MindSpore报错ValueError: For ‘AvgPool‘ 输出形状的每一维都要大于零

1 报错描述

1.1 系统环境

Hardware Environment(Ascend/GPU/CPU): Ascend
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AvgPool的单算子网络，对输入的多维数据进行二维平均池化运算。脚本如下：

 01 class Net(nn.Cell):
 02     def __init__(self):
 03         super(Net, self).__init__()
 04         self.avgpool_op = ops.AvgPool(pad_mode="VALID", kernel_size=32, strides=1)
 05 
 06     def construct(self, x):
 07         result = self.avgpool_op(x)
 08         return result
 09 
 10 x = Tensor(np.arange(128 * 20 * 32 * 65).reshape(65, 32, 20, 128),mindspore.float32)
 11 net = Net()
 12 output = net(x)
 13 print(output)

1.2.2 报错

这里报错信息如下：

Traceback (most recent call last):
  File "avgpool.py", line 17, in <module>
    output = net(x)
  File "/root/archiconda3/envs/lilinjie_high/lib/python3.7/site-packages/mindspore/nn/cell.py", line 573, in __call__
    out = self.compile_and_run(*args)
  File "/root/archiconda3/envs/lilinjie_high/lib/python3.7/site-packages/mindspore/nn/cell.py", line 956, in compile_and_run
    self.compile(*inputs)
  File "/root/archiconda3/envs/lilinjie_high/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/archiconda3/envs/lilinjie_high/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/archiconda3/envs/lilinjie_high/lib/python3.7/site-packages/mindspore/ops/primitive.py", line 575, in __infer__
    out[track] = fn(*(x[track] for x in args))
  File "/root/archiconda3/envs/lilinjie_high/lib/python3.7/site-packages/mindspore/ops/operations/nn_ops.py", line 1572, in infer_shape
    raise ValueError(f"For '{self.name}', the each element of the output shape must be larger than 0, "
ValueError: For 'AvgPool', the each element of the output shape must be larger than 0, but got output shape: [65, 32, -3, 105]. The input shape: [65, 32, 20, 128], kernel size: (1, 1, 24, 24), strides: (1, 1, 1, 1).Please check the official api documents for more information about the output.

原因分析

我们看报错信息，在ValueError中，写到For’AvgPool’, the each element of the output shape must be larger than 0, but got output shape: [65, 32, -3, 105].意思是输出形状的每一个值都应该大于零， 实际出现了负数， 一般输出数据范围不合理很可能是因为输入没有满足要求导致， 因此我们可以从排查各个输入的有效性着手。仔细查看官网API描述， 再结合我们输入的数据， strides=1在合理范围内， 但是kernel_size=32不合理，文档说明了在Ascend环境下，kernel_size的高度和宽度相乘应小于256, 而32*32=1024。因此我们可以调小kernel_size的高度或者宽度来解决这个问题。

2 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.avgpool_op = ops.AvgPool(pad_mode="VALID", kernel_size=15, strides=1)
 05 
 06     def construct(self, x):
 07         result = self.avgpool_op(x)
 08         return result
 09 
 10 x = Tensor(np.arange(128 * 20 * 32 * 65).reshape(65, 32, 20, 128),mindspore.float32)
 11 net = Net()
 12 output = net(x)
 13 print(output)

此时执行成功，输出如下：

(65, 32, 6, 114)

3 总结

定位报错问题的步骤：

1、找到报错的用户代码行：output = net(x);

2、 根据日志报错信息中的关键字，缩小分析问题的范围* the each element of the output shape must be larger than 0* ;

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 AvgPool算子API接口