1 报错描述
1.1 系统环境
Hardware Environment(Ascend/GPU/CPU): GPU
Software Environment:
MindSpore version (source or binary): 1.5.2
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):
2 基本信息
2.1 脚本
训练脚本是通过构建BatchNorm单算子网络，对Tensor做归一化处理。脚本如下：
01 class Net(nn.Cell):
02 def __init__(self):
03 super(Net, self).__init__()
04 self.batch_norm = ops.BatchNorm()
05 def construct(self,input_x, scale, bias, mean, variance):
06 output = self.batch_norm(input_x, scale, bias, mean, variance)
07 return output
08
09 net = Net()
10 input_x = Tensor(np.ones([2, 2]), mindspore.float16)
11 scale = Tensor(np.ones([2]), mindspore.float16)
12 bias = Tensor(np.ones([2]), mindspore.float16)
13 bias = Tensor(np.ones([2]), mindspore.float16)
14 mean = Tensor(np.ones([2]), mindspore.float16)
15 variance = Tensor(np.ones([2]), mindspore.float16)
16 output = net(input_x, scale, bias, mean, variance)
17 print(output)
2.2 报错
这里报错信息如下：
Traceback (most recent call last):
File "116945.py", line 22, in <module>
output = net(input_x, scale, bias, mean, variance)
File "/data2/llj/mindspores/r1.5/build/package/mindspore/nn/cell.py", line 407, in __call__
out = self.compile_and_run(*inputs)
File "/data2/llj/mindspores/r1.5/build/package/mindspore/nn/cell.py", line 734, in compile_and_run
self.compile(*inputs)
File "/data2/llj/mindspores/r1.5/build/package/mindspore/nn/cell.py", line 721, in compile
_cell_graph_executor.compile(self, *inputs, phase=self.phase, auto_parallel_mode=self._auto_parallel_mode)
File "/data2/llj/mindspores/r1.5/build/package/mindspore/common/api.py", line 551, in compile
result = self._graph_executor.compile(obj, args_list, phase, use_vm, self.queue_name)
TypeError: mindspore/ccsrc/runtime/device/gpu/kernel_info_setter.cc:355 PrintUnsupportedTypeException] Select GPU kernel op[BatchNorm] fail! Incompatible data type!
The supported data types are in[float32 float32 float32 float32 float32], out[float32 float32 float32 float32 float32]; in[float16 float32 float32 float32 float32], out[float16 float32 float32 float32 float32]; , but get in [float16 float16 float16 float16 float16 ] out [float16 float16 float16 float16 float16 ]
3 原因分析
我们看报错信息，在TypeError中，写到Select GPU kernel op[BatchNorm] fail! Incompatible data type! The supported data types are in[float32 float32 float32 float32 float32], out[float32 float32 float32 float32 float32]; in[float16 float32 float32 float32 float32], out[float16 float32 float32 float32 float32]; , but get in [float16 float16 float16 float16 float16 ] out [float16 float16 float16 float16 float16 ]，大概意思是GPU环境下， 不支持当前输入的数据类型组合， 并说明了支持的数据类型组合是怎样的：全部为float32或者input_x为float16， 其余为float32。检查脚本的输入发现全部为float16类型， 因此报错。
4 解决方法
基于上面已知的原因，很容易做出如下修改：
01 class Net(nn.Cell):
02 def __init__(self):
03 super(Net, self).__init__()
04 self.batch_norm = ops.BatchNorm()
05 def construct(self,input_x, scale, bias, mean, variance):
06 output = self.batch_norm(input_x, scale, bias, mean, variance)
07 return output
08
09 net = Net()
10 input_x = Tensor(np.ones([2, 2]), mindspore.float16)
11 scale = Tensor(np.ones([2]), mindspore.float32)
12 bias = Tensor(np.ones([2]), mindspore.float32)
13 mean = Tensor(np.ones([2]), mindspore.float32)
14 variance = Tensor(np.ones([2]), mindspore.float32)
15
16 output = net(input_x, scale, bias, mean, variance)
17 print(output)
此时执行成功，输出如下：
output: (Tensor(shape=[2, 2], dtype=Float16, value=
[[ 1.0000e+00, 1.0000e+00],
[ 1.0000e+00, 1.0000e+00]]), Tensor(shape=[2], dtype=Float32, value= [ 0.00000000e+00, 0.00000000e+00]), Tensor(shape=[2], dtype=Float32, value= [ 0.00000000e+00, 0.00000000e+00]), Tensor(shape=[2], dtype=Float32, value= [ 0.00000000e+00, 0.00000000e+00]), Tensor(shape=[2], dtype=Float32, value= [ 0.00000000e+00, 0.00000000e+00]))
5 总结
定位报错问题的步骤：
1、找到报错的用户代码行： 16 output = net(input_x, scale, bias, mean, variance);
2、 根据日志报错信息中的关键字，缩小分析问题的范围：The supported data types are in[float32 float32 float32 float32 float32], out[float32 float32 float32 float32 float32]; in[float16 float32 float32 float32 float32], out[float16 float32 float32 float32 float32]; , but get in [float16 float16 float16 float16 float16 ] out [float16 float16 float16 float16 float16 ]
3、需要重点关注变量定义、初始化的正确性。