Dev gzf exp (#1705)

zhifu gao

2024-05-08 a7bc099548576d6d855a3837523f51cadd0e8910

Dev gzf exp (#1705)

* 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

* decoding

* wechat

* decoding key

* decoding key

* decoding key

* decoding key

* decoding key

* Gcf (#1704)

* 添加富文本解码约束

* special token

* bug fix

* fix

---------

Co-authored-by: 常材 <gaochangfeng.gcf@alibaba-inc.com>

* decoding key

---------

Co-authored-by: 常材 <gaochangfeng.gcf@alibaba-inc.com>

 docs/images/wechat.png

 funasr/models/sense_voice/model.py

@@ -514,6 +514,20 @@
        self.beam_search.sos = sos_int
        self.beam_search.eos = eos_int[0]

        # Paramterts for rich decoding
        self.beam_search.emo_unk = tokenizer.encode(
            DecodingOptions.get("emo_unk_token", "<|SPECIAL_TOKEN_1|>"), allowed_special="all")[0]
        self.beam_search.emo_unk_score = 1
        self.beam_search.emo_tokens = tokenizer.encode(
            DecodingOptions.get("emo_target_tokens", "<|HAPPY|><|SAD|><|ANGRY|>"), allowed_special="all")
        self.beam_search.emo_scores = DecodingOptions.get("emo_target_threshold", [0.1, 0.1, 0.1])

        self.beam_search.event_bg_token = tokenizer.encode(
            DecodingOptions.get("gain_tokens_bg", "<|Speech|><|BGM|><|Applause|><|Laughter|>"), allowed_special="all")
        self.beam_search.event_ed_token = tokenizer.encode(
            DecodingOptions.get("gain_tokens_ed", "<|/Speech|><|/BGM|><|/Applause|><|/Laughter|>"), allowed_special="all")
        self.beam_search.event_score_ga = DecodingOptions.get("gain_tokens_score", [1, 1, 1, 1])

        encoder_out, encoder_out_lens = self.encode(
            speech[None, :, :].permute(0, 2, 1), speech_lengths
        )
@@ -843,6 +857,20 @@
        self.beam_search.sos = sos_int
        self.beam_search.eos = eos_int[0]

        # Paramterts for rich decoding
        self.beam_search.emo_unk = tokenizer.encode(
            DecodingOptions.get("emo_unk_token", "<|SPECIAL_TOKEN_1|>"), allowed_special="all")[0]
        self.beam_search.emo_unk_score = 1
        self.beam_search.emo_tokens = tokenizer.encode(
            DecodingOptions.get("emo_target_tokens", "<|HAPPY|><|SAD|><|ANGRY|>"), allowed_special="all")
        self.beam_search.emo_scores = DecodingOptions.get("emo_target_threshold", [0.1, 0.1, 0.1])

        self.beam_search.event_bg_token = tokenizer.encode(
            DecodingOptions.get("gain_tokens_bg", "<|Speech|><|BGM|><|Applause|><|Laughter|>"), allowed_special="all")
        self.beam_search.event_ed_token = tokenizer.encode(
            DecodingOptions.get("gain_tokens_ed", "<|/Speech|><|/BGM|><|/Applause|><|/Laughter|>"), allowed_special="all")
        self.beam_search.event_score_ga = DecodingOptions.get("gain_tokens_score", [1, 1, 1, 1])

        encoder_out, encoder_out_lens = self.encode(
            speech[None, :, :].permute(0, 2, 1), speech_lengths
        )

 funasr/models/sense_voice/search.py

@@ -1,4 +1,5 @@
from itertools import chain
from dataclasses import field
import logging
from typing import Any
from typing import Dict
@@ -8,6 +9,7 @@
from typing import Union

import torch
import numpy as np

from funasr.metrics.common import end_detect
from funasr.models.transformer.scorers.scorer_interface import PartialScorerInterface
@@ -42,6 +44,17 @@
        vocab_size: int,
        sos=None,
        eos=None,
        # NOTE add rich decoding parameters
        # [SPECIAL_TOKEN_1, HAPPY, SAD, ANGRY, NEUTRAL]
        emo_unk: int = 58964,
        emo_unk_score: float = 1.0,
        emo_tokens: List[int] = field(default_factory=lambda: [58954, 58955, 58956, 58957]),
        emo_scores: List[float] = field(default_factory=lambda: [0.1, 0.1, 0.1, 0.1]),
        # [Speech, BGM, Laughter, Applause]
        event_bg_token: List[int] = field(default_factory=lambda: [58946, 58948, 58950, 58952]),
        event_ed_token: List[int] = field(default_factory=lambda: [58947, 58949, 58951, 58953]),
        event_score_ga: List[float] = field(default_factory=lambda: [1, 1, 5, 25]),

        token_list: List[str] = None,
        pre_beam_ratio: float = 1.5,
        pre_beam_score_key: str = None,
@@ -110,6 +123,14 @@
            and len(self.part_scorers) > 0
        )

        self.emo_unk = emo_unk
        self.emo_unk_score = emo_unk_score
        self.emo_tokens = emo_tokens
        self.emo_scores = emo_scores
        self.event_bg_token = event_bg_token
        self.event_ed_token = event_ed_token
        self.event_score_ga = event_score_ga

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

@@ -170,10 +191,48 @@
        """
        scores = dict()
        states = dict()

        def get_score(yseq, sp1, sp2):
            score = [0, 0]
            last_token = yseq[-1]
            last_token2 = yseq[-2] if len(yseq) > 1 else yseq[-1]
            sum_sp1 = sum([1 if x == sp1 else 0 for x in yseq])
            sum_sp2 = sum([1 if x == sp2 else 0 for x in yseq])
            if sum_sp1 > sum_sp2 or last_token in [sp1, sp2]:
                score[0] = -np.inf
            if sum_sp2 >= sum_sp1:
                score[1] = -np.inf
            return score

        def struct_score(yseq, score):
            import math

            last_token = yseq[-1]
            if last_token in self.emo_tokens + [self.emo_unk]:
                # prevent output event after emotation token 
                score[self.event_bg_token] = -np.inf

            for eve_bg, eve_ed, eve_ga in zip(self.event_bg_token, self.event_ed_token, self.event_score_ga):
                score_offset = get_score(yseq, eve_bg, eve_ed)
                score[eve_bg] += score_offset[0]
                score[eve_ed] += score_offset[1]
                score[eve_bg] += math.log(eve_ga)


            score[self.emo_unk] += math.log(self.emo_unk_score)
            for emo, emo_th in zip(self.emo_tokens, self.emo_scores):
                if score.argmax() == emo and score[emo] < math.log(emo_th):
                    score[self.emo_unk] = max(score[emo], score[self.emo_unk])
                    score[emo] = -np.inf
            return score

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

        return scores, states


    def score_partial(
        self, hyp: Hypothesis, ids: torch.Tensor, x: torch.Tensor
    ) -> Tuple[Dict[str, torch.Tensor], Dict[str, Any]]: