Onnxruntime错误9:未实现LeakyRelu

Question

我已经将一个简单的pytorch模型转换为onnx格式，并试图通过onnx在单独的文件中加载和评估该模型，但失败了。它给出了错误消息：

NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for the node LeakyRelu_1:LeakyRelu(6)

但是文档清楚地列出了leakyrelu运算符：https://github.com/microsoft/onnxruntime/blob/master/docs/OperatorKernels.md

转换的脚本非常简单。

import torch.onnx
import torch

torch.set_default_dtype(torch.float64)

Model = torch.load("Test.pth")
inputs = torch.randn(1,2)
torch.onnx.export(Model,inputs, "Test.onnx", export_params=True)

加载转换后的模型的脚本也非常简单，

import onnx
import onnxruntime as rt
import numpy as np


onnx_model = onnx.load("Test.onnx")
onnx.checker.check_model(onnx_model)

#print(onnx_model.graph)
print("[Graph Input] name: {}, shape: {}".format(onnx_model.graph.input[0].name,
                                                 [dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim]))
print("[Graph Output] name: {}, shape: {}".format(onnx_model.graph.output[0].name,
                                                  [dim.dim_value for dim in onnx_model.graph.output[0].type.tensor_type.shape.dim]))


print(onnx.helper.printable_graph(onnx_model.graph))

sess = rt.InferenceSession("Test.onnx")

打印的图表是

graph torch-jit-export (
  %input.1[DOUBLE, 1x2]
) initializers (
  %0.weight[DOUBLE, 228x2]
  %0.bias[DOUBLE, 228]
  %2.weight[DOUBLE, 70x228]
  %2.bias[DOUBLE, 70]
  %4.weight[DOUBLE, 1x70]
  %4.bias[DOUBLE, 1]
) {
  %7 = Gemm[alpha = 1, beta = 1, transB = 1](%input.1, %0.weight, %0.bias)
  %8 = LeakyRelu[alpha = 0.00999999977648258](%7)
  %9 = Gemm[alpha = 1, beta = 1, transB = 1](%8, %2.weight, %2.bias)
  %10 = LeakyRelu[alpha = 0.00999999977648258](%9)
  %11 = Gemm[alpha = 1, beta = 1, transB = 1](%10, %4.weight, %4.bias)
  return %11
}

我想知道leakyrelu是否已经实现，或者我只是在转换过程中遗漏了一些东西。谢谢你的帮助！

Answer 1

根据文档https://github.com/microsoft/onnxruntime/blob/master/docs/OperatorKernels.md，LeakyRelu只为float (32位)类型实现，而您有double (64位)。

您可以尝试在PyTorch代码中的LeakyRely之前将其转换为32位浮点数。也许会在ONNX Runtime Github上创建一个问题，以增加对LeakyRelu的双重支持。