when loss div batchsize, change lr, more epoch, loss can reduce more and cer lower than before

examples/aishell/README.md

 # Aishell-1
 
 ## CTC
-| Model | Config | Test set |  CER |
-| --- | --- | --- | --- |
-| DeepSpeech2 | conf/deepspeech2.yaml | test | 0.078977 |
-| DeepSpeech2 | release 1.8.5 | test | 0.080447 |
+| Model | Config | Test Set |  CER | Valid Loss |
+| --- | --- | --- | --- | --- |
+| DeepSpeech2 | conf/deepspeech2.yaml | test | 0.078786 | 7.036566 |
+| DeepSpeech2 | release 1.8.5 | test | 0.080447 | - |

examples/aishell/conf/deepspeech2.yaml

@@ -29,8 +29,8 @@ model:
   use_gru: True 
   share_rnn_weights: False
 training:
-  n_epoch: 30
-  lr: 5e-4
+  n_epoch: 50
+  lr: 2e-3
   lr_decay: 0.83
   weight_decay: 1e-06
   global_grad_clip: 5.0

examples/aishell/local/train.sh

@@ -2,7 +2,7 @@
 
 # train model
 # if you wish to resume from an exists model, uncomment --init_from_pretrained_model
-export FLAGS_sync_nccl_allreduce=0
+#export FLAGS_sync_nccl_allreduce=0
 
 ngpu=$(echo ${CUDA_VISIBLE_DEVICES} | python -c 'import sys; a = sys.stdin.read(); print(len(a.split(",")));')
 echo "using $ngpu gpus..."