fix imagenet nondeterminism when seed is set (pytorch#1056) (pytorch#…

…1057)
mpc001 · Sep 14, 2022 · 32c15f2 · 32c15f2
1 parent d5d9de6
commit 32c15f2
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/imagenet/main.py b/imagenet/main.py
@@ -88,6 +88,7 @@ def main():
         random.seed(args.seed)
         torch.manual_seed(args.seed)
         cudnn.deterministic = True
+        cudnn.benchmark = False
         warnings.warn('You have chosen to seed training. '
                       'This will turn on the CUDNN deterministic setting, '
                       'which can slow down your training considerably! '
@@ -204,7 +205,6 @@ def main_worker(gpu, ngpus_per_node, args):
         else:
             print("=> no checkpoint found at '{}'".format(args.resume))
 
-    cudnn.benchmark = True
 
     # Data loading code
     if args.dummy: