Should we modify tracked buffer in this norm layer to support symbolic tracing?

[super-dainiu]

https://github.com/rwightman/pytorch-image-models/blob/94a91598c30125f6d95b08c7f2edcafd72b7dfa6/timm/layers/norm_act.py#L80
might be replaced with the PyTorch implementation
self.num_batches_tracked.add_(1) # type: ignore[has-type] source
because it is more tracer friendly.
This way, the tracer can detect the add_ operand with Proxy and create a new FX Node without assigning a Proxy to the buffer.

[rwightman]

@super-dainiu good point, PR welcome!

[rayennecandido]

Thank, you.

[rwightman]

FYI it's in #1633 to be merged in day or so