Added gradient clipping instructions

YangFei1990 · Jun 5, 2019 · 600cc22 · 600cc22
1 parent 1c9d4c5
commit 600cc22
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/RESULTS.md b/RESULTS.md
@@ -16,4 +16,8 @@ You can use a improved bounding box regression weight (`cfg.FRCNN.BBOX_REG_WEIGH
 
 ### SyncBN
 
-You can use SyncBN to train with very large batch sizes without getting NaN losses. However, currently the accuracy is generally lower than when using FreezeBN and the throughput is significantly worse.
+You can use SyncBN to train with very large batch sizes without getting NaN losses. However, currently the accuracy is generally lower than when using FreezeBN and the throughput is significantly worse.
+
+### Large batch size
+
+When training in the 32x4 configuration, you will get NaN ~5% of the time if you do not use gradient clipping. To enable gradient clipping, you need to add `TRAIN.GRADIENT_CLIP=1.5` to the config. This has a minor throughput impact, but eliminates NaN runs.