torch2onnx

问题一：torch模型中有条件支路

ONNX export failure: Exporting the operator silu to ONNX opset version 12 is not supported. Please open a bug to request ONNX export support for the missing operator.

参考link 修改silu实现

C:\Users\admin\.conda\envs\torch\Lib\site-packages\torch\nn\modules

class SiLU(Module):__constants__ = [&＃39;inplace&＃39;]inplace: booldef __init__(self, inplace: bool = False):super(SiLU, self).__init__()self.inplace = inplacedef forward(self, input: Tensor) -> Tensor:# ------------------------------------- ## 把F.silu替换掉，修改后如下return input * torch.sigmoid(input)#原来的代码return F.silu(input, inplace=self.inplace)

2  trt

ERROR: builtin_op_importers.cpp:1602 In function importIf:
[8] Assertion failed: cond.is_weights() && cond.weights().count() == 1 && "If condition must be a initializer!"

vino

There is no registered "infer" function for node "If_156" with op = "If". Please implement this function in the extensions.

参考link

link

错误是发生在importIf，即if节点，可是我forward函数里没有写任何判断，研究了一下转onnx的输出，发现问题出自squeeze()函数，torch1.7版本在onnx里加了一个判断操作，判断要squeeze的维度是不是1：