!9797 Move the set_context operation before init.

From: @linqingke
Reviewed-by: @guoqi1024,@wuxuejian
Signed-off-by: @guoqi1024

model_zoo/official/cv/yolov4/train.py

@@ -112,6 +112,10 @@ args.lr_epochs = list(map(int, args.lr_epochs.split(',')))
 args.data_root = os.path.join(args.data_dir, 'train2017')
 args.annFile = os.path.join(args.data_dir, 'annotations/instances_train2017.json')
 
+device_id = int(os.getenv('DEVICE_ID', '0'))
+context.set_context(mode=context.GRAPH_MODE, enable_auto_mixed_precision=True,
+                    device_target=args.device_target, save_graphs=False, device_id=device_id)
+
 # init distributed
 if args.is_distributed:
     if args.device_target == "Ascend":
@@ -154,9 +158,6 @@ class BuildTrainNetwork(nn.Cell):
 
 
 if __name__ == "__main__":
-    device_id = int(os.getenv('DEVICE_ID', '0'))
-    context.set_context(mode=context.GRAPH_MODE, enable_auto_mixed_precision=True,
-                        device_target=args.device_target, save_graphs=False, device_id=device_id)
     if args.need_profiler:
         profiler = Profiler(output_path=args.outputs_dir, is_detail=True, is_show_op_path=True)