From 96bae0153cb04c82d6e7ca7cb9654d55eb987567 Mon Sep 17 00:00:00 2001
From: aky15 <ankeyu.aky@11.17.44.249>
Date: 星期三, 15 三月 2023 17:34:34 +0800
Subject: [PATCH] rnnt bug fix
---
funasr/models_transducer/encoder/blocks/conv_input.py | 9 ++-------
1 files changed, 2 insertions(+), 7 deletions(-)
diff --git a/funasr/models_transducer/encoder/blocks/conv_input.py b/funasr/models_transducer/encoder/blocks/conv_input.py
index 931d0f0..c68c73b 100644
--- a/funasr/models_transducer/encoder/blocks/conv_input.py
+++ b/funasr/models_transducer/encoder/blocks/conv_input.py
@@ -120,7 +120,7 @@
self.create_new_mask = self.create_new_conv2d_mask
self.vgg_like = vgg_like
- self.min_frame_length = 2
+ self.min_frame_length = 7
if output_size is not None:
self.output = torch.nn.Linear(output_proj, output_size)
@@ -218,9 +218,4 @@
: Number of frames before subsampling.
"""
- if self.subsampling_factor > 1:
- if self.vgg_like:
- return ((size * 2) * self.stride_1) + 1
-
- return ((size + 2) * 2) + (self.kernel_2 - 1) * self.stride_2
- return size
+ return size * self.subsampling_factor
--
Gitblit v1.9.1