Add symbolic for SwishJitAutoFn to support onnx

[tigert1998]

No description provided.

[tigert1998]

There's a series of jit activation functions but I only changed one of them. I am here to kindly remind you that simply adding a symbolic function can support exporting to onnx format and make it easy for edge deployment.

[rwightman]

@tigert1998 thanks for pointing that out, will keep it in mind... there are reasons besides onnx export to not use the AutoGrad + Jit version, the activation factory allows switching between different variants of those activations, including ones that are onnx exportable. Now that SiLU is in PyTorch (as of 1.7) it's moot for swish.

[tigert1998]

But I also suggest adding this workaround since not everyone can use torch 1.7+. In my case, I have to share GPUs with my colleagues. The servers use cuda 10.1 and people use torch 1.6. I am not the administrator so I do not have permission to update cuda version.
I think adding the symbolic workaround with a warning might be a not-so-bad choice.

[rwightman]

@tigert1998 you're aware you can create the model with create_model('efficientnet_b0', exportable=True) ? As mentioned, there are other reasons besides onnx exporting that might need to avoid the autograd.Function or jit completely