Matteo Salvarezza

> Just to make sure, I understand this correct: what happens to Squeeze and Unsqueeze upgrade when there is no Loop? Is the axes attribute turned into an initializer? If...

What I don't like with solution (2) or with using `Constant` ops is that it doesn't automatically take care of the problem: all future adapters should also not add graph...

Ok so there is a case for adopting both solutions simultaneously: (3) to make sure that the version converter never produces an invalid model and (2) as a good practice...

I have changed all adapters adding initializer + input to add Constant ops instead: https://github.com/onnx/onnx/pull/3344

I would consider this as being still relevant because solution (3) is still a desirable thing to have

@DickJC123 1) Training would obviously be the same, but there is a corner case when one might want to perform several forward passes in training mode without doing backwards. In...

@DickJC123 you were in fact right about biased vs unbiased variance computation. This script tests such claim by letting a non-cudnn batchnorm and a cudnn-batchnorm update their moving variance, and...

So, to summarize the issues I found with the cuDNN implementation: 1) Moving arrays are normally updated only if a BACKWARD pass in training mode is performed, but on GPU...

@leezu unfortunately we are still tied to 1.x versions, and I'd like to avoid switching to 2.0 master for the time being (but we'll make the switch at some point)....

@leezu I've managed to get master to build on Mac-ARM (with a small fix: https://github.com/apache/incubator-mxnet/pull/20072). I will now look into the changes needed by 1.x @mseth10 I'm doing native compilation...