I exported original AlexNet to .onnx using a function: torch.onnx.export().
The original model has LocalResponseNormalization. This one layer is converted to ~30 layers of combination with different names such as Mul, Unsqueeze, Gather, etc.
On the other hand, when I use chainer model and convert the original model uisng onnx-chainer. The LocalResponseNormalization layer is still LocalResponseNormalization, which is actuuly supported by the onnx-chainer package accrodibg to onnx-chainer · PyPI
Pytorhc is not supporting LocalResponseNormalization itself?