idx, hyp in enumerate(nbest_hyps):
                ibest_writer = None
                if kwargs.get("output_dir") is not None:
                    if not hasattr(self, "writer"):
                        self.writer = DatadirWriter(kwargs.get("output_dir"))
                    ibest_writer = self.writer[f"{nbest_idx + 1}best_recog"]

                # 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 != self.eos and x != self.sos and x != self.blank_id, token_int
                #     )
                # )

                # Change integer-ids to tokens
                # token = tokenizer.ids2tokens(token_int)
                text = tokenizer.decode(token_int)

                result_i = {"key": key[i], "text": text}
                results.append(result_i)

                if ibest_writer is not None:
                    # ibest_writer["token"][key[i]] = " ".join(token)
                    ibest_writer["text"][key[i]] = text

        return results, meta_data

 funasr/models/sense_voice/search.py

New file
@@ -0,0 +1,453 @@
from itertools import chain
import logging
from typing import Any
from typing import Dict
from typing import List
from typing import NamedTuple
from typing import Tuple
from typing import Union

import torch

from funasr.metrics.common import end_detect
from funasr.models.transformer.scorers.scorer_interface import PartialScorerInterface
from funasr.models.transformer.scorers.scorer_interface import ScorerInterface


class Hypothesis(NamedTuple):
    """Hypothesis data type."""

    yseq: torch.Tensor
    score: Union[float, torch.Tensor] = 0
    scores: Dict[str, Union[float, torch.Tensor]] = dict()
    states: Dict[str, Any] = dict()

    def asdict(self) -> dict:
        """Convert data to JSON-friendly dict."""
        return self._replace(
            yseq=self.yseq.tolist(),
            score=float(self.score),
            scores={k: float(v) for k, v in self.scores.items()},
        )._asdict()


class BeamSearch(torch.nn.Module):
    """Beam search implementation."""

    def __init__(
        self,
        scorers: Dict[str, ScorerInterface],
        weights: Dict[str, float],
        beam_size: int,
        vocab_size: int,
        sos=None,
        eos=None,
        token_list: List[str] = None,
        pre_beam_ratio: float = 1.5,
        pre_beam_score_key: str = None,
    ):
        """Initialize beam search.

        Args:
            scorers (dict[str, ScorerInterface]): Dict of decoder modules
                e.g., Decoder, CTCPrefixScorer, LM
                The scorer will be ignored if it is `None`
            weights (dict[str, float]): Dict of weights for each scorers
                The scorer will be ignored if its weight is 0
            beam_size (int): The number of hypotheses kept during search
            vocab_size (int): The number of vocabulary
            sos (int): Start of sequence id
            eos (int): End of sequence id
            token_list (list[str]): List of tokens for debug log
            pre_beam_score_key (str): key of scores to perform pre-beam search
            pre_beam_ratio (float): beam size in the pre-beam search
                will be `int(pre_beam_ratio * beam_size)`

        """
        super().__init__()
        # set scorers
        self.weights = weights
        self.scorers = dict()
        self.full_scorers = dict()
        self.part_scorers = dict()
        # this module dict is required for recursive cast
        # `self.to(device, dtype)` in `recog.py`
        self.nn_dict = torch.nn.ModuleDict()
        for k, v in scorers.items():
            w = weights.get(k, 0)
            if w == 0 or v is None:
                continue
            # assert isinstance(
            #     v, ScorerInterface
            # ), f"{k} ({type(v)}) does not implement ScorerInterface"
            self.scorers[k] = v
            if isinstance(v, PartialScorerInterface):
                self.part_scorers[k] = v
            else:
                self.full_scorers[k] = v
            if isinstance(v, torch.nn.Module):
                self.nn_dict[k] = v

        # set configurations
        self.sos = sos
        self.eos = eos
        if isinstance(self.eos, (list, tuple)):
            self.eos = eos[0]
        self.token_list = token_list
        self.pre_beam_size = int(pre_beam_ratio * beam_size)
        self.beam_size = beam_size
        self.n_vocab = vocab_size
        if (
            pre_beam_score_key is not None
            and pre_beam_score_key != "full"
            and pre_beam_score_key not in self.full_scorers
        ):
            raise KeyError(f"{pre_beam_score_key} is not found in {self.full_scorers}")
        self.pre_beam_score_key = pre_beam_score_key
        self.do_pre_beam = (
            self.pre_beam_score_key is not None
            and self.pre_beam_size < self.n_vocab
            and len(self.part_scorers) > 0
        )

    def init_hyp(self, x: torch.Tensor) -> List[Hypothesis]:
        """Get an initial hypothesis data.

        Args:
            x (torch.Tensor): The encoder output feature

        Returns:
            Hypothesis: The initial hypothesis.

        """
        init_states = dict()
        init_scores = dict()
        for k, d in self.scorers.items():
            init_states[k] = d.init_state(x)
            init_scores[k] = 0.0
        if not isinstance(self.sos, (list, tuple)):
            self.sos = [self.sos]
        return [
            Hypothesis(
                score=0.0,
                scores=init_scores,
                states=init_states,
                yseq=torch.tensor(self.sos, device=x.device),
            )
        ]

    @staticmethod
    def append_token(xs: torch.Tensor, x: int) -> torch.Tensor:
        """Append new token to prefix tokens.

        Args:
            xs (torch.Tensor): The prefix token
            x (int): The new token to append

        Returns:
            torch.Tensor: New tensor contains: xs + [x] with xs.dtype and xs.device

        """
        x = torch.tensor([x], dtype=xs.dtype, device=xs.device)
        return torch.cat((xs, x))

    def score_full(
        self, hyp: Hypothesis, x: torch.Tensor
    ) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:
        """Score new hypothesis by `self.full_scorers`.

        Args:
            hyp (Hypothesis): Hypothesis with prefix tokens to score
            x (torch.Tensor): Corresponding input feature

        Returns:
            Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: Tuple of
                score dict of `hyp` that has string keys of `self.full_scorers`
                and tensor score values of shape: `(self.n_vocab,)`,
                and state dict that has string keys
                and state values of `self.full_scorers`

        """
        scores = dict()
        states = dict()
        for k, d in self.full_scorers.items():
            scores[k], states[k] = d.score(hyp.yseq, hyp.states[k], x)
        return scores, states

    def score_partial(
        self, hyp: Hypothesis, ids: torch.Tensor, x: torch.Tensor
    ) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]:
        """Score new hypothesis by `self.part_scorers`.

        Args:
            hyp (Hypothesis): Hypothesis with prefix tokens to score
            ids (torch.Tensor): 1D tensor of new partial tokens to score
            x (torch.Tensor): Corresponding input feature

        Returns:
            Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: Tuple of
                score dict of `hyp` that has string keys of `self.part_scorers`
                and tensor score values of shape: `(len(ids),)`,
                and state dict that has string keys
                and state values of `self.part_scorers`

        """
        scores = dict()
        states = dict()
        for k, d in self.part_scorers.items():
            scores[k], states[k] = d.score_partial(hyp.yseq, ids, hyp.states[k], x)
        return scores, states

    def beam(
        self, weighted_scores: torch.Tensor, ids: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute topk full token ids and partial token ids.

        Args:
            weighted_scores (torch.Tensor): The weighted sum scores for each tokens.
            Its shape is `(self.n_vocab,)`.
            ids (torch.Tensor): The partial token ids to compute topk

        Returns:
            Tuple[torch.Tensor, torch.Tensor]:
                The topk full token ids and partial token ids.
                Their shapes are `(self.beam_size,)`

        """
        # no pre beam performed
        if weighted_scores.size(0) == ids.size(0):
            top_ids = weighted_scores.topk(self.beam_size)[1]
            return top_ids, top_ids

        # mask pruned in pre-beam not to select in topk
        tmp = weighted_scores[ids]
        weighted_scores[:] = -float("inf")
        weighted_scores[ids] = tmp
        top_ids = weighted_scores.topk(self.beam_size)[1]
        local_ids = weighted_scores[ids].topk(self.beam_size)[1]
        return top_ids, local_ids

    @staticmethod
    def merge_scores(
        prev_scores: Dict[str, float],
        next_full_scores: Dict[str, torch.Tensor],
        full_idx: int,
        next_part_scores: Dict[str, torch.Tensor],
        part_idx: int,
    ) -> Dict[str, torch.Tensor]:
        """Merge scores for new hypothesis.

        Args:
            prev_scores (Dict[str, float]):
                The previous hypothesis scores by `self.scorers`
            next_full_scores (Dict[str, torch.Tensor]): scores by `self.full_scorers`
            full_idx (int): The next token id for `next_full_scores`
            next_part_scores (Dict[str, torch.Tensor]):
                scores of partial tokens by `self.part_scorers`
            part_idx (int): The new token id for `next_part_scores`

        Returns:
            Dict[str, torch.Tensor]: The new score dict.
                Its keys are names of `self.full_scorers` and `self.part_scorers`.
                Its values are scalar tensors by the scorers.

        """
        new_scores = dict()
        for k, v in next_full_scores.items():
            new_scores[k] = prev_scores[k] + v[full_idx]
        for k, v in next_part_scores.items():
            new_scores[k] = prev_scores[k] + v[part_idx]
        return new_scores

    def merge_states(self, states: Any, part_states: Any, part_idx: int) -> Any:
        """Merge states for new hypothesis.

        Args:
            states: states of `self.full_scorers`
            part_states: states of `self.part_scorers`
            part_idx (int): The new token id for `part_scores`

        Returns:
            Dict[str, torch.Tensor]: The new score dict.
                Its keys are names of `self.full_scorers` and `self.part_scorers`.
                Its values are states of the scorers.

        """
        new_states = dict()
        for k, v in states.items():
            new_states[k] = v
        for k, d in self.part_scorers.items():
            new_states[k] = d.select_state(part_states[k], part_idx)
        return new_states

    def search(self, running_hyps: List[Hypothesis], x: torch.Tensor) -> List[Hypothesis]:
        """Search new tokens for running hypotheses and encoded speech x.

        Args:
            running_hyps (List[Hypothesis]): Running hypotheses on beam
            x (torch.Tensor): Encoded speech feature (T, D)

        Returns:
            List[Hypotheses]: Best sorted hypotheses

        """
        best_hyps = []
        part_ids = torch.arange(self.n_vocab, device=x.device)  # no pre-beam
        for hyp in running_hyps:
            # scoring
            weighted_scores = torch.zeros(self.n_vocab, dtype=x.dtype, device=x.device)
            scores, states = self.score_full(hyp, x)
            for k in self.full_scorers:
                weighted_scores += self.weights[k] * scores[k]
            # partial scoring
            if self.do_pre_beam:
                pre_beam_scores = (
                    weighted_scores
                    if self.pre_beam_score_key == "full"
                    else scores[self.pre_beam_score_key]
                )
                part_ids = torch.topk(pre_beam_scores, self.pre_beam_size)[1]
            part_scores, part_states = self.score_partial(hyp, part_ids, x)
            for k in self.part_scorers:
                weighted_scores[part_ids] += self.weights[k] * part_scores[k]
            # add previous hyp score
            weighted_scores += hyp.score

            # update hyps
            for j, part_j in zip(*self.beam(weighted_scores, part_ids)):
                # will be (2 x beam at most)
                best_hyps.append(
                    Hypothesis(
                        score=weighted_scores[j],
                        yseq=self.append_token(hyp.yseq, j),
                        scores=self.merge_scores(hyp.scores, scores, j, part_scores, part_j),
                        states=self.merge_states(states, part_states, part_j),
                    )
                )

            # sort and prune 2 x beam -> beam
            best_hyps = sorted(best_hyps, key=lambda x: x.score, reverse=True)[
                : min(len(best_hyps), self.beam_size)
            ]
        return best_hyps

    def forward(
        self, x: torch.Tensor, maxlenratio: float = 0.0, minlenratio: float = 0.0
    ) -> List[Hypothesis]:
        """Perform beam search.

        Args:
            x (torch.Tensor): Encoded speech feature (T, D)
            maxlenratio (float): Input length ratio to obtain max output length.
                If maxlenratio=0.0 (default), it uses a end-detect function
                to automatically find maximum hypothesis lengths
                If maxlenratio<0.0, its absolute value is interpreted
                as a constant max output length.
            minlenratio (float): Input length ratio to obtain min output length.

        Returns:
            list[Hypothesis]: N-best decoding results

        """
        # set length bounds
        if maxlenratio == 0:
            maxlen = x.shape[0]
        elif maxlenratio < 0:
            maxlen = -1 * int(maxlenratio)
        else:
            maxlen = max(1, int(maxlenratio * x.size(0)))
        minlen = int(minlenratio * x.size(0))
        logging.info("decoder input length: " + str(x.shape[0]))
        logging.info("max output length: " + str(maxlen))
        logging.info("min output length: " + str(minlen))

        # main loop of prefix search
        running_hyps = self.init_hyp(x)
        ended_hyps = []
        for i in range(maxlen):
            logging.debug("position " + str(i))
            best = self.search(running_hyps, x)
            # post process of one iteration
            running_hyps = self.post_process(i, maxlen, maxlenratio, best, ended_hyps)
            # end detection
            if maxlenratio == 0.0 and end_detect([h.asdict() for h in ended_hyps], i):
                logging.info(f"end detected at {i}")
                break
            if len(running_hyps) == 0:
                logging.info("no hypothesis. Finish decoding.")
                break
            else:
                logging.debug(f"remained hypotheses: {len(running_hyps)}")

        nbest_hyps = sorted(ended_hyps, key=lambda x: x.score, reverse=True)
        # check the number of hypotheses reaching to eos
        if len(nbest_hyps) == 0:
            logging.warning(
                "there is no N-best results, perform recognition " "again with smaller minlenratio."
            )
            return (
                []
                if minlenratio < 0.1
                else self.forward(x, maxlenratio, max(0.0, minlenratio - 0.1))
            )

        # report the best result
        best = nbest_hyps[0]
        for k, v in best.scores.items():
            logging.info(f"{v:6.2f} * {self.weights[k]:3} = {v * self.weights[k]:6.2f} for {k}")
        logging.info(f"total log probability: {best.score:.2f}")
        logging.info(f"normalized log probability: {best.score / len(best.yseq):.2f}")
        logging.info(f"total number of ended hypotheses: {len(nbest_hyps)}")
        if self.token_list is not None:
            logging.info(
                "best hypo: " + "".join([self.token_list[x] for x in best.yseq[1:-1]]) + "\n"
            )
        return nbest_hyps

    def post_process(
        self,
        i: int,
        maxlen: int,
        maxlenratio: float,
        running_hyps: List[Hypothesis],
        ended_hyps: List[Hypothesis],
    ) -> List[Hypothesis]:
        """Perform post-processing of beam search iterations.

        Args:
            i (int): The length of hypothesis tokens.
            maxlen (int): The maximum length of tokens in beam search.
            maxlenratio (int): The maximum length ratio in beam search.
            running_hyps (List[Hypothesis]): The running hypotheses in beam search.
            ended_hyps (List[Hypothesis]): The ended hypotheses in beam search.

        Returns:
            List[Hypothesis]: The new running hypotheses.

        """
        logging.debug(f"the number of running hypotheses: {len(running_hyps)}")
        if self.token_list is not None:
            logging.debug(
                "best hypo: " + "".join([self.token_list[x] for x in running_hyps[0].yseq[1:]])
            )
        # add eos in the final loop to avoid that there are no ended hyps
        if i == maxlen - 1:
            logging.info("adding <eos> in the last position in the loop")
            running_hyps = [
                h._replace(yseq=self.append_token(h.yseq, self.eos)) for h in running_hyps
            ]

        # add ended hypotheses to a final list, and removed them from current hypotheses
        # (this will be a problem, number of hyps < beam)
        remained_hyps = []
        for hyp in running_hyps:
            if hyp.yseq[-1] == self.eos:
                # e.g., Word LM needs to add final <eos> score
                for k, d in chain(self.full_scorers.items(), self.part_scorers.items()):
                    s = d.final_score(hyp.states[k])
                    hyp.scores[k] += s
                    hyp = hyp._replace(score=hyp.score + self.weights[k] * s)
                ended_hyps.append(hyp)
            else:
                remained_hyps.append(hyp)
        return remained_hyps

 funasr/train_utils/trainer.py

@@ -308,6 +308,7 @@
                    checkpoint["step_in_epoch"] if "step_in_epoch" in checkpoint else 0
                )
                self.step_in_epoch = 0 if self.step_in_epoch is None else self.step_in_epoch
                print(checkpoint["train_acc_avg"])
                self.train_acc_avg = (
                    checkpoint["train_acc_avg"] if "train_acc_avg" in checkpoint else 0
                )
@@ -464,7 +465,8 @@
                    batch_num_epoch = len(dataloader_train)
                self.log(
                    epoch,
                    batch_idx + kwargs.get("start_step", 0),
                    batch_idx,
                    log_step=batch_idx + kwargs.get("start_step", 0),
                    step_in_epoch=self.step_in_epoch,
                    batch_num_epoch=batch_num_epoch,
                    lr=lr,
@@ -633,11 +635,12 @@
        tag="train",
        data_split_i=0,
        data_split_num=1,
        log_step=None,
        **kwargs,
    ):

        if (batch_idx + 1) % self.log_interval == 0:

            batch_idx = log_step if log_step is not None else batch_idx
            gpu_info = (
                "GPU, memory: usage: {:.3f} GB, "
                "peak: {:.3f} GB, "