symbolically traced variables cannot be used as inputs[BUG] Issue title...

[kanlions]

Describe the bug
A clear and concise description of what the bug is.
This happens when we try to integrate TorchFX with TIMM models. For example try 'regnety_032'
To Reproduce
Steps to reproduce the behavior:

1. from torchvision.models.feature_extraction import get_graph_node_names, create_feature_extractor
   2.model = timm.create_model('regnety_032', pretrained=True, features_only=True)
2. nodes,_ = get_graph_node_names(model)

Expected behavior
A clear and concise description of what you expected to happen.
We would have got the layers

Screenshots
If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. Windows 10, Ubuntu 18.04]
• This repository version [e.g. pip 0.3.1 or commit ref]
• PyTorch version w/ CUDA/cuDNN [e.g. from conda list, 1.7.0 py3.8_cuda11.0.221_cudnn8.0.3_0]

Additional context
Add any other context about the problem here.

[rwightman]

@kanlion there are some incompatibilities with some leaf modules for various models that prevent symoblic tracing from working without handling those nodes specially, using the torchvision classes directly like that will fail for them. I hope to resolve them over time

Your use case isn't clear to me as you'd typically use the create_feature_extractor functionality for doing the equivalent of features_only=True ... so not sure why you want both?

You can use FX based feature extraction in place of the timm feature extractors by doing somethign like
model = timm.create_model('regnety_032', pretrained=True, features_only=True, feature_cfg=dict(feature_cls='fx')) which uses the create_feature_extractor internally but ignores problematic leaf nodes...

[rwightman]

@kanlions tracing without using the leaf exceptions (problematic modules not to trace through, https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/fx_features.py#L26-L44) was causing an issue with BatchNormAct2d layer, used in regnets and some other networks in timm.

I'm fixing BatchNormAct2d and hopefully getting BlurPool and DropPath off the list too so that fewer networks/layers will have tracing issues when not using the timm helpers ... will be checked in soon...