问题现象描述

系统环境：

Hardware Environment(Ascend/GPU/CPU): Ascend Software Environment: -- MindSpore version (source or binary): 1.6.0 -- 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):

训练脚本是通过构建SoftmaxCrossEntropyWithLogits的单算子网络，计算两个变量softmax 交叉熵的例子。脚本如下：

class Net(nn.Cell):
  def __init__(self):
    super(Net, self).__init__()
    self.loss = nn.SoftmaxCrossEntropyWithLogits(sparse=False)

  def construct(self, logits, labels):
    output = self.loss(logits, labels)
    return output

net = Net()
logits = Tensor(np.array([[[2, 4, 1, 4, 5], [2, 1, 2, 4, 3]]]), mindspore.float32)
labels = Tensor(np.array([[[0, 0, 0, 0, 1], [0, 0, 0, 1, 0]]]).astype(np.float32))
out = net(logits, labels)
print('out',out)

报错信息：

Traceback (most recent call last):
File 'demo.py', line 13, in <module>
​    out = net(logits, labels)
…
ValueError: mindspore/core/utils/check_convert_utils.cc:395 CheckInteger] For primitive[SoftmaxCrossEntropyWithLogits], the dimension of logits must be equal to 2, but got 3.
The function call stack (See file ' rank_0/om/analyze_fail.dat' for more details):
\# 0 In file demo.py(04)
​    output = self.loss(logits, labels)

原因分析

在报错信息ValueError中，For primitive[SoftmaxCrossEntropyWithLogits], the dimension of logits must be equal to 2, but got 3，表示传入logits应该等于2维，但实际传入的logits的Shape是3维，查看官网对logits的描述可知，支持的Shape为(N,C)。如下图所示。

对于3维数据，建议先reshape成2维(N*L, C)，然后再调用nn.SoftmaxCrossEntropyWithLogits接口，执行完后再reshape回 (N, 1)。

参考文档：SoftmaxCrossEntropyWithLogits算子API接口

解决措施

基于以上原因，修改训练脚本：

class Net(nn.Cell):
    def __init__(self):
        super(Net, self).__init__()
        self.loss = nn.SoftmaxCrossEntropyWithLogits(sparse=False)

    def construct(self, logits, labels):
        output = self.loss(logits, labels)
        return output

net = Net()
logits = Tensor(np.array([[[2, 4, 1, 4, 5], [2, 1, 2, 4, 3]]]), mindspore.float32)
labels = Tensor(np.array([[[0, 0, 0, 0, 1], [0, 0, 0, 1, 0]]]).astype(np.float32))
L, N, C = logits.shape
logits,labels = logits.reshape(L*N, C),labels.reshape(L*N, C)
out = net(logits, labels)
out = out.reshape(N,1)
print('out',out)

执行训练。成功后输出如下：

out [[0.5899297 ]
[0.52374405]]