From f151c4612cb722fec21d7ef6ccc50b04e9f9c4a6 Mon Sep 17 00:00:00 2001
From: "sean.narenthiran" <sean.narenthiran@digitalreasoning.com>
Date: Thu, 20 Feb 2020 16:45:31 +0000
Subject: [PATCH] Set loss-scale to 1 to prevent using dynamic scaling

---
 train.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/train.py b/train.py
index 83e618ff..18c07156 100644
--- a/train.py
+++ b/train.py
@@ -80,7 +80,8 @@
 parser.add_argument('--seed', default=123456, type=int, help='Seed to generators')
 parser.add_argument('--opt-level', type=str)
 parser.add_argument('--keep-batchnorm-fp32', type=str, default=None)
-parser.add_argument('--loss-scale', type=str, default=None)
+parser.add_argument('--loss-scale', default=1,
+                    help='Loss scaling used by Apex. Default is 1 due to warp-ctc not supporting scaling of gradients')
 
 torch.manual_seed(123456)
 torch.cuda.manual_seed_all(123456)