MindSpore:图算融合loss出现异常

问题描述：

在Ascend 910A 、mindspore1.1.2 环境下运行图算融合，网络架构是Resnet 50 ，同样的参数条件下，程序正常运行没有任何问题，但是程序中加上（enable_graph_kernel=True） 之后，有时候训练时 loss变为负无穷大，有时候会变成nan，然后都会报错。

具体报错的内容如下：

amp_level: O2
WARNING: 'ControlDepend' is deprecated from version 1.1 and will be removed in a future version, use 'Depend' instead.
epoch: 1 step: 1, loss is 2.2937918
WARNING: 'ControlDepend' is deprecated from version 1.1 and will be removed in a future version, use 'Depend' instead.
epoch: 1 step: 71, loss is 4803.1064
epoch: 1 step: 72, loss is 4034.7324
epoch: 1 step: 142, loss is -5.3169115e+37
epoch: 1 step: 143, loss is 2.3025851
epoch: 1 step: 213, loss is -3.4028235e+38
epoch: 1 step: 214, loss is 2.3025851
[ERROR] RUNTIME(75982)model execute error, retCode=0x91, [the model stream execute failed].