From 23bc5dee4e88ef8b5d0c0d64d2e188c054422e8b Mon Sep 17 00:00:00 2001
From: speech_asr <wangjiaming.wjm@alibaba-inc.com>
Date: 星期二, 11 四月 2023 00:21:45 +0800
Subject: [PATCH] update
---
funasr/models/e2e_uni_asr.py | 14 +++++++-------
1 files changed, 7 insertions(+), 7 deletions(-)
diff --git a/funasr/models/e2e_uni_asr.py b/funasr/models/e2e_uni_asr.py
index 03fbca9..0c26b8e 100644
--- a/funasr/models/e2e_uni_asr.py
+++ b/funasr/models/e2e_uni_asr.py
@@ -23,6 +23,7 @@
from funasr.models.postencoder.abs_postencoder import AbsPostEncoder
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.layers.abs_normalize import AbsNormalize
from funasr.torch_utils.device_funcs import force_gatherable
from funasr.train.abs_espnet_model import AbsESPnetModel
@@ -38,7 +39,7 @@
yield
-class UniASR(AbsESPnetModel):
+class UniASR(FunASRModel):
"""
Author: Speech Lab, Alibaba Group, China
"""
@@ -198,16 +199,15 @@
# for data-parallel
text = text[:, : text_lengths.max()]
- speech = speech[:, :speech_lengths.max(), :]
+ speech = speech[:, :speech_lengths.max()]
ind = self.encoder.overlap_chunk_cls.random_choice(self.training, decoding_ind)
- speech_raw = speech.clone().to(speech.device)
# 1. Encoder
if self.enable_maas_finetune:
with torch.no_grad():
- encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
+ speech_raw, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
else:
- encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
+ speech_raw, encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, ind=ind)
intermediate_outs = None
if isinstance(encoder_out, tuple):
@@ -486,7 +486,7 @@
# 3. Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
if self.normalize is not None:
feats, feats_lengths = self.normalize(feats, feats_lengths)
-
+ speech_raw = feats.clone().to(feats.device)
# Pre-encoder, e.g. used for raw input data
if self.preencoder is not None:
feats, feats_lengths = self.preencoder(feats, feats_lengths)
@@ -523,7 +523,7 @@
if intermediate_outs is not None:
return (encoder_out, intermediate_outs), encoder_out_lens
- return encoder_out, encoder_out_lens
+ return speech_raw, encoder_out, encoder_out_lens
def encode2(
self,
--
Gitblit v1.9.1