From 6427c834dfd97b1f05c6659cdc7ccf010bf82fe1 Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期一, 24 四月 2023 19:50:07 +0800
Subject: [PATCH] update
---
funasr/models/encoder/conformer_encoder.py | 13 ++++++++++---
1 files changed, 10 insertions(+), 3 deletions(-)
diff --git a/funasr/models/encoder/conformer_encoder.py b/funasr/models/encoder/conformer_encoder.py
index 2df2ba6..e649eca 100644
--- a/funasr/models/encoder/conformer_encoder.py
+++ b/funasr/models/encoder/conformer_encoder.py
@@ -14,7 +14,6 @@
from typeguard import check_argument_types
from funasr.models.ctc import CTC
-from funasr.models.encoder.abs_encoder import AbsEncoder
from funasr.modules.attention import (
MultiHeadedAttention, # noqa: H301
RelPositionMultiHeadedAttention, # noqa: H301
@@ -41,7 +40,7 @@
from funasr.modules.subsampling import Conv2dSubsampling8
from funasr.modules.subsampling import TooShortUttError
from funasr.modules.subsampling import check_short_utt
-
+from funasr.modules.subsampling import Conv2dSubsamplingPad
class ConvolutionModule(nn.Module):
"""ConvolutionModule in Conformer model.
@@ -277,7 +276,7 @@
return x, mask
-class ConformerEncoder(AbsEncoder):
+class ConformerEncoder(torch.nn.Module):
"""Conformer encoder module.
Args:
@@ -381,6 +380,13 @@
)
elif input_layer == "conv2d":
self.embed = Conv2dSubsampling(
+ input_size,
+ output_size,
+ dropout_rate,
+ pos_enc_class(output_size, positional_dropout_rate),
+ )
+ elif input_layer == "conv2dpad":
+ self.embed = Conv2dSubsamplingPad(
input_size,
output_size,
dropout_rate,
@@ -546,6 +552,7 @@
or isinstance(self.embed, Conv2dSubsampling2)
or isinstance(self.embed, Conv2dSubsampling6)
or isinstance(self.embed, Conv2dSubsampling8)
+ or isinstance(self.embed, Conv2dSubsamplingPad)
):
short_status, limit_size = check_short_utt(self.embed, xs_pad.size(1))
if short_status:
--
Gitblit v1.9.1