🐛 [Bug] Log print misleading error messages when there is unsupported operator

[bowang007]

Bug Description

When we have some unsupported operator in a network, it will automatically fallback with some messages showing that this operator is not supported. However, currently the message is misleading because it shows as an error which may make users think the compilation failed.

To Reproduce

For example, have many errors such as this we I support MaskRCNN:

ERROR: [Torch-TensorRT - Debug Build] - Unsupported operator: aten::max.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)
/home/bo/Projects/detectron2/detectron2/structures/image_list.py(83): from_tensors
/home/bo/Projects/detectron2/detectron2/modeling/meta_arch/rcnn.py(226): preprocess_image
/home/bo/Projects/detectron2/detectron2/modeling/meta_arch/rcnn.py(199): inference
export_model.py(112): inference
/home/bo/Projects/detectron2/detectron2/export/flatten.py(294): forward
/home/bo/.local/lib/python3.6/site-packages/torch/nn/modules/module.py(1090): _slow_forward
/home/bo/.local/lib/python3.6/site-packages/torch/nn/modules/module.py(1102): _call_impl
/home/bo/.local/lib/python3.6/site-packages/torch/jit/_trace.py(965): trace_module
/home/bo/.local/lib/python3.6/site-packages/torch/jit/_trace.py(750): trace
export_model.py(121): export_tracing
export_model.py(221): <module>
Serialized   File "code/__torch__/detectron2/export/flatten.py", line 24
    _1 = ops.prim.NumToTensor(torch.size(t, 2))
    image_size = torch.stack([_0, _1])
    max_size, _2 = torch.max(torch.stack([image_size]), 0)
                   ~~~~~~~~~ <--- HERE
    _3 = torch.div(torch.add(max_size, CONSTANTS.c0), CONSTANTS.c1, rounding_mode="floor")
    max_size0 = torch.mul(_3, CONSTANTS.c1)

Expected behavior

Give a warning, rather than error.

Additional context

will raise a PR for this soon.

[narendasan]

I have a fix for this as part of #918

[chaoz-dev]

For vis, related to #803