From df5f263e5fe3d7961b1aeb3589012400a9905a8f Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期一, 24 四月 2023 16:17:41 +0800
Subject: [PATCH] update
---
funasr/models/e2e_asr_paraformer.py | 102 +++++++++++++++++++++++++++++++++++++++-----------
1 files changed, 79 insertions(+), 23 deletions(-)
diff --git a/funasr/models/e2e_asr_paraformer.py b/funasr/models/e2e_asr_paraformer.py
index b57c8e2..288f469 100644
--- a/funasr/models/e2e_asr_paraformer.py
+++ b/funasr/models/e2e_asr_paraformer.py
@@ -12,24 +12,20 @@
import numpy as np
from typeguard import check_argument_types
-from funasr.layers.abs_normalize import AbsNormalize
from funasr.losses.label_smoothing_loss import (
LabelSmoothingLoss, # noqa: H301
)
from funasr.models.ctc import CTC
from funasr.models.decoder.abs_decoder import AbsDecoder
from funasr.models.e2e_asr_common import ErrorCalculator
-from funasr.models.encoder.abs_encoder import AbsEncoder
-from funasr.models.frontend.abs_frontend import AbsFrontend
from funasr.models.postencoder.abs_postencoder import AbsPostEncoder
from funasr.models.predictor.cif import mae_loss
from funasr.models.preencoder.abs_preencoder import AbsPreEncoder
-from funasr.models.specaug.abs_specaug import AbsSpecAug
+from funasr.models.base_model import FunASRModel
from funasr.modules.add_sos_eos import add_sos_eos
from funasr.modules.nets_utils import make_pad_mask, pad_list
from funasr.modules.nets_utils import th_accuracy
from funasr.torch_utils.device_funcs import force_gatherable
-from funasr.train.abs_espnet_model import AbsESPnetModel
from funasr.models.predictor.cif import CifPredictorV3
@@ -42,7 +38,7 @@
yield
-class Paraformer(AbsESPnetModel):
+class Paraformer(FunASRModel):
"""
Author: Speech Lab, Alibaba Group, China
Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
@@ -53,11 +49,11 @@
self,
vocab_size: int,
token_list: Union[Tuple[str, ...], List[str]],
- frontend: Optional[AbsFrontend],
- specaug: Optional[AbsSpecAug],
- normalize: Optional[AbsNormalize],
+ frontend: Optional[torch.nn.Module],
+ specaug: Optional[torch.nn.Module],
+ normalize: Optional[torch.nn.Module],
preencoder: Optional[AbsPreEncoder],
- encoder: AbsEncoder,
+ encoder: torch.nn.Module,
postencoder: Optional[AbsPostEncoder],
decoder: AbsDecoder,
ctc: CTC,
@@ -620,11 +616,11 @@
self,
vocab_size: int,
token_list: Union[Tuple[str, ...], List[str]],
- frontend: Optional[AbsFrontend],
- specaug: Optional[AbsSpecAug],
- normalize: Optional[AbsNormalize],
+ frontend: Optional[torch.nn.Module],
+ specaug: Optional[torch.nn.Module],
+ normalize: Optional[torch.nn.Module],
preencoder: Optional[AbsPreEncoder],
- encoder: AbsEncoder,
+ encoder: torch.nn.Module,
postencoder: Optional[AbsPostEncoder],
decoder: AbsDecoder,
ctc: CTC,
@@ -898,11 +894,11 @@
self,
vocab_size: int,
token_list: Union[Tuple[str, ...], List[str]],
- frontend: Optional[AbsFrontend],
- specaug: Optional[AbsSpecAug],
- normalize: Optional[AbsNormalize],
+ frontend: Optional[torch.nn.Module],
+ specaug: Optional[torch.nn.Module],
+ normalize: Optional[torch.nn.Module],
preencoder: Optional[AbsPreEncoder],
- encoder: AbsEncoder,
+ encoder: torch.nn.Module,
postencoder: Optional[AbsPostEncoder],
decoder: AbsDecoder,
ctc: CTC,
@@ -1025,16 +1021,76 @@
# 1. Encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
+ intermediate_outs = None
+ if isinstance(encoder_out, tuple):
+ intermediate_outs = encoder_out[1]
+ encoder_out = encoder_out[0]
+ loss_att, acc_att, cer_att, wer_att = None, None, None, None
+ loss_ctc, cer_ctc = None, None
+ loss_pre = None
stats = dict()
+
+ # 1. CTC branch
+ if self.ctc_weight != 0.0:
+ loss_ctc, cer_ctc = self._calc_ctc_loss(
+ encoder_out, encoder_out_lens, text, text_lengths
+ )
+
+ # Collect CTC branch stats
+ stats["loss_ctc"] = loss_ctc.detach() if loss_ctc is not None else None
+ stats["cer_ctc"] = cer_ctc
+
+ # Intermediate CTC (optional)
+ loss_interctc = 0.0
+ if self.interctc_weight != 0.0 and intermediate_outs is not None:
+ for layer_idx, intermediate_out in intermediate_outs:
+ # we assume intermediate_out has the same length & padding
+ # as those of encoder_out
+ loss_ic, cer_ic = self._calc_ctc_loss(
+ intermediate_out, encoder_out_lens, text, text_lengths
+ )
+ loss_interctc = loss_interctc + loss_ic
+
+ # Collect Intermedaite CTC stats
+ stats["loss_interctc_layer{}".format(layer_idx)] = (
+ loss_ic.detach() if loss_ic is not None else None
+ )
+ stats["cer_interctc_layer{}".format(layer_idx)] = cer_ic
+
+ loss_interctc = loss_interctc / len(intermediate_outs)
+
+ # calculate whole encoder loss
+ loss_ctc = (
+ 1 - self.interctc_weight
+ ) * loss_ctc + self.interctc_weight * loss_interctc
+
+ # 2b. Attention decoder branch
+ if self.ctc_weight != 1.0:
+ loss_att, acc_att, cer_att, wer_att, loss_pre = self._calc_att_loss(
+ encoder_out, encoder_out_lens, text, text_lengths
+ )
loss_pre2 = self._calc_pre2_loss(
encoder_out, encoder_out_lens, text, text_lengths
)
- loss = loss_pre2
+ # 3. CTC-Att loss definition
+ if self.ctc_weight == 0.0:
+ loss = loss_att + loss_pre * self.predictor_weight + loss_pre2 * self.predictor_weight * 0.5
+ elif self.ctc_weight == 1.0:
+ loss = loss_ctc
+ else:
+ loss = self.ctc_weight * loss_ctc + (1 - self.ctc_weight) * loss_att + loss_pre * self.predictor_weight + loss_pre2 * self.predictor_weight * 0.5
+ # Collect Attn branch stats
+ stats["loss_att"] = loss_att.detach() if loss_att is not None else None
+ stats["acc"] = acc_att
+ stats["cer"] = cer_att
+ stats["wer"] = wer_att
+ stats["loss_pre"] = loss_pre.detach().cpu() if loss_pre is not None else None
stats["loss_pre2"] = loss_pre2.detach().cpu()
+
stats["loss"] = torch.clone(loss.detach())
# force_gatherable: to-device and to-tensor if scalar for DataParallel
@@ -1051,11 +1107,11 @@
self,
vocab_size: int,
token_list: Union[Tuple[str, ...], List[str]],
- frontend: Optional[AbsFrontend],
- specaug: Optional[AbsSpecAug],
- normalize: Optional[AbsNormalize],
+ frontend: Optional[torch.nn.Module],
+ specaug: Optional[torch.nn.Module],
+ normalize: Optional[torch.nn.Module],
preencoder: Optional[AbsPreEncoder],
- encoder: AbsEncoder,
+ encoder: torch.nn.Module,
postencoder: Optional[AbsPostEncoder],
decoder: AbsDecoder,
ctc: CTC,
--
Gitblit v1.9.1