From 9fcb3cc06b4e324f0913d2f61b89becc2baeef1b Mon Sep 17 00:00:00 2001
From: hnluo <haoneng.lhn@alibaba-inc.com>
Date: 星期一, 11 九月 2023 17:40:03 +0800
Subject: [PATCH] Merge pull request #932 from alibaba-damo-academy/dev_lhn

---
 egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml |   65 +++++++++++++++++++-------------
 1 files changed, 39 insertions(+), 26 deletions(-)

diff --git a/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml b/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml
index 60f796c..a1f27a3 100644
--- a/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml
+++ b/egs/aishell/rnnt/conf/train_conformer_rnnt_unified.yaml
@@ -1,54 +1,57 @@
+encoder: chunk_conformer
 encoder_conf:
-    main_conf:
-      pos_wise_act_type: swish
-      pos_enc_dropout_rate: 0.5
-      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: 0
-    input_conf:
-      block_type: conv2d
-      conv_size: 512
+      left_chunk_size: 1
+      embed_vgg_like: false
       subsampling_factor: 4
-      num_frame: 1
-    body_conf:
-    - block_type: conformer
-      linear_size: 2048
-      hidden_size: 512
-      heads: 8
+      linear_units: 2048
+      output_size: 512
+      attention_heads: 8
       dropout_rate: 0.5
-      pos_wise_dropout_rate: 0.5
-      att_dropout_rate: 0.5
-      conv_mod_kernel_size: 15
+      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: 0.5
     dropout_rate: 0.5
-
 joint_network_conf:
     joint_space_size: 512
 
+# frontend related
+frontend: wav_frontend
+frontend_conf:
+    fs: 16000
+    window: hamming
+    n_mels: 80
+    frame_length: 25
+    frame_shift: 10
+    lfr_m: 1
+    lfr_n: 1
+
+
 # Auxiliary CTC
+model: rnnt_unified
 model_conf:
     auxiliary_ctc_weight: 0.0
 
 # minibatch related
 use_amp: true
-batch_type: unsorted
-batch_size: 16
-num_workers: 16
 
 # optimization related
 accum_grad: 1
 grad_clip: 5
-max_epoch: 200
+max_epoch: 120
 val_scheduler_criterion:
     - valid
     - loss
@@ -65,8 +68,6 @@
 scheduler_conf:
    warmup_steps: 25000
 
-normalize: None
-
 specaug: specaug
 specaug_conf:
     apply_time_warp: true
@@ -83,4 +84,16 @@
     - 50
     num_time_mask: 5
 
+dataset_conf:
+    data_names: speech,text
+    data_types: sound,text
+    shuffle: True
+    shuffle_conf:
+        shuffle_size: 2048
+        sort_size: 500
+    batch_conf:
+        batch_type: token
+        batch_size: 16000
+    num_workers: 8
+
 log_interval: 50

--
Gitblit v1.9.1