MindSpore报错 For ‘MirrorPad‘, paddings must be a Tensor with *

1 报错描述

1.1 系统环境

Hardware Environment(Ascend/GPU/CPU): CPU
Software 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 基本信息

1.2.1 脚本

训练脚本是通过构建MirrorPad的单算子网络，使用镜像值填充张量。脚本如下：

01  context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
02  class Net(nn.Cell):
03      def __init__(self):
04          super(Net, self).__init__()
05          self.pad = ops.MirrorPad(mode="REFLECT")
06      def construct(self, x, paddings):
07          return self.pad(x, paddings)
08  
09  x = Tensor(np.random.random(size=(2, 3)).astype(np.float32))
10  paddings = Tensor([[1, 1], [2, 2]])
11  pad = Net()
12  output = pad(x, paddings)
13  print(output.shape)

1.2.2 报错

这里报错信息如下：

Traceback (most recent call last):
  File "99553.py", line 20, in <module>
    output = pad(x, paddings)
  File "/root/miniconda3/envs/high_llj/lib/python3.7/site-packages/mindspore/nn/cell.py", line 573, in __call__
    out = self.compile_and_run(*args)
  File "/root/miniconda3/envs/high_llj/lib/python3.7/site-packages/mindspore/nn/cell.py", line 956, in compile_and_run
    self.compile(*inputs)
  File "/root/miniconda3/envs/high_llj/lib/python3.7/site-packages/mindspore/nn/cell.py", line 929, in compile
    _cell_graph_executor.compile(self, *inputs, phase=self.phase, auto_parallel_mode=self._auto_parallel_mode)
  File "/root/miniconda3/envs/high_llj/lib/python3.7/site-packages/mindspore/common/api.py", line 1063, in compile
    result = self._graph_executor.compile(obj, args_list, phase, self._use_vm_mode())
  File "/root/miniconda3/envs/high_llj/lib/python3.7/site-packages/mindspore/ops/operations/nn_ops.py", line 4189, in __infer__
    raise ValueError(f"For '{self.name}', paddings must be a Tensor with type of int64, "
ValueError: For 'MirrorPad', paddings must be a Tensor with type of int64, but got None.

原因分析

我们看报错信息，在ValueError中，写到ValueError: For 'MirrorPad', paddings must be a Tensor with type of int64, but got None.，意思是paddings必须是一个Tensor[INT64], 但是实际得到的是None， 结合脚本第10和12行发现paddings初始化值明明不是None，这是为什么呢。其实这是因为， 在mindspore的图模式下， 常量输入必须在构图的阶段传入， 否则在执行的时候传入就只能得到None。Ps. 在PYNATIVE_MODE不存在这个问题。

2 解决方法

基于上面已知的原因，可以做出两种修改解决此问题。

# 第一种， 在PYNATIVE_MODE下跑脚本：
from mindspore import context

context.set_context(mode=context.PYNATIVE_MODE, device_target="CPU")
class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.pad = ops.MirrorPad(mode="REFLECT")
    def construct(self, x, paddings):
        return self.pad(x, paddings)

x = Tensor(np.random.random(size=(2, 3)).astype(np.float32))
paddings = Tensor([[1, 1], [2, 2]])
pad = Net()
output = pad(x, paddings)
print(output shape: output.shape)

此时执行成功，输出如下：

output shape: (4, 7)

# 第二种，将paddings在构图的时候当作常量传入：
class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.pad = ops.MirrorPad(mode="REFLECT")
        self.paddings = Tensor([[1, 1], [2, 2]])

    def construct(self, x):
        return self.pad(x, self.paddings)

x = Tensor(np.random.random(size=(2, 3)).astype(np.float32))
pad = Net()
output = pad(x)
print("output shape: "output.shape)

此时执行成功，输出如下：

output shape: (4, 7)

3 总结

定位报错问题的步骤：

1、找到报错的用户代码行：output = pad(x, paddings);

2、 ValueError: For 'MirrorPad', paddings must be a Tensor with type of int64, but got None;

3、需要重点关注变量定义、初始化的正确性。

4 参考文档

4.1 MirrorPad算子API接口