From 798a7396f1851f06f3ab6971b73f48baefff9b0c Mon Sep 17 00:00:00 2001
From: Glenn Jocher <glenn.jocher@ultralytics.com>
Date: Mon, 26 Aug 2019 16:24:19 +0200
Subject: [PATCH] weight_decay fix

---
 train.py | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/train.py b/train.py
index f2c3f4f2..0a56a8c7 100644
--- a/train.py
+++ b/train.py
@@ -261,9 +261,8 @@ def train():
                 print('WARNING: nan loss detected, ending training')
                 return results
 
-            # Divide by accumulation count
-            if accumulate > 1:
-                loss /= accumulate
+            # Scale loss by nominal batch_size of 64
+            loss *= batch_size / 64
 
             # Compute gradient
             if mixed_precision: