pytorch网络转libtorch常见问题


一、All inputs of range must be ints, found Tensor in argument 0:

问题
参数类型不正确,函数的默认参数是tensor

解决措施
函数传入参数不是tensor需要注明类型
我的问题是传入参数npoint是一个int类型,没有注明会报错,更改如下:

def test(npoint):
  ...

更改为

def test(npoint: int):
  ...

二、Sliced expression not yet supported for subscripted assignment. File a bug if you want this:

问题
不支持赋值给切片表达式

解决措施
根据自己需求,进行修改,可利用循环替代

我将view_shape[1:] = [1] * (len(view_shape) - 1)更改为

    for i in range(1, len(view_shape)):
        view_shape[i] = 1

三、Tried to access nonexistent attribute or method 'len' of type 'torch.torch.nn.modules.container.ModuleList'. Did you forget to initialize an attribute in init()?

问题
forward函数中好像不支持len(nn.ModuleList())和下标访问

解决措施
如果是一个ModuleList()可以用enumerate函数,多个同维度的可以用zip函数

我这里有两个ModuleList(),所以采用zip函数,更改如下:

   for i, conv in enumerate(self.mlp_convs):
      bn = self.mlp_bns[i]
      new_points = F.relu(bn(conv(new_points)))

更改为

    for conv, bn in zip(self.mlp_convs, self.mlp_bns):
        new_points = F.relu(bn(conv(new_points)))

ref: https://github.com/pytorch/pytorch/issues/16123


四、Expected integer literal for index

问题和解决方法类似第三个


五、Arguments for call are not valid. The following variants are available

Expected a value of type 'List[Tensor]' for argument 'indices' but instead found type 'List[Optional[Tensor]]'

问题
赋值类型不对,需求是tensor,但给的是int

解决措施

  • 方法1
    int类型的数Ntorch.tensor(N)代替
mask = sqrdists > radius ** 2
group_idx[mask] = N

变为

mask = sqrdists > radius ** 2
group_idx[mask] = torch.tensor(N)
  • 方法2 (速度较慢)
    for循环替代`
mask = sqrdists > radius ** 2
group_idx[mask] = N

变为

B, rows, cols = sqrdists.shape
ref_redius = radius ** 2
for b in range(B):
    for r in range(rows):
        print("r: ", r)
        for c in range(cols):
            if sqrdists[b][r][c] > ref_redius:
                group_idx[b][r][c] = N
原文地址:https://www.cnblogs.com/xiaxuexiaoab/p/15555066.html