From e1ba6bc138b4e73875c64f35f98f3b15a0560e92 Mon Sep 17 00:00:00 2001
From: 嘉渊 <wangjiaming.wjm@alibaba-inc.com>
Date: 星期三, 17 五月 2023 15:16:06 +0800
Subject: [PATCH] Merge branch 'dev_infer' of https://github.com/alibaba/FunASR into dev_infer

---
 egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml |   56 ++++++++++++++++++++++++++------------------------------
 1 files changed, 26 insertions(+), 30 deletions(-)

diff --git a/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml b/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml
index ef37b97..8a1c40c 100644
--- a/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml
+++ b/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml
@@ -1,36 +1,30 @@
+encoder: chunk_conformer
 encoder_conf:
-    main_conf:
-      pos_wise_act_type: swish
-      pos_enc_dropout_rate: 0.3
-      conv_mod_act_type: swish
+      activation_type: swish
+      positional_dropout_rate: 0.5
       time_reduction_factor: 2
       unified_model_training: true
       default_chunk_size: 16
       jitter_range: 4
-      left_chunk_size: 1
-    input_conf:
-      block_type: conv2d
-      conv_size: 512
+      left_chunk_size: 0
+      embed_vgg_like: false
       subsampling_factor: 4
-      num_frame: 1
-    body_conf:
-    - block_type: conformer
-      linear_size: 2048
-      hidden_size: 512
-      heads: 8
-      dropout_rate: 0.3
-      pos_wise_dropout_rate: 0.3
-      att_dropout_rate: 0.3
-      conv_mod_kernel_size: 15
+      linear_units: 2048
+      output_size: 512
+      attention_heads: 8
+      dropout_rate: 0.5
+      positional_dropout_rate: 0.5
+      attention_dropout_rate: 0.5
+      cnn_module_kernel: 15
       num_blocks: 12    
 
 # decoder related
-decoder: rnn
-decoder_conf:
+rnnt_decoder: rnnt
+rnnt_decoder_conf:
     embed_size: 512
     hidden_size: 512
-    embed_dropout_rate: 0.2
-    dropout_rate: 0.1
+    embed_dropout_rate: 0.5
+    dropout_rate: 0.5
 
 joint_network_conf:
     joint_space_size: 512
@@ -41,14 +35,14 @@
 
 # minibatch related
 use_amp: true
-batch_type: numel
-batch_bins: 1600000
+batch_type: unsorted
+batch_size: 16
 num_workers: 16
 
 # optimization related
 accum_grad: 1
 grad_clip: 5
-max_epoch: 80
+max_epoch: 200
 val_scheduler_criterion:
     - valid
     - loss
@@ -56,11 +50,11 @@
 -   - valid
     - cer_transducer_chunk
     - min
-keep_nbest_models: 5
+keep_nbest_models: 10
 
 optim: adam
 optim_conf:
-   lr: 0.0003
+   lr: 0.001
 scheduler: warmuplr
 scheduler_conf:
    warmup_steps: 25000
@@ -75,10 +69,12 @@
     apply_freq_mask: true
     freq_mask_width_range:
     - 0
-    - 30
+    - 40
     num_freq_mask: 2
     apply_time_mask: true
     time_mask_width_range:
     - 0
-    - 40
-    num_time_mask: 2
+    - 50
+    num_time_mask: 5
+
+log_interval: 50

--
Gitblit v1.9.1