diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
index a1d0061ef..8da246c7c 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py
@@ -858,7 +858,8 @@ def train_one_epoch(
             # NOTE: We use reduction==sum and loss is computed over utterances
             # in the batch and there is no normalization to it so far.
             scaler.scale(loss).backward()
-            set_batch_count(model, params.batch_idx_train)
+            if params.batch_idx_train % 10 == 0:
+                set_batch_count(model, params.batch_idx_train)
             scheduler.step_batch(params.batch_idx_train)
 
             scaler.step(optimizer)