[mindspore ascend] [user defined operator] graph_ In mode, customize how to traverse tensor

Because you need to use belt weights Of SoftmaxCrossEntropyLoss, It's just mindspore Provided nn.SoftmaxCrossEntropyLoss An operator is customized on .

In this operator , I need to traverse one dimension Tensor. stay PYNATIVE In the pattern , I will Tensor To numpy Array implements this operation . But in GRAPH_MODE There is no way to Tensor To numpy or list, How to operate ？

【 Related codes 】

# labels_int: (n,) dimension Tensor
# self.ignore_label: int
# self.cls_weight: (c, ) dimension Tensor
weights_np = np.ones((labels_int.shape[0]))
labels_np = labels_int.asnumpy()
weights_np[labels_np == self.ignore_label] = 0
cls_weight_np = self.cls_weight.asnumpy()
for idx, v in enumerate(cls_weight_np, 0):
    weights_np[labels_np == idx] *= v

Your problem is that you want to be in graph mode , Traverse one dimension Tensor, Currently, there are two ways to traverse Tensor.

1. The first way to get Tensor The length of the first dimension , Then traverse through the integer index , as follows ：

   for i in range(x.shape[0]):
   
       x[i]

2. The second way is through enumerate Interface , Direct traversal , as follows ：

   for i, ele in enumerate(x, 0):
   
       ele