Implement LRN

[salvacarrion]

This layer is used in many early DNNs such as AlexNet, GoogLeNet, Inception, etc. so we need to implement it to support these models through ONNX.

For now, we have two workarounds to be made:

• Option 1.: Bypass this layer in ONNX showing a warning
• Option 2.: Replace it with BatchNorm where its parameters are all ones and zeros

Info:

• PyTorch: https://pytorch.org/docs/stable/generated/torch.nn.LocalResponseNorm.html
• Blog: https://towardsdatascience.com/difference-between-local-response-normalization-and-batch-normalization-272308c034ac

[salvacarrion]

I think it is supported by cuDNN: https://docs.nvidia.com/deeplearning/cudnn/api/index.html#cudnnSetLRNDescriptor