Failed to export an ONNX attribute, since it's not constant, please try to make things (e.g., kernel size) static if possible

[dayekuaipao]

So it only supports the model in torchvision,not my own?

[puru-d]

I am also facing this issue with custom model. So you are telling it does not support custom models. Because i tried on pre-trained squeezenet and it worked well for me.

[sytelus]

We support any model. If you had like to investigate this issue then please add detailed steps that would allow us to reproduce the issue. Include what you are doing, what is your model etc.

[Max-ChenFei]

Hi, I meet some problem using torchvision

model = torchvision.models.segmentation.deeplabv3_resnet101(pretrained=True)
model.eval()
trace, _ =  torch.jit.get_trace_graph(model, torch.randn(1, 3, 224, 224))
torch.onnx._optimize_trace(trace, torch.onnx.OperatorExportTypes.ONNX)

Could you please give some hits to save this problem?

[sruthikesh-MU]

Facing the same issue with torchvision mobilenet. The issue is in onnx._optimize_trace

is there a solution for this? resnet_50 and googlenet worked fine, some networks have this issue