When exporting a PyTorch model to ONNX, if the model has the range indices selection operator (e.g. x[2:]) and didn't specify the ending index, the "end" attribute in the exported Slice operator will be an extremely large number. This happens with PyTorch 1.0 1.1 and 1.2. However, in 0.4, there isn't such a case. I wonder if this behavior is intended?

e.g. the model is like this:

class testMod(nn.Module):
    def forward(self, x):
        return x[2:]

model = testMod()
x = torch.rand(10)
torch.onnx.export(model, x, "slice.onnx", verbose=True, input_names=["input"], output_names=["output"], opset_version = 10)

The exported Slice op is:
%output : Float(8) = onnx::Slice[axes=[0], ends=[**9223372036854775807**], starts=[2]](%input), scope: testMod

If using v0.4 to export, the ending index could be inferred:
%output : Float(8) = onnx::Slice[axes=[0], ends=[**10**], starts=[2]](%input), scope: testMod

Many Thanks