From e24dbdc496debec225414d4d2c760f5775e64f2a Mon Sep 17 00:00:00 2001
From: 天地 <tiandiweizun@gmail.com>
Date: 星期三, 26 三月 2025 13:44:41 +0800
Subject: [PATCH] 感觉应该从文件读取更合适，因为上面判断了文件存在，且可以读取，如果本身是文本的话，下面也会有逻辑进行处理 (#2452)

---
 examples/aishell/branchformer/conf/branchformer_12e_6d_2048_256.yaml |    8 ++++++--
 1 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/examples/aishell/branchformer/conf/branchformer_12e_6d_2048_256.yaml b/examples/aishell/branchformer/conf/branchformer_12e_6d_2048_256.yaml
index aefd2b9..acb7946 100644
--- a/examples/aishell/branchformer/conf/branchformer_12e_6d_2048_256.yaml
+++ b/examples/aishell/branchformer/conf/branchformer_12e_6d_2048_256.yaml
@@ -79,8 +79,9 @@
 train_conf:
   accum_grad: 1
   grad_clip: 5
-  max_epoch: 150
+  max_epoch: 180
   keep_nbest_models: 10
+  avg_keep_nbest_models_type: acc
   log_interval: 50
 
 optim: adam
@@ -96,7 +97,7 @@
     index_ds: IndexDSJsonl
     batch_sampler: EspnetStyleBatchSampler
     batch_type: length # example or length
-    batch_size: 25000 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
+    batch_size: 10000 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
     max_token_length: 2048 # filter samples if source_token_len+target_token_len > max_token_length,
     buffer_size: 1024
     shuffle: True
@@ -116,3 +117,6 @@
     reduce: true
     ignore_nan_grad: true
 normalize: null
+
+beam_size: 10
+decoding_ctc_weight: 0.4
\ No newline at end of file

--
Gitblit v1.9.1