change default enable_auto_mixed_precision from true to false

example/resnet101_imagenet2012/train.py

@@ -48,7 +48,8 @@ args_opt = parser.parse_args()
 
 device_id = int(os.getenv('DEVICE_ID'))
 
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id)
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
+                    enable_auto_mixed_precision=True)
 
 if __name__ == '__main__':
     if not args_opt.do_eval and args_opt.run_distribute:

example/resnet50_cifar10/train.py

@@ -41,8 +41,8 @@ args_opt = parser.parse_args()
 
 device_id = int(os.getenv('DEVICE_ID'))
 
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False)
-context.set_context(device_id=device_id)
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
+                    enable_auto_mixed_precision=True)
 
 if __name__ == '__main__':
     if not args_opt.do_eval and args_opt.run_distribute:

example/resnet50_imagenet2012/train.py

@@ -43,8 +43,9 @@ args_opt = parser.parse_args()
 
 device_id = int(os.getenv('DEVICE_ID'))
 
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False)
-context.set_context(device_id=device_id)
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", save_graphs=False, device_id=device_id,
+                    enable_auto_mixed_precision=True)
 
 if __name__ == '__main__':
     if not args_opt.do_eval and args_opt.run_distribute:

mindspore/ccsrc/utils/context/ms_context.cc

@@ -72,7 +72,7 @@ MsContext::MsContext(const std::string &policy, const std::string &target) {
   enable_mem_reuse_ = true;
   enable_gpu_summary_ = true;
   precompile_only_ = false;
-  auto_mixed_precision_flag_ = true;
+  auto_mixed_precision_flag_ = false;
   enable_pynative_infer_ = false;
   enable_dynamic_mem_pool_ = true;
   graph_memory_max_size_ = "0";