共计 4010 个字符，预计需要花费 11 分钟才能阅读完成。
1 报错形容 1.1 零碎环境 Hardware Environment(Ascend/GPU/CPU): AscendSoftware Environment:– MindSpore version (source or binary): 1.8.0– Python version (e.g., Python 3.7.5): 3.7.6– OS platform and distribution (e.g., Linux Ubuntu 16.04): Ubuntu 4.15.0-74-generic– GCC/Compiler version (if compiled from source):1.2 根本信息 1.2.1 脚本训练脚本是通过构建 ReduceMean 算子网络，对 axis1 进行求平均值归约。脚本如下：01 class Net(nn.Cell):
02 def __init__(self, axis, keep_dims):
03 super().__init__()
04 self.reducemean = ops.ReduceMean(keep_dims=keep_dims)
05 self.axis = axis
06 def construct(self, input_x):
07 return self.reducemean(input_x, self.axis)
08 net = Net(axis=(1,), keep_dims=True)
09 x = Tensor(np.random.randn(1, 2, 3, 4, 5, 6, 7, 8, 9), mindspore.float32)
10 out = net(x)
11 print(“out shape: “, out.shape)
1.2.2 报错这里报错信息如下：Traceback (most recent call last):
File “test.py”, line 18, in <module>
out = net(x)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/mindspore/nn/cell.py”, line 574, in call
out = self.compile_and_run(*args)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/mindspore/nn/cell.py”, line 975, in compile_and_run
self.compile(*inputs)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/mindspore/nn/cell.py”, line 948, in compile
jit_config_dict=self._jit_config_dict)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/mindspore/common/api.py”, line 1092, in compile
result = self._graph_executor.compile(obj, args_list, phase, self._use_vm_mode())RuntimeError: Single op compile failed, op: reduce_mean_d_1629966128061146056_6
except_msg: 2022-07-15 01:36:29.720449: Query except_msg:Traceback (most recent call last):
File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/te_fusion/parallel_compilation.py”, line 1469, in run
relation_param=self._relation_param)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/te_fusion/fusion_manager.py”, line 1283, in build_single_op
compile_info = call_op()File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/te_fusion/fusion_manager.py”, line 1270, in call_op
opfunc(*inputs, *outputs, *new_attrs, **kwargs)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/tbe/common/utils/para_check.py”, line 537, in _in_wrapper
formal_parameter_list[i][1], op_name)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/tbe/common/utils/para_check.py”, line 516, in _check_one_op_param
_check_input(op_param, param_name, param_type, op_name)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/tbe/common/utils/para_check.py”, line 299, in _check_input
_check_input_output_dict(op_param, param_name, op_name)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/tbe/common/utils/para_check.py”, line 223, in _check_input_output_dict
param_name=param_name)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/tbe/common/utils/para_check.py”, line 689, in check_shape
_check_shape_range(max_rank, min_rank, param_name, shape)File “/root/archiconda3/envs/lh37_ascend/lib/python3.7/site-packages/tbe/common/utils/para_check.py”, line 727, in _check_shape_range
% (error_info['param_name'], min_rank, max_rank, len(shape)))RuntimeError: ({‘errCode’: ‘E80012’, ‘op_name’: ‘reduce_mean_d’, ‘param_name’: ‘input_x’, ‘min_value’: 0, ‘max_value’: 8, ‘real_value’: 9}, ‘In op, the num of dimensions of input/output[input_x] should be inthe range of [0, 8], but actually is [9].’)
起因剖析咱们看报错信息，在 RuntimeError 中，’In op, the num of dimensions of input/output[input_x] should be inthe range of [0, 8], but actually is [9].’ 意思是 ReduceMean 的输出为维度应该大于等于 0，小于等于 8，但理论值为 9，显然超过了 ReduceMean 算子反对的维度。在官网中对 ReduceSum 也做了输出维度限度阐明：
2 解决办法基于下面已知的起因，很容易做出如下批改：01 class Net(nn.Cell):
02 def __init__(self, axis, keep_dims):
03 super().__init__()
04 self.reducemean = ops.ReduceMean(keep_dims=keep_dims)
05 self.axis = axis
06 def construct(self, input_x):
07 return self.reducemean(input_x, self.axis)
08 net = Net(axis=(1,), keep_dims=True)
09 x = Tensor(np.random.randn(2, 3, 4, 5, 6, 7, 8, 9), mindspore.float32)
10 out = net(x)
11 print(“out shape: “, out.shape)
此时执行胜利，输入如下：out shape: (2, 1, 4, 5, 6, 7, 8, 9)
3 总结定位报错问题的步骤：1、找到报错的用户代码行：out = net(x)；2、依据日志报错信息中的关键字，放大剖析问题的范畴：should be in the range of [0, 8], but actually is [10];3、须要重点关注变量定义、初始化的正确性。4 参考文档 4.1 ReduceMean 算子 API 接口