From 6f2128a511c4fb6ebc0a056ed63a416d07e24145 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Piotr=20=C5=BBelasko?= <petezor@gmail.com>
Date: Sat, 12 Aug 2023 20:46:25 -0400
Subject: [PATCH] Speed up yesno training to finish in ~10s on CPU

Setting `persistent_workers=True` when creating DataLoader removes a significant overhead at the start of each epoch and validation.
---
 egs/yesno/ASR/tdnn/asr_datamodule.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/egs/yesno/ASR/tdnn/asr_datamodule.py b/egs/yesno/ASR/tdnn/asr_datamodule.py
index 3c1682fa1..ada8c1a6c 100644
--- a/egs/yesno/ASR/tdnn/asr_datamodule.py
+++ b/egs/yesno/ASR/tdnn/asr_datamodule.py
@@ -209,7 +209,7 @@ class YesNoAsrDataModule(DataModule):
             sampler=train_sampler,
             batch_size=None,
             num_workers=self.args.num_workers,
-            persistent_workers=False,
+            persistent_workers=True,
         )
 
         return train_dl
@@ -236,6 +236,7 @@ class YesNoAsrDataModule(DataModule):
             batch_size=None,
             sampler=sampler,
             num_workers=self.args.num_workers,
+            persistent_workers=True,
         )
         return test_dl